AI in social care
Published 20 November 2026
Applies to England
Summary
Guidance from the Department for Education (DfE) and the Department of Health and Social Care (DHSC) on the safe, responsible and appropriate use of artificial intelligence (AI) in children’s and adult social care in England.
It is intended to help local authorities, social care organisations and practitioners consider how AI may be used responsibly in practice.
It is not a technical guide and does not replace existing legal, regulatory or professional requirements.
DfE and DHSC recognise that, used appropriately, AI can support children’s and adult social care.
AI has the potential to:
- support better outcomes
- create more time for meaningful engagement with individuals and families
- help practitioners develop a more informed and reflective understanding
- identify gaps in information and help organisations make better use of information
- support resource planning
- strengthen decision-making and strategic oversight
Social care is fundamentally relationship-based. AI should always strengthen, not replace, human connection, professional judgement and accountability.
AI use raises ethical questions around:
- accuracy and bias
- transparency
- consent
- accountability
- data protection
- security risks
Local authorities and social care organisations should ensure AI use is governed appropriately. There should be clear guardrails in place, including:
- acceptable use
- data handling rules
- training and supervision
- monitoring and review
Responsibility and accountability remain with practitioners and their organisations.
Organisations should provide clarity on the AI tools that can be used to support practice, and the parameters that govern this use. No personal data should be entered into publicly available AI tools operating outside the organisation’s secure environment or enterprise agreements that protect onward sharing or use of data.
Decisions about whether or how to adopt AI remain with organisations and should be taken in line with their:
- legal duties
- local policies
- data protection arrangements
- professional standards and responsibilities
This guidance is intended to provide a foundation for DfE and DHSC to develop further resources (jointly where appropriate) to support ongoing responsible implementation, use and governance of AI tools in social care.
Who this publication is for
This guidance highlights universal themes applicable across organisations within children’s and adult social care. We understand that approaches will vary across local areas and organisations. The application of the guidance is at the discretion of the organisation.
This guidance is for:
- frontline practitioners
- supervisors
- managers
- digital and data teams
- transformation teams
- data protection officers
- Caldicott guardians
- procurement staff
- other organisations providing social care and related services
The principles of this guidance will also apply to other organisations providing social care and related services.
Using this guidance
This guidance is intended to set out DfE and DHSC’s position on AI in social care and provide high-level guidance on its opportunities, risks and ethics.
We would recommend that the whole document is read and understood, but the following sections may help with specific queries:
- how AI may support children’s and adult social care may help transformation or technical staff identify where AI could support practice
- considerations for organisations and practitioners may be useful for practitioners and supervisors in understanding how to safely use AI in practice
- considerations for the appropriate use of AI may also help managers and technical or transformation staff create the guardrails to support the safe use of AI
- ethics, risks and challenges may help practitioners and supervisors understand the different risks and ethical dilemmas they may encounter when using AI
- Using AI in social care: a quick guide for practitioners may help those new or unfamiliar with AI understand the basics
DfE has guidance on Generative artificial intelligence (AI) in education.
Main points
AI in this document means the tools used to:
- generate, summarise, transcribe or organise information
- support service improvement, communication, accessibility and analysis in support of children’s and adult social care practice
AI, used responsibly, can support social care practice by:
- strengthening relationship-based and reflective practice
- improving how information is used
- testing assumptions
- identifying gaps and reducing administrative burdens
- strengthening services and supporting better decisions
AI is not infallible – it makes mistakes and practitioners should challenge AI outputs.
Responsibility and accountability remain with practitioners and their organisations. AI outputs should be treated as a first draft to be reviewed and amended before use.
Risks including bias, inaccuracy or hallucinations, privacy and security concerns, and automation bias (where AI outputs may be given undue weight by practitioners) should be identified and managed.
Practitioners should be able to clearly explain the AI tools being used, why, and what will be done with the outputs. This helps build trust and transparency.
Local authorities and social care organisations should have or be working to implement governance (including AI use policies), assurance and practical guardrails so AI is used safely, lawfully and appropriately.
Only platforms that are secure and data protection compliant should be used, and in line with organisational policies. Personal data should not be entered into AI tools that have not been approved by the organisation, as these are not secure.
The use of AI in social care should not jeopardise or impede the relational values of care, individual rights and dignities.
AI is not sentient and does not understand the personal contexts of a child, family, individual or their lived experiences. AI should not replace practitioners’ judgement or make automated decisions about risk, need, eligibility or intervention.
DfE and DHSC’s position: 4 principles for the use of AI in social care
This guidance is non-statutory and should be applied in line with local policies, governance arrangements and professional standards. DfE’s and DHSC’s view is that:
1. Practitioners remain responsible for their casework, recording and decision-making. AI-generated content should be treated as a first draft and reviewed, challenged and amended where necessary to ensure it reflects context, observations and critical reflections.
2. AI should not make decisions that directly affect people. Decisions relating to risk, need, eligibility, safeguarding, care planning and intervention should remain the responsibility of practitioners exercising professional judgement.
3. AI should support better experiences and outcomes for children, adults and families. It should strengthen relationship-based practice by improving communications, reducing the burden of administrative tasks, supporting access to information and helping practitioners spend more time working directly with people.
4. Personal information should not be entered into publicly accessible AI tools. AI should only be used in line with organisational policies, using approved tools and according to data protection policies.
AI is not sentient and does not understand the personal contexts of a child, family or individual or their lived experiences. AI should not replace practitioners’ judgement or make automated decisions about risk, need, eligibility or intervention.
AI and human relationships
Social care is relationship-based, and AI has the potential to influence human connection both positively and negatively. It may support practitioners to reduce administrative tasks and reflect on their practice, but if implemented poorly it may reduce opportunities for meaningful engagement and undermine trust.
Responsible and appropriate use of AI should consider how AI supports the values of care and does not erode them. The adoption of AI in social care should not be judged solely by efficiency gains but also by how it contributes to better experiences and stronger human connections.
Organisations and practitioners should be aware of the impact that the increasing focus on the use of AI (and other digital tools) may have on those experiencing digital exclusion. Some children, families, carers or individuals drawing on care and support may:
- have low levels of digital confidence
- have limited access to devices or connectivity
- require support to engage with technology
- experience cognitive impairments that affect digital participation
Taking an ethically robust, transparent and well-governed approach to implementing AI, built around the relationships and importance of human connection may help mitigate against the use of AI widening the digital exclusion divide.
Organisations developing tools or shaping implementation may want to consider how young people, families or individuals drawing on care can be involved in service design or governance. This can help mitigate against other forms of potential digital exclusion such as literacy barriers, under-representation and inaccessible outputs.
How AI may support children’s and adult social care
Social care has a significant impact on the lives of the individuals and families supported by practitioners. Children, families or individuals drawing on care should be at the heart of every decision made about them.
AI should support practitioners and relationship-based, person-centred care – not replace their professional judgement.
Support and reduce administrative burden
AI tools have the potential to reduce the time spent on routine and repetitive tasks, particularly recording and paperwork, although their effectiveness and safety can vary. For example, AI tools may be able to:
- support speech‑to‑text transcription of visits, meetings and supervision notes, with practitioners reviewing, ensuring factors such as body language, physical environment, tone of voice and other areas not picked up by AI are reflected
- help draft case notes, reports, chronologies and letters, which practitioners review, amend and finalise
- structure and format records in line with agreed templates and recording standards, with practitioners checking for accuracy and bias
- help ensure children, families and carers, and those who draw on care are heard and their voice and needs are accurately reflected in their care notes
Case study: Kingston Council
A local authority introduced an AI-enabled tool to support them to draft case notes, assessments and supervision records in adult social care. The initiative was driven by increasing demand for services and a desire to reduce the amount of time practitioners spent on administrative tasks. This would allow them to focus more time on direct work with people who draw on care and support.
The tool helped practitioners transcribe and summarise conversations into draft records, which remained subject to practitioner review and approval. Social workers reported an average of 50 to 60 per cent reduction in time spent completing case notes and assessments. Supervision write-ups reduced from 40 minutes to less than 10 minutes.
Kingston Council: Using AI in Adult Social Care administration Local Government Association has more information.
AI tools, used effectively, also have the potential to support improvements to workforce wellbeing by reducing frustration linked to high workloads and heavy recording requirements, which can contribute to stress and burnout.
Case study: Together for Children Sunderland
When evaluating the impact of AI during the implementation of an AI-enabled transcription tool, Together for Children Sunderland worked with social workers to understand the impact on workforce wellbeing.
The evaluation involved:
- journey mapping practitioners’ day-to-day experiences before and after adopting AI
- baseline data collection
- time-motion studies
- practitioner feedback
- post-pilot surveys
The results found:
- reductions in the time spent on administrative tasks
- improvements to worker wellbeing and mental health
- reductions in stress
- productivity benefits
- improved quality of conversations and case recordings
Improving the quality and consistency of case recording and care planning
AI tools may support higher quality case records and care planning by:
- reducing unnecessary repetition across assessments, plans and meeting records
- supporting clearer and more concise writing, while allowing practitioners to apply their professional judgement and narrative
- supporting the removal of judgemental or victim-blaming language that can be re-traumatising to individuals reviewing their records
- helping improve the quality of writing and use of appropriate language for example:
- using ‘language that cares’ and writing ‘to’ individuals
- identifying bias, subjectivity and errors
- identifying areas where additional information or reflection may be needed, such as analysis, next steps and the voice of the individual in records
- offering the potential to pre-fill data fields in case-records
- helping staff to better use information in their professional judgement
This can support better-informed decision-making by helping practitioners:
- identify relevant information
- understand patterns across complex case records
- highlight gaps for further enquiry
- access evidence more efficiently
The final decision should always remain with the practitioner.
Supporting good practice
AI tools may be able to support relational and strengths-based practice by:
- allowing practitioners to be more focused on direct work and engagement with people who draw on care, children, families and carers
- supporting practitioners to prepare for visits, interactions and conversations
- supporting practitioners’ professional curiosity and thinking by structuring information, testing analysis, identifying gaps, and challenging assumptions
- helping practitioners test analysis and challenge assumptions
Case study: Norfolk County Council fostering teams
Norfolk County Council’s fostering team used a Microsoft Copilot custom-built supervision agent. This was embedded with National Minimum Standards, relevant legislation and local guidance to support reflective drafting and structured analysis.
The agent analyses transcripts alongside the supervising social worker’s notes, observations and reflections to produce a structured first draft of a supervision record.
This supports deeper reflection, sharper analysis and clearer professional judgement, reinforcing rather than replacing professional expertise. This has led to clear, more consistent and more analytically rigorous supervision records.
Supporting the use of information in cases, case records and care planning
AI may help practitioners find and use information more easily, especially in long or complex case records. For example, AI enabled tools have the potential to:
- help staff quickly understand and search case histories to find relevant information;
- produce summaries of lengthy records to support preparation for visits and reviews
- highlight missing sections or gaps that may need further consideration
- retrieve and summarise information from multiple systems
This can support better‑informed decision‑making, while decisions remain with practitioners.
Supporting reflection, supervision and workforce wellbeing
AI’s ability to process and summarise complex information has the potential to enhance practitioner reflection and supervision. This includes:
- helping practitioners prepare reflective summaries for supervision, for example through posing reflective questions, or prompting curiosity
- supporting supervisors to identify themes across workloads, while maintaining accountability for decisions
Supporting reasonable adjustments
AI may help local authorities support practitioners, children, parents and carers and people who draw on care and support, by removing disability-related barriers to:
- communication
- comprehension
- drafting
- information processing
This can support:
- the creation of accessible versions of information for children, young people and adults with different communication needs
- reasonable adjustments for practitioners, making the career more accessible
Supporting integration
AI has the potential to support integration between systems, both within organisations and across organisational boundaries. This may include improving the sharing of information between practitioners and other agencies, extending beyond individual interactions with social care services. This will help to deliver more integrated care across the wider health and care ecosystem.
For this to be effective, tools will need to be tested for accuracy and security, grounded in relevant materials, and subject to effective governance. Where information or systems span organisational boundaries, appropriate agreements will also need to be in place.
Strategic analysis, commissioning and planning
AI’s ability to process and analyse large datasets may help organisations develop strategic insights to support planning, resource management and strategic decision-making.
AI may also help organisations identify gaps in provision and highlight emerging pressures or risks that may require further consideration. This can support more informed planning and decision making across the social care system.
For example, AI used in this capacity may help organisations:
- analyse demand, referral patterns, waiting times or identify changes in demand for services
- identify gaps in local support and provisions or unmet needs
- help leaders understand workforce pressures and service bottlenecks
- identify emerging pressures and risks on service provision
- identify commissioning needs and support procurement decisions
- evaluate the impact of services
- support earlier planning for placement sufficiency, care packages or community support
Exploring AI’s capability to analyse large amounts of data could also support better strategic commissioning through the development of more accurate, data-driven and evidence-led needs assessments and forecasts.
There is limited evidence of the use of this aspect of AI across social care, there are examples of local authorities using this capability to support other areas, such as housing and waste.
Case study: AI in Swindon Borough Council’s Adult Social Care
Swindon piloted an AI tool to support social workers. They recognised the impact that administrative tasks have on the time available for social workers to focus on person-centred care. This can be amplified for colleagues with learning difficulties or other protected characteristics.
The pilot found:
- a 63% reduction in admin time, helping them to allocate more time to supporting individuals
- faster assessments – the time to conduct a Care Act Conversation fell from 1.5 hours to 30 minutes, and time taken to write up assessments fell from 4 hours to 1.5 hours
The tool has been particularly beneficial for colleagues with learning difficulties, visual impairments, and those where English is not their first language. Dyslexic social workers and those with learning difficulties reported benefits in their practice and recording.
More information is available from the news article: Council completes AI trial to enhance Adult Social Care services Swindon Borough Council.
Translation
AI-enabled translation may support practitioners to undertake direct work with families where English is not their first, or spoken, language. This may reduce barriers and improve timeliness where there could be delays in finding a translator.
Case study: using AI to improve accessibility
A significant proportion of care leavers, particularly unaccompanied asylum-seeking children (UASC), have English as a second language or speak no English at all. Approximately 25% of care leavers cannot access the digital information and support provided. Research found many care leavers were unaware of the support available when leaving care.
AI Solution
The Support for Care Leavers website uses AI-powered translation to provide content in more than 100 languages. The tool generates translations instantly, allowing users to access information in their preferred language without requiring manual translation.
Outcome
AI has enabled a more accessible and inclusive service, helping more care leavers understand their rights and the support available as they transition to independent living.
Implications
AI-powered translation can enhance digital accessibility and, where integrated with wider services, help people to find the right support earlier. It can also help to facilitate informed decision-making, reduce pressure on services and can improve outcomes for councils and communities across children’s and adult social care.
Where AI-enabled translation tools are being used, organisations should ensure processes are in place to ensure and evaluate accuracy – particularly where translations are likely to be used in a statutory process. In these cases, organisations should consider when a qualified interpreter is required, rather than relying solely on AI-enabled translation.
Translation tools are also supporting social workers who have English as a second language to check and improve their written English. This may carry some risks relating to over-reliance, but has the potential to improve the quality of case recording and care planning.
Improving how families, carers and individuals engage with services
Implemented appropriately, with effective guardrails, AI has the potential to improve how citizens engage with social care services. For example, it has the potential to:
- support plain-English explanations of assessments, plans and the statutory process
- reduce power imbalances by improving how families and individuals understand statutory frameworks and the experience, support or care they are likely to receive
- enable children, parents, carers and people who draw on care to access records in a more appropriate and compassionate way
- analyse service user feedback to improve services
- enable tools to support citizens to understand what support may be available.
- capture the thoughts, feelings, aspirations and challenges a child, family, carer or individual drawing on care may face
Some of these use cases are still emerging in social care and should be evaluated before wider adoption.
Evaluation should consider:
- accuracy
- accessibility
- equality
- user experience
- data protection
- information governance
- whether the tool improves people’s ability to understand, participate in and shape their care and support
Improving access to services
AI’s ability to process natural language and generate accurate responses when trained on trusted, approved information and advice is improving how citizens access services.
With appropriate governance and safeguards, chatbots are being used to help citizens access the support and solutions they need. Translation capabilities increase equality of access – having access available 24 hours a day, 7 days a week reduces barriers to getting the right support.
Case study: Bradford Council – supporting adult social care front door with AI digital assistants
Bradford Council has introduced a digital assistant, ‘Annie’, accessed online or through WhatsApp. Annie offers a new access channel available 24 hours a day, 7 days a week – helping people get the right information at the right time. Annie operates from a ‘locked box’. This means that it is specifically trained on trusted, approved information and advice from content that is curated by the council. It can process over 40 languages across both text inputs and voice notes, reducing barriers to use and services.
Built into Annie is the concept that if a person asks more than 4 questions, they probably have a complex situation, need to talk to a human and are directed to the ‘front door’. Behind Annie there is an active monitoring layer which flags any potential safeguarding issues. Users are advised not to enter personal information through the user interface. If any personal data is shared, this is identified and redacted in the backend.
Bradford Council: supporting the ASC front door with AI digital assistants has more information.
Some of these use cases are yet to be explored in a social care context, but they are being seen across other sectors. They should be carefully evaluated before adoption, regularly reviewed and monitored for the accuracy of outputs and the governance and protection of information.
Limitations on AI use
Where AI should not be used
Engagement with local authorities and other sector organisations highlights that it would be helpful to know where AI should not be used in practice. DfE’s view is that AI, in its current form, should be approached with significant extra caution or avoided in the following high-risk or high-consequence situations.
AI alone should not determine risk, need, eligibility, safeguarding thresholds or interventions, or independently make or suggest decisions that directly affect an individual, child, young person or family
AI cannot properly understand context, nuance or inference, and decisions should remain with practitioners. For example, AI should not determine what intervention should be taken in a case. AI may support practitioners by organising information, identifying possible patterns, highlighting gaps or inconsistencies and prompting further professional enquiry, helping practitioners consider risk systematically and decide what further information or action may be needed.
Directly recording into case records and care planning without meaningful human oversight
This may lead to inaccuracies perpetuating throughout case records and ultimately presenting an inaccurate version of events.
Making an independent assessment of individual needs or risks that are not subject to the reflections and judgement of the practitioner
AI tools should not be making independent assessments of risks or care and support needs. They are better placed in helping the practitioner gather, summarise and process evidence to help them make their professionally informed assessment of needs or risks. Predictive tools may support strategic planning or professional enquiry but should not determine decisions affecting individuals without professional judgement and oversight.
Where AI should be used with caution
Making predictions about individual needs
Although some AI tools can use predictive analytics to process large amounts of data and generate forecasts and predictions on an individual’s likely needs, caution should be used when considering whether to use this capability.
If AI tools are being used for predictive analytics, organisations should ensure models are well-documented and understood, tested for accuracy and bias, and subject to ongoing evaluation to assess their impact and effectiveness.
Automating elements of processes and workflows
AI, particularly Agentic AI (for example an AI model that is capable of completing tasks independently or alongside other AI models) has the potential to streamline workflows and automate processes.
Agentic AI could reduce time spent on:
- communication
- administrative tasks
- information retrieval
Use caution to ensure that this is appropriate, accurate, and does not undermine the professional integrity of practitioners.
A lower-risk use may include using simple agents to support practitioners to challenge their reasoning and reflect on their practice.
A more cautious and considered approach may be needed when using agentic AI for more complex cases, such as directly engaging with people who draw on care or completing workflows within case management and care planning systems.
Robust local governance, feedback loops and policies should be in place to safeguard quality, consistency and accuracy. Agentic AI may need to:
- be grounded in specific, context-relevant content
- operate within clearly defined instructions and boundaries
- remain subject to meaningful human oversight and intervention
A range of tools and guides are available to help public sector bodies govern and openly document their use of AI, including the:
- Responsible AI Toolkit
- AI Risk Management Toolkit
- Algorithmic Transparency Recording Standard (ATRS)
Supporting or assisting with triage of cases
There are examples of AI being effectively used to support and assist with triaging cases. Caution should be used to ensure that:
- AI models are not making decisions that will impact the care or wellbeing of individuals without meaningful human oversight
- organisations understand the accuracy and limitations of the model, including any potential for bias
- consideration is given to how oversight is maintained, including for example, through auditability of individuals’ journeys through the ‘system’
Predictive analytics, AI-enabled strategic insights and triage tools may support strategic planning, prioritisation for professional review or the identification of issues for human consideration.
Some of the greatest value of these uses of AI comes from helping practitioners or leaders ask better questions, identify patterns and target further enquiry. They should not determine outcomes for individuals without practitioner judgement, governance and meaningful human oversight.
There are examples across government and the public sector where AI-assisted decision making is proving effective:
- the NHS is proving AI’s ability to support with diagnosis and medical image interpretation
- policing is using facial recognition and digital forensics to support CCTV analysis and sorting through digital evidence
- some local authorities are using AI to improve waste collections and streamline customer services
This reflects the current position of AI, and its use in children’s and adult social care. We will continue to review and revisit this as evidence emerges and technology continues to develop.
Ethics, risks and challenges of AI
AI offers a range of opportunities to support practitioners and the social care sector but brings a set of new risks, challenges and ethical considerations. Some of these stem from the way AI models are developed and work, others from how we interact with them. For example, over-reliance on AI outputs and reduced professional judgement may degrade decision-making.
The use of AI may, at times, involve trade-offs between potential benefits and ethical values. These should be:
- documented
- governed transparently
- considered with practitioners and people affected by the system, where appropriate
The rise of easily accessible AI in recent years introduces vulnerabilities that extend beyond the ethical questions and risks raised in this guidance. For example, the Independent Inquiry into Child Sexual Abuse highlights that AI is being used to create artificial child sexual abuse material and manipulate still or moving images of individuals.
Annex A has information on ethical considerations, risks and challenges and potential mitigations, alongside a checklist for practitioners.
Considerations for the appropriate use of AI
The following section sets out some of the main considerations for the appropriate use of AI in children’s and adult social care. Many of these are not specific to the sector but carry more importance due to the sensitivity of ‘data’ relating to people who draw on care, children, families and carers supported through social care and the potential impacts of decisions made with (and without) the use of AI.
AI tools should work within an interconnected system where the individual, child, young person or family is at the heart, and the practitioner is responsible and accountable.
Governance
We recommend that local authorities and social care organisations should have, or be working to implement, governance arrangements for the use of AI in frontline practice. This should extend beyond corporate level governance and should be reflective of the high-risk and high-consequence nature of social care and how AI works as a technology.
Without effective governance, oversight and accountability remain vague. This can reduce the effectiveness of AI and increase the organisational risks associated with inappropriate use of AI. Organisations may want to provide clarity (for example through clear and accessible policies) for practitioners on what tools can be used and for what purposes.
Robust governance of AI tools should consider:
- data governance – for example, where data processed by AI tools sits and retention periods
- people governance such as:
- usage policies
- impacts on practice
- human oversight and meaningful human control
- training and learning
Main points for organisations
Effective governance should, as a minimum:
- identify and manage ethical, legal, safeguarding and operational risks
- assess and manage data protection implications, including legal bases for AI use in the collection and processing of data
- set clear rules on approved, restricted and prohibited use (for example AI usage policy)
- support staff to understand how AI may and may not be used
- establish approaches to training and supervision and oversight
- define roles, responsibilities and escalation routes
- provide assurance, monitoring and review arrangements, and escalation routes
- include procurement and supplier assurance
- clarify:
- consent
- agreements
- transparency arrangements
Annex B includes a governance checklist intended to help organisations assess their approach to governing AI tools in social care.
AI tools should comply with relevant legislation and regulation, including data protection legislation. Local authorities and organisations should also align with established transparency and data protection standards (such as Article 5 of UK General Data Protection Regulation (GDPR)).
The Information Commissioner’s Office (ICO) has published guidance on AI and data protection.
Algorithmic Transparency Recording Standard (ATRS) from Government Digital Service helps public sector organisations to publish information about the algorithmic tools they are using and why they are using them.
AI Readiness and self-assessment
Organisations that have not yet started to adopt AI or are at an early stage of their AI journey, may find it helpful to assess their readiness before adopting or scaling AI tools in children’s and adult social care. This includes considering whether they have the right governance, data protection arrangements, technical infrastructure, workforce capability, leadership and assurance in place to use AI safely and effectively.
Local authorities can use the Adult Social Care Assessments Improvement Toolkit Self-assessment Tool to help identify areas of their practice that could be improved through digital and AI-enabled solutions.
The AI Readiness framework and checklist was developed by Bright (the Birmingham AI & Digital Health and Care Group based at the University of Birmingham and University Hospitals Birmingham NHS Foundation Trust) as part of their AI Readiness Project.
It:
- provides a practical framework to support organisations in health and care settings to consider these issues
- can help organisations identify strengths, gaps and areas for further development before implementing AI in practice
The AI Playbook for the UK government and AI Knowledge Hub provides additional guidance on how to govern AI programmes and initiatives.
Implementation and monitoring
AI adoption should be treated as a change programme, rather than a ‘traditional’ technology rollout. It needs ongoing support, clear messaging, visible leadership, workforce capability-building and opportunities for practitioners to shape and test tools. This helps ensure tools reflect user needs, increasing effectiveness and maximising the benefit for the organisation and the citizens they support.
Organisations should consider how to continue to develop the right capabilities and culture to enable the safe use of AI in frontline practice. This could include hands-on training, simple guidance, GDPR and data governance workshops, acceptable use policies, peer learning, AI champions and feedback loops for sharing good practice, issues and concerns.
AI tools are not static. Both the tools themselves, and the models they are built on are likely to be constantly developing and evolving. Reflecting this, organisations should be aware of updates and factor ongoing monitoring in their deployment and adoption of AI, using appropriate governance and oversight arrangements as tools and their use develop.
Case study: Brent Council’s system approach to implementing AI in practice
Brent Council’s children’s social care set out to implement AI in a way that really supports practice: reducing administrative pressure, improving access to information, and creating more time for meaningful work with children and families. The aim was clear: AI should enhance practice, not replace professional judgement.
Approach
Brent’s children’s social care department took a coordinated and carefully governed approach. Practitioners, managers, practice leaders, learning and development colleagues, digital services and information governance teams worked together to ensure AI was introduced with the right safeguards, training and oversight. This balanced approach has helped staff explore the benefits of AI while staying aligned with Brent’s values, responsibilities and ‘Ways of Working’ (Brent’s practice framework).
Impact
- Staff report greater confidence in using AI safely, responsibly and purposefully.
- Practitioners are spending less time on drafting, summarising and routine administrative tasks.
- Teams are creating more capacity for direct, relationship-based work with children and families.
- Managers have additional capacity to support reflective supervision, quality assurance and service improvement ideas.
- Improvements in the department’s understanding of safe, ethical use of AI.
Organisations should be supporting their practitioners to develop their understanding of and confidence in using AI tools.
Robust learning and development offers should both consider tool-specific training and courses and modules on the AI in its broader sense. This may help practitioners understand how AI works, and its limitations – underlining the importance of:
- reviewing outputs
- the ‘human in the loop’
- retaining professional judgement
- avoiding over reliance
Supporting practitioners to develop their AI skills and maturity can help ensure the value of AI tools is increased.
How to Test and Evaluate AI has useful prompts for organisations to consider when implementing and monitoring AI tools.
Case study: Essex County Council’s adult social care approach to implementation
When exploring the implementation of AI in adult social care, a proof-of-concept tested AI transcription and summarisation in adult social care practice produced learning that was relevant to the whole organisation.
It showed how AI can support practice while also highlighting the conditions required to use AI safely, legally and confidently across the organisation.
What Essex County Council learned:
- AI can improve clarity, reduce recording burden and support presence in conversations when strong human oversight is in place
- secure purpose-built tools performed best for high-risk statutory work
- secure general-purpose tools are useful for low risk
Lawful data processing
Article 5 of UK GDPR sets out seven main principles for data protection:
- lawfulness
- fairness and transparency
- purpose limitation
- data minimisation
- accuracy
- storage limitation
- integrity and confidentiality (security)
- accountability
These should be considered when implementing and adopting AI in children’s social care to provide the foundations for good data protection practice.
When processing personal data, whether AI is being used or not, a legal basis under Article 6 of the UK GDPR must be identified. Additionally, where that personal data is special category data (such as health data, racial or ethnic origin and sexual orientation), at least one of the conditions in Article 9 must be satisfied.
The common law duty of confidentiality (CLDC) must also be considered. More information is available from the duty of confidentiality.
The basic principle underlying the CLDC is that a person who receives information in confidence must not take unfair advantage of it or use or disclose it to the prejudice of the confider without consent.
In the absence of consent, the CLDC may be overridden on the basis of legal requirement, statutory power or overriding public interest.
Where automated processing is being used to make individual decisions, organisations must also consider whether the provisions in Section 4A of Chapter III of the UK GDPR apply. These provisions apply where automated processing is used to make a significant decision without meaningful human involvement in the taking of the decision.
Organisations should ensure that they have identified the relevant legal power and undertaken the analysis set out above when using AI and reflect this in their guidance and documentation as part of their implementation and governance approaches.
Ultimately, it is your responsibility to ensure you comply with all data protection legislation, and organisations may want to consider how they support practitioner compliance.
Trust and choice
Trust is fundamental to effective social care, and this should not be jeopardised by using AI or any other tool.
Organisations should support their frontline staff to understand the organisation’s policies on being transparent and what to do when people disagree to the use of AI when working with people who draw on care. This includes:
- children
- young people and families
- people who may not be able to disagree themselves, for example during mental capacity assessments
This may include providing clear and accessible guidance and policies on transparency and the choices people have about the use of AI in practice and how to handle objections. This should be in line with data protection law, safeguarding responsibilities and local, organisational policy.
If practitioners understand their organisation’s policies on choice, and are able to explain the tools they use, this may help to:
- build trust with individuals, children, young people, families and partner organisations
- ensure alignment with data protection and safeguarding obligations
- give individuals the opportunity to object to its use
There may be times when those families or individuals will not agree to AI being used in their care and support. They are within their rights to do so, and organisations should ensure that there are clear alternative processes in place for where individuals refuse permission to use AI.
When used with other organisations (for example in multi-agency meetings), it may be worth considering how agreement to use AI can be addressed through governance or at an organisational level. This is most likely to be relevant when using AI for recording or minuting meetings.
Practitioners should ensure that they adhere to different organisations’ policies on AI when working in partnership with them.
Transparency
Organisations should support transparency by ensuring staff understand what tools are in use, what they do, where they may be used, and how outputs should be used and reviewed. Where appropriate, organisations should also support practitioners to explain AI use clearly to children, families, carers, individuals and partner agencies.
This can be supported by ensuring practitioners:
- understand, at a non-technical level, the AI tools that they are using, and how the output is generated
- can explain how AI will be used in their practice, and outputs reviewed, to individuals, children, families and partner organisations where appropriate
- can explain why AI is being used, understand where and when it can be used and the limitations and hard boundaries
- understand when people should be informed and can disagree to the use of AI
- are clear on how objections to AI use should be managed
- understand and follow relevant organisational guidance
- can feedback issues and challenges and share learning and practice with peers
- are supported by organisational transparency about systems in use
Organisations may find the following resources useful for providing plain English guidance to staff around AI to support organisational transparency:
Organisations may also want to consider how they can provide transparency more generally on the use of AI in social care, for example through accessible information for the public.
Buying and deploying tools appropriately
Organisations should use this policy paper, alongside wider advice and guidance, to create the conditions in which practitioners feel confident and supported to use AI:
- appropriately
- ethically
- in the best interests of children, families and people who draw on care without diminishing their professional judgement and discretion
Organisations are responsible for creating the conditions for safe, appropriate and lawful use of AI. This should include:
- procuring safe and secure tools
- approving use cases for tools, and innovations
- providing staff with appropriate, engaging and accurate guidance that is updated to reflect emerging AI technologies as they are implemented
- providing training on tools and necessary capabilities
- reviewing AI use and impact
- challenging inappropriate use and escalating where necessary.
This should address both how individuals use AI tools and how organisations decide which tools to develop or procure. For local government, councils can use the Local Government Association (LGA) guide on responsibly buying AI.
When procuring AI tools, organisations should consider and assess tools based on their functionality and security.
Consider asking more grounded questions about what the tool actually does, and whether it is focused on an area where AI can be beneficial.
Grounding
Where possible, organisations may want to consider grounding AI tools with authoritative and relevant source material.
Grounding refers to providing an AI system with trusted information and evidence on which to base its response, helping to reduce the risk of inaccurate, incomplete or fabricated outputs. Several of the case studies outlined in this guidance reflect the benefit of grounding AI tools in authoritative material, such as national statutory or non-statutory guidance and local guidance or policies.
Grounding does not guarantee accuracy. Practitioners should continue to review and verify AI-generated outputs and apply professional judgements.
Case study: North Yorkshire Council ‘Policy Buddy’
Focused on a practical challenge practitioners face every day: finding the right information quickly to inform professional practice, North Yorkshire introduced an AI-powered knowledge assistant ‘Policy Buddy’ to help children’s social care practitioners quickly access and interpret policies, procedures, statutory guidance and legislation.
The tool is built on local and statutory guidance and designed to only provide factual information rather than recommendations, ensuring professional judgement remains with social workers.
Over 800 practitioners were using ‘Policy Buddy’ in the first 3 months of implementation. The evaluation found that:
- practitioners can save 5 to 7 hours each week
- there were increases in practice consistency and professional confidence
- managers reported improved compliance
Find out more about Policy Buddy .
Innovation
Some local authorities and social care organisations are innovating with use cases for AI, either on their own or in collaboration with private or third sector organisations to solve practice-led problems.
DfE and DHSC support responsible innovation in the use of AI, provided that it is accompanied by appropriate safeguards, governance and oversight, and that organisations are mindful of the risks and ethics associated with the use of AI.
Organisations are responsible for ensuring compliance with all relevant legislation and statutory guidance such as:
- data protection legislation (including the UK GDPR and Data Protection Act 2018)
- Care Act 2014
- Children Act 1989
- Children Act 2004
- Working together to safeguard children
Case study: Peterborough City Council’s ‘Hey Geraldine’
In Peterborough Council, staff frequently relied on a highly experienced occupational therapist for advice, creating bottlenecks and limiting access to expertise.
AI Solution
They developed ‘Hey Geraldine’, a bespoke AI assistant trained on the knowledge of an experienced practitioner, to provide instant guidance to staff.
Outcome
‘Hey Geraldine’ answered over 1,200 questions and saved around 15 minutes per interaction, freeing staff to focus on more complex work. This demonstrates how AI can scale access to specialist knowledge and expert advice without increasing staff workload and resource.
The quality of tools developed ‘in-house’ are likely to depend on the quality of the underlying data and how easily it can be accessed and shared across systems. AI will not compensate for incomplete, inconsistent or poorly structured information. When considering innovation, it may be worth considering wider data and information readiness alongside AI readiness.
Organisations should be aware that some AI systems’ capability to generate web apps via prompts is enabling individuals to create apps and tools without prior knowledge or experience.
This may result in the development of solutions to real-world problems, but consideration should be given to how secure these tools might be. Many of these apps are likely to be hosted by the generating product (both OpenAI and Claude currently offer this functionality) and may not be covered by the organisations’ secure environments or enterprise agreements.
Considerations for practitioners and organisations
The following sections set out the main considerations that are relevant to practitioners in using AI to support frontline practice, and organisations in supporting the safe use of AI.
Annex C has a checklist to help practitioners consider these aspects when using AI.
The role of the practitioner
AI tools offer great opportunities to support practitioners and their use should be guided by:
- transparency
- the ‘human in the loop’
- the principle that professional judgement remains with practitioners
- organisational policies and guidance
Final decisions on any aspect of casework decision-making relating to children, young people, families and people who draw on care should be made by the practitioner (alongside partner agencies where relevant). The practitioner is ultimately accountable for their casework.
Where AI tools are being used to support frontline practice, it is important that practitioners’ own thoughts, judgements and assessments are reflected in the final content.
This can be the difference between an accurate assessment of need and an AI-generated output that lacks context and human analysis. AI tools cannot understand body language, tone and inflections or accurately weight information in order of importance (relevant to the context).
Case records, care planning and related documentation should reflect the wider engagement, rather than just a summary of a direct transcript or other AI generated content. Where appropriate, practitioners should allow the child, family, carer or person drawing on care to have sight of the AI transcript or summary produced for them to ensure their voice and needs have been accurately captured.
It is expected that practitioners using AI to generate content will:
- check the content for basic accuracy including spelling, grammar, conversation accuracy
- carefully review the output to prevent bias and hallucinations (generation of outputs that are inaccurate, misleading or incorrect) and ensure it is contextually accurate
- ensure, where appropriate, the output is amended to include additional information or analysis that AI tools cannot pick up on, for example body language, tone, facial expressions, physical observations
- adhere to their organisation’s AI usage policy where relevant
Close attention should be paid to outputs that will be included or used in:
- case records
- documentation or care planning
- legal or statutory settings
Main point for practitioners
AI outputs should always be treated as a first draft, not a final answer. Case records and reports should reflect practitioner thinking, critical reflection and professional judgement.
AI can be more effective when practitioners develop their own critical reflections and judgements before using AI to generate a first draft, so they can interrogate the draft from an informed place. Think first, then draft.
Reviewing drafts should be considered a professional task, rather than a proof-reading exercise. Practitioners should test whether content is accurate, evidenced, balanced and reflective of their own judgement. Practitioners will need to add things AI cannot capture, including observations, context, uncertainty, professional curiosity and the significance of what was not said as well as what was.
Accountability
Local authorities and social care organisations should support their practitioners to understand accountability for the use of AI. This could be through specific training and local guidance. Managers can support this by providing oversight and ensuring that their teams understand how to use AI appropriately and safely.
Clear accountability is essential to ensuring the safe, appropriate and effective use of AI in social care:
- practitioners remain accountable for their work, the decisions they make and should be able to justify their decisions
- managers are accountable for supervision and oversight
- organisations remain accountable for the governance, policy and assurance that enables the safe and appropriate use of AI
Ensuring accuracy
AI may present opportunities to help identify inaccuracies and biases that individuals may be prone to. Used well, it can also help identify inconsistencies and errors and highlight evidence gaps.
AI-generated outputs and content, however, may be inaccurate, incomplete, misleading or overly confident. They reflect the underlying model or the user’s prompt and are prone to confidently presenting incorrect information as fact, sometimes oversimplifying complex situations or downplaying risks.
They will return a credible response unless told otherwise and are, therefore, vulnerable to hallucinations and inaccuracies. A helpful point to remember here is that AI models and tools do not understand the outputs they are generating.
These risks do not remove the potential value of AI but do underline the need for careful review and oversight, ensuring any gaps are filled and inaccuracies corrected before use in a professional context.
Practitioners, supervisors and managers should be mindful that recording inaccuracies can have both short- and long-term consequences. Inaccuracies in the short-term may result in missed opportunities to achieve a positive outcome.
In the long-term, inaccuracies may be reiterated, becoming seen as the ‘truth’. They have enduring implications for the child, young person, family, or person drawing on care and their understanding of events and relationships when accessing and looking back over their records in the future.
Prompting: asking the right questions
Many GenAI tools (those that generate new outputs) rely on prompts to shape the output. Some GenAI tools that develop outputs to meet a specific purpose may also have prompts built in.
Clearer instructions usually result in better outputs to use as a first draft.
Main point for practitioners
Good prompts ask the tool to summarise evidence. Poor prompts ask the tool to make professional judgements.
Well prompted AI tools should support record-writing, not determine needs, risks, eligibility, safeguarding concerns or other statutory decisions.
Practitioners remain responsible for reviewing, editing and approving all outputs before they are added to case records.
Where organisations choose to support the use of GenAI tools, they may find it helpful to provide staff with simple prompting frameworks to improve quality. A frequent example used in a public sector context is the GCSE framework:
- Goal: clearly define what you want the AI to accomplish
- Context: provide background information so the AI understands the purpose and audience
- Source: specify where the AI should retrieve information from
- Expectation: describe the desired format, tone, style, or length of the output
Ensuring safety and data privacy
AI is a data-driven tool. As a sector, we need to be mindful of the data protection and safety risks when procuring, implementing, governing or using AI.
Organisations should define clearly which tools are approved for use, what categories of information may be entered into them, and what safeguards apply.
A general guide is included here, but organisations may have more specific stances.
Publicly accessible AI platforms
Platforms that have not been procured as part of an organisation’s corporate licensing are not secure. No personal, identifiable or case-specific information should be put into these tools. Information entered may be stored, processed or reused by the tool. Many of these AI systems may learn from this data, and any information entered could be replicated elsewhere, creating a data breach.
Corporately licensed tools
For example Microsoft Copilot as part of an organisation’s Microsoft 365 package. These tools are generally more secure but may still have some data protection implications if inputs are used to train the organisation-specific model. Personal information could potentially resurface elsewhere in the organisation through a response to a prompt. If a corporate GenAI assistant is being considered for use in social care, corporate usage policies should clearly state the limitations on use.
Use-case-specific-tools, procured for a specific purpose
For example a social care specific transcription tool. These tools are likely to be more secure, and in many cases are less likely to use data entered into them for training and development.
We recommend that organisations develop a simple way to help practitioners understand what data can be entered into which tool as this will vary based on the procurement and contractual arrangements.
More information on AI data protection principles and risks is available from the National Cyber Security Centre: ChatGPT and LLMs: what’s the risk and Principles for security of Machine learning ML.
Main point for practitioners and organisations
AI tools included in corporate packages may not include the necessary guardrails to guarantee data protection of personal information.
We recommend that organisations provide clarity on whether corporate AI tools have the necessary protections in place to effectively protect personal information in line with:
- data protection regulations
- security requirements
- local policies
If in doubt, check – if you are still in doubt, do not add personal information to the tool.
The regulators’ positions on AI
Ofsted’s position on the use of AI
Ofsted supports innovation and the use of AI where it improves the education and care of children.
Ofsted do not look at AI as a stand-alone part of our inspections, and do not directly evaluate the use of AI, nor of any AI tool. However, inspectors can consider the impact that the use of AI has on the outcomes and experiences of children and learners. That includes both how providers themselves use AI, and how they respond to the use of AI by others – for example parents, children and learners.
Similarly, Ofsted do not consider the use of AI in and of itself in their regulation practice. However, Ofsted’s evaluation of the impact of AI use in regulated settings may be used to help inform any enforcement action.
Ofsted do not require their inspectors to actively look for AI use, and their inspection reports will not refer to AI unless it is crucial to their broader inspection and regulatory decisions. If inspectors find that a provider’s use of, and response to AI has a significant impact on children and learners, they will record it in the same way as any other evidence gathered during inspection.
Ofsted do not expect inspectors to understand how AI software has been designed – any more than, for example, they need to understand the programming of a provider’s firewall software to evaluate their data security arrangements.
How Ofsted looks at AI during inspection and regulation has more information.
Social Work England’s position on AI
Social Work England recognises the potential benefits of AI in social work practice. Where appropriate, social workers can use artificial intelligence (AI) as a tool to support, rather than replace, their professional judgment and values. AI should be used in a way that strengthens ethical, reflective, and relationship-based practice consistent with the core values underpinning Social Work England’s professional standards.
Social workers remain accountable for their decisions and the quality of their practice, regardless of whether AI tools are used, in line with the professional standards. Emerging technology, such as AI can support human judgement, but it should never replace it.
Where AI is used, it is essential that practitioners critically evaluate AI outputs, recognising that these tools can reflect biases and inaccuracies. Social workers are responsible for ensuring that:
- any tools comply with legal and organisational standards
- those they work with understand how the tool is being used
- any use of AI is always aligned with ethical practice
Ongoing digital literacy and critical reflection are essential for ethical AI use in social work. Practitioners should understand both the capabilities and limitations of AI, as well as opportunities to reflect on its impact on practice and people with lived experience.
Care Quality Commission’s (CQC) position on the use of AI
As England’s independent regulator of health and social care, the CQC encourages the use of innovative technologies, including AI, where the technology benefits people and results in more effective and efficient services. However, it acknowledges that AI is not without risk and the use of AI should contribute to safe, equitable and person-centred care.
The CQC:
- recognises that AI is rapidly expanding across health and social care, offering clear benefits, for example efficiency, earlier risk identification, improved monitoring, but also significant risks
- does not regulate specific AI tools but ensures that existing regulations are applied so AI supports safe, effective, equitable and person‑centred care
- expects providers to use AI responsibly, with human oversight, transparency, accountability, and strong governance, ensuring it enhances – not replaces – professional judgement
- will integrate AI into its regulatory approach, support innovation, and continue developing guidance to help the sector use AI safely and effectively
Artificial intelligence in health and social care: CQC’s role, expectations and plans has more information.
Acknowledgements
The Department for Education and Department of Health and Social Care would like to thank the following for input and involvement in developing this guidance:
- The Ada Lovelace Institute
- AI in Care Alliance (AIIA)
- The Association of Directors of Adult Social Services (ADASS)
- Digital Care Hub
- Barnsley Council
- Brent Borough Council
- BRIGHT (the Birmingham AI & Digital Health and Care Group) & University of Birmingham
- British Association of Social Workers (BASW)
- Buckinghamshire Council
- Cambridgeshire County Council
- City of London Council
- The Care Quality Commission (CQC)
- Centre for Excellence for Children’s Care and Protection (CELCIS)
- Coram
- Enfield Council
- Essex County Council
- Hammersmith and Fulham Council
- Hampshire County Council
- Hertfordshire Council
- Hillington Council
- Islington Council
- Kent County Council
- Kingston Council
- Kirklees Council
- Leeds City Council
- Leicestershire County Council
- Lincolnshire County Council
- Luton Council
- The Local Government Association (LGA)
- London Borough of Hammersmith and Fulham
- London Borough of Hillingdon Council
- London Borough of Redbridge Council
- London Borough of Waltham Forest Council
- NVEST
- NHS IG Engagement
- Partners in Care and Health (PCH)
- National Children’s Bureau
- Norfolk County Council
- North Yorkshire Council
- Oldham Council
- Peterborough City Council
- Redcar & Cleveland Borough Council
- Rotherham Council
- Salford Council
- Skills for Care (SfC)
- Slough Borough Council
- South Gloucestershire Council
- Social Work England (SWE)
- The Social Care Institute for Excellence (SCIE)
- South Gloucestershire Council
- Sunbridge Childcare
- Sunderland City Council and Together for Children
- Swindon Council
- Telford & Wrekin Council
- Think Local Act Personal (TLAP)
- Waltham Forest Council
- West Northants Council
- Wiltshire Council
Glossary: types of AI
AI is a broad and open term used to encompass a range of technologies. These technologies often work together in AI tools – for example transcription tools often draw on ‘Ambient Voice Technology’, Natural Language Processing and GenAI.
Most AI tools currently used in social care are based on machine learning models. These models are trained on large amounts of data to recognise patterns and generate outputs. Rather than understanding information in the way humans do, they identify relationships in data and produce responses that are likely to be relevant based on their training.
Some machine learning models, known as deep learning models, are able to learn complex patterns from large datasets, making them capable of speech recognition, image analysis and generative AI.
Common AI use cases
Speech recognition (SR) also known as automatic speech recognition (ASR) and ambient voice technology (AVT)
A technology that converts spoken audio into text (or speech-to-speech). Often referred to as ‘transcription’ and sits alongside GenAI to create summarised transcriptions and populate templates.
Computer vision (CV)
AI that interprets images and video by analysing pixels to detect and classify objects, people and scenes – for example medical imaging, surveillance, robotics.
Natural language processing (NLP)
Techniques that represent language numerically so models can understand, analyse and generate text – for example translation, sentiment, chatbots.
Generative AI (GenAI)
Models trained on large datasets that generate new text, images, audio or video by sampling learned patterns – for example content creation and modelling. Some GenAI tools include the ability to develop ‘agents’ to complete/support with simple tasks.
Agentic AI
Systems or ‘agents’ built on foundation models that can plan, decide and act using tools to achieve goals with minimal human input. These can be task-specific agents, workflow agents or tool-using agents. Different types of agents specialising in completing different tasks might be used together to automate end-to-end processes. It is worth noting this differs from ‘AI Agents’ – these are more likely to be GenAI-centred assistants, like chatbots.
Across the children’s and adult social care sector, we are most readily seeing Speech Recognition, Natural Language Processing and GenAI used. These technologies work together in the case of transcription tools to provide summaries of discussions and meetings.
Annex A: ethics, risks and challenges of AI
A summary of some of the ethical risks and challenges associated with the use of AI, and examples of how these risks might be mitigated.
Professional judgement and human oversight
AI undermines professional integrity, judgement and ethical decision-making.
Reduced critical thinking and over-reliance on AI outputs, leading to poorer practice and recording.
AI tools may be prone to automation bias, meaning outputs are given undue trust or weight because the tool appears authoritative.
Mitigating the risks
AI should be used to support (not replace) professional judgement (used well AI can support practice, learning and help practitioners extend their capability) but the practitioner remains responsible for decisions and final records and should be applying their own analysis and reflections to AI output where appropriate.
Practitioners may want to first ask themselves what their own critical reflection of the situation is, so they can interrogate an AI generated draft from an informed place and to mitigate against authority bias.
AI outputs should be treated as a first draft: review, edit and add your own analysis, reflections, judgements and context before use.
Robust audit arrangements, alongside post-deployment monitoring, on the use of AI will strengthen the organisation’s oversight of AI use and the application of professional judgement.
Over-reliance and deskilling
Practitioners may come to rely too much on AI rather than professional judgement.
Over-reliance on AI may lead to practitioners losing or not fully developing (in the case of new entry practitioners) their analytical, reflective and critical thinking skills.
Mitigating the risk
AI should be used to strengthen practice, not replace it. Good practice in social care relies on professional judgements, and we should be using AI in a way that supports and complements this, rather than replaces it.
Accountability and governance
AI outputs are accepted without challenge, leading to errors being recorded within documentation and practitioners not taking ownership of their decisions.
Mitigating the risk
Organisations should be clear on accountability and governance and ensure that practitioners are clear on where accountability sits. Managers and supervisors should be aware of their responsibility for overseeing AI within practice.
Audit arrangements should factor in and account for the use of AI and support managers and supervisors to identify where practitioners might be inappropriately using AI.
Accuracy and hallucination
AI produces incorrect (but often plausible) or incomplete content, including misunderstandings of transcripts or context.
AI may also infer conclusions from the conversations that are inaccurate or false. These are often presented with confidence by the AI.
AI may also produce a coherent summary that discourages further questioning.
Inaccuracies and hallucinations, left unchecked, misinform decision-making.
Hallucinations are a feature of AI tools, and not a bug. The risk of hallucinations cannot be entirely eliminated but can be mitigated against.
Mitigating the risk
Outputs should be verified against source information and amended before saving; AI text should not be copied into records unchecked by a practitioner.
Any conclusions included within outputs should be reviewed and tested for accuracy. Practitioners may want to consider how prompting can reduce inferences.
Practitioners should feel confident in challenging and testing AI outputs, reducing the risk of AI impeding critical thinking and professional curiosity. Practitioners and the individuals and families they support should be cognisant that AI is not infallible, and outputs should be tested and challenged for inaccuracies. Practitioners should also consider the implications of errors being included in the final document when reviewing.
AI can be prompted to give a confidence rating on its outputs. This may help understand the initial accuracy of outputs.
Bias, unfair treatment, fairness and equity
AI systems may reflect or amplify patterns of bias present in training data, design choices or deployment context. This can often be in ways that are difficult to detect. This creates risk of unfair treatment, distorted analysis or unequal impacts on different groups.
This means AI may exacerbate existing inequalities, including gender and racial bias, and can sometimes influence how GenAI responds.
Research has found some AI tools produce different summaries depending on individual characteristics, such as gender. Examples include men’s needs being described more directly, while women’s needs were sometimes less clear or under-emphasised.
Mitigating the risk
Practitioners and managers should be alert to bias when reviewing AI outputs, as this can affect how needs are understood and acted on, and should not rely on AI alone. Outputs should be checked carefully and anything that seems unclear, biased or unfair should be questioned.
Where AI is used at scale, organisations should test and monitor tools or models for bias, fairness and unequal impact, and take actions where issues are identified.
Organisations should ensure they:
- have robust equalities impact assessments in place (including Public Sector Equalities Duty (PSED) and Human Rights checks)
- follow internal policies
- set clear expectations for practitioners to review outputs for accuracy, fairness and alignment with professional judgement and value
Where concerns arise, feedback loops can help ensure that these issues are addressed and understood by peers.
Transparency, explainability and trust
AI may be used without staff understanding how it works, its limitations, or how data is handled. This can impact their ability to clearly explain how tools are being used and how content will be reviewed.
There are risks around the explainability of AI tools themselves. It is often unclear how AI tools have generated their outputs.
Using AI without explanation (or consent or agreement in certain situations) may damage trust in both the practitioners and the organisation.
Mitigating the risk
Practitioners should be clear about when and why AI is used. Steps should be taken to ensure users understand limitations and remain accountable for outputs.
Practitioners should be able to explain both the tool they are using and how outputs will be used and next steps. They should be supported to develop a basic understanding of how AI works and organisations should provide clear, accessible guidance to help practitioners understand the tools they use.
Organisations should have clear local policies and guidance to outline:
- the way that AI is used
- what staff need to be aware of
- how they can escalate issues and concerns within the local authority
Practitioner confidence will help grow trust in the use of AI in practice. Organisations should support practitioners to develop their AI skills to both confidently use, and explain the use of, AI in their practice.
Privacy, security and data protection
Personal and sensitive information may be unintentionally shared (through prompts or AI outputs) or stored and reused by third‑party tools, risking data protection and security breaches and creating uncertainty about ownership and Intellectual Property.
Mitigating the risk
Only approved tools should be used, and personal or case data should not be entered into publicly available AI tools. Corporately purchased tools should confirm storage, retention and training use. Anyone using AI should follow their organisation’s data protection guidance and organisations should keep terms up to date.
Organisations should have clear local policies and data protection impact assessments (DPIAs) which outline:
- the way that AI is used and data is processed
- what staff need to be aware of
- how they can escalate issues or concerns within the local authority
AI in multi-disciplinary settings
Different organisations may have different policies on how and when AI can be used. This can lead to situations where AI cannot be used.
Mitigating the risk
AI use should be agreed upon prior to meetings including more than one organisation.
Partner organisations may also want to agree what use of AI will be accepted in evidence or assessment and planning documentation. Where AI has been used, any decision-making and professional judgement should remain with the practitioner, and it would remain their responsibility if there were any complaints or court proceedings.
Legal risk
If invalid or incorrect AI outputs are left unchecked and used to produce legal effects on a person within statutory decision-making or to different outcomes than would have otherwise occurred, this could lead to:
- legal repercussions for the local authority or professional responsible
- long-term impacts on the child, family or individual drawing on care
This risk is particularly relevant where there is statutory decision-making.
Mitigating the risk
Practitioners should ensure they read and check AI outputs for accuracy and should not use AI tools to make decisions. They are responsible for the AI-generated content and should be able to justify reasoning.
Organisations may wish to take extra care when supporting newly qualified practitioners, including through learning, development and supervision. Local authorities may want to consider how AI is factored in as part of learning, development and training – including how to both critique AI outputs and use AI to enhance critical thinking.
AI Drift
An AI model’s performance may get worse over time because the world it is working in has changed. For example it keeps using historical data, and does not update to reflect current data, new policies and processes or external events. This means the model’s assumptions and outputs are no longer as accurate.
Mitigating the risk
Practitioners should ensure they continue to check outputs for accuracy and ensure external factors (that the AI may not be aware of) are considered.
Organisations should ensure they are aware of updates and patches to tools and that these are clearly communicated. Organisations should be aware that updates and patches to AI systems and tools are not always unambiguously beneficial. For example, an update may enable a tool to make or advise on a decision where the previous version did not. Organisations should also ensure that any data inputs are as accurate as possible.
Annex B: organisational governance checklist for AI in social care
Purpose
Use this checklist to establish proportionate organisational oversight and assurance across the AI lifecycle. Record your answers, evidence, accountable owner and review date.
This is intended to provide a starting point for organisations to build upon to reflect their own situation, organisational needs and established governance.
Proportionate assurance
Set the level of scrutiny according to the sensitivity of the data, the degree of autonomy and the potential impact on people, practice or statutory functions. Higher risk uses may require enhanced multidisciplinary review, senior approval, stronger auditability and more frequent monitoring. The AI Risk Management Tool Kit provides advice on implementing a simple tiered approach – children’s safeguarding and statutory uses should be considered as high-consequence by default.
AI should not make autonomous decisions about assessment, risk, need, eligibility, safeguarding, care planning or intervention, or write unchecked content into case records.
1. Governance structure and accountability
- A senior responsible owner is accountable for the organisation’s AI approach, with named operational, information governance, security and practice owners.
- A defined governance forum has authority to approve, condition, reject, pause or withdraw AI tools and use cases.
- Terms of reference set decision rights, membership, quorum, reporting arrangements, escalation routes and review frequency.
- An AI policy defines approved, restricted and prohibited uses, required controls, practitioner accountability and exceptions.
- A central register records every tool and use case, supplier, purpose, data used, risk rating, approvals, control owners, incidents and review dates. This should also include clarity on how risks, owners and mitigations or treatments are logged. The AI Risk Management toolkit has more information.
- Organisational, managerial and practitioner responsibilities are clear, including who accepts residual risk.
2. Risk-based approval and assurance
- Each use case is risk assessed in line with a risk appetite statement including a impact or likelihood score before procurement, pilot, deployment or material change.
- Approval evidence covers intended benefit, non-AI alternatives, lawfulness, safeguarding, equality, ethics, privacy, security, accessibility, human oversight and business continuity.
- Data protection, equality impact assessments, PSED (Public Sector Equality Duty) and Human Rights checks are completed where required and linked to the approval record.
- Conditions of approval, acceptance thresholds, residual risks, review dates and stop or withdrawal criteria are documented.
- Material changes to models, prompts, data, integrations, suppliers, terms or purpose trigger reassessment and reapproval.
3. Supplier and contract oversight
- Due diligence tests supplier claims on functionality, limitations, accuracy, bias, security, data use and regulatory compliance.
- Contracts cover data access, location, retention, deletion, reuse, model training, subcontractors, incidents, auditing, performance reporting and exit.
- Suppliers must notify the organisation of material model, service, security or contractual changes.
- Named contract owners monitor service levels, risks, incidents, updates, assurance evidence and corrective actions.
4. Operational oversight and human control
- Approved workflows identify where human review, professional judgement and sign-off are mandatory.
- Outputs entering records or informing practice are appropriately reviewed by the practitioner.
- Managers use supervision and quality assurance to identify over-reliance, boundary breaches or inappropriate use.
- Access controls, usage controls and technical restrictions support policy compliance.
- Staff have clear routes for advice, challenge, incident reporting and protected escalation.
5. Audit and compliance assurance
- An audit plan defines scope, frequency, sampling method, evidence requirements, responsible assurance function and reporting route.
- Audits test whether only approved tools and uses are operating, required human review occurred, data rules were followed and records are accurate and complete.
- Audit sampling is proportionate to risk and includes case records, outputs, user activity, exceptions, complaints and decisions influenced by AI.
- Findings are reported to the governance forum and senior owner, with severity ratings, named actions, deadlines and follow-up verification.
- Repeated, systemic or high-severity findings trigger escalation, increased sampling, control changes, suspension or withdrawal.
- Independent assurance is considered for higher risk uses or where internal assurance is insufficient.
6. Monitoring, incidents and change control
- The governance forum receives a regular oversight report covering usage, accuracy, corrections, bias, unequal impact, accessibility, security, incidents, complaints, benefits and control compliance.
- Performance is reviewed across relevant groups, teams and settings to identify drift, exclusion or unequal outcomes.
- Thresholds define when issues require investigation, corrective action, escalation, suspension or supplier remediation.
- Incidents and near misses are logged, assessed, contained, investigated and reported through existing safeguarding, data, security and corporate risk routes.
- Model drift, updates and changes in law, policy, practice or local context trigger reassessment.
- Each tool has a scheduled review and explicit continuation, modification, suspension and withdrawal decisions.
7. Feedback, transparency and learning
- Practitioners, managers, service users, and partner organisations have accessible routes to provide feedback, challenge use or raise concerns.
- Organisations should consider how children, parents, carers and individuals who draw on care can inform the design and implementation of AI tools where they may affect people drawing on care and support.
- Feedback, complaints, incidents, audit findings and evidence of benefit are logged, categorised, assigned and reviewed together by the governance forum.
- Governance records show what was heard, the decision taken, action owner, timescale and, where no action is taken, the rationale.
- Recording the tool on the Algorithmic Transparency Recording Standards (ATRS) has been considered – where relevant, this is best practice.
- Learning informs approval conditions, controls, tools, contracts, policy, training, audit plans and monitoring thresholds.
- Limitations and accountability are communicated transparently and a safe non-AI route is available where appropriate.
Quick assurance test for senior leaders
Before approval or continued use, senior leaders should be able to answer yes to the following:
- Is there clear ownership, decision authority and documented risk acceptance?
- Is the level of oversight and assurance proportionate to potential impact?
- Are human control, auditability and accountability built into the workflow?
- Do audit and monitoring provide reliable evidence of compliance, safety, fairness and benefit?
- Are findings, incidents and feedback acted on and tracked to closure?
- Are supplier changes and performance subject to effective oversight?
- Can the organisation pause or withdraw the tool quickly?
Annex C checklist: considering ethics, risks and challenges when using AI in social care
Purpose
This checklist is intended to support practitioners in considering the risks and challenges associated with AI in their use in practice. This may be used as a starting point for organisations to develop their own checklists.
Professional judgement, human oversight and over-reliance
Before using or accepting an AI output, practitioners should check that AI is supporting, not replacing their professional judgement. This helps mitigate risks around reduced critical thinking, over-reliance and automation bias where AI outputs may be given undue weight because the tool appears authoritative.
- Have I formed my own professional view before looking at, or relying on, the AI output?
- Have I reviewed, challenged and amended the output where needed?
- Have I added my own analysis, reflections, observations, context and professional judgement?
- Have I considered whether I am giving too much weight to the AI output because it appears confident or authoritative?
- Would I be able to explain and justify the final content without referring to AI as the decision-maker?
Accountability and governance
Accepting AI outputs without challenge, testing or review can lead to errors in case notes, records and documentation and may result in practitioners not fully owning their decisions. Practitioners should remain accountable for their use of AI.
- Am I clear that I remain accountable for any final record, assessment, plan, letter or decision?
- Is my use of AI in line with my organisation’s AI policy or local guidance?
- Have I recorded or evidenced my own reasoning where appropriate?
- Would my manager or supervisor be able to understand how AI supported the work and how I checked it?
- Can I justify my decisions?
Accuracy and hallucination
AI may produce incorrect, incomplete or plausible-sounding content, including misunderstandings of transcripts or context. Outputs should be verified against source information and amended before being saved or used.
- Have I checked the AI output against the original source information?
- Have I checked names, dates, facts, chronology, actions and key details?
- Have I removed or corrected anything that the AI appears to have invented, overstated or misunderstood?
- Have I checked whether the AI has inferred conclusions that are not supported by the evidence?
- Have I considered whether the AI summary is too neat, certain or simplified?
- Have I checked whether important uncertainty, disagreement, risk or context has been lost?
- Have I considered the consequences if an error were entered into a record or used in later work?
- If appropriate, have I asked the AI to identify uncertainty or distinguish between evidence and possible inference?
Bias, fairness and equity
AI systems may also reflect or amplify bias in training data, design choices or deployment context, creating risks of unfair treatment, distorted analysis or unequal impacts. Practitioners and managers should be alert to bias when reviewing AI outputs.
- Have I checked whether the output could unfairly represent, stereotype or disadvantage a person or group?
- Have I considered whether the language used is judgemental, stigmatising, victim-blaming or deficit-focused?
- Have I checked whether the person’s voice, wishes, feelings and lived experience are accurately reflected?
- Have I checked whether different characteristics or circumstances may have influenced the way the output has been framed?
- If something seems unclear, biased or unfair, have I questioned and amended it?
- Where concerns about bias arise, do I know how to raise them through local feedback or governance routes?
Transparency, explainability and trust
Where AI is used by practitioners who are not able to explain (at a high level) what tools are being used, how they work and how outputs will be used, it risks undermining trust between practitioner and those they support. Understanding the basics of how AI tools work, and their limitations, will help practitioners get the most out of the tools.
- Can I explain what information the tool captures, and how outputs will be used?
- Do I understand the limitations of the tool?
- Have I followed local policy on transparency, consent, agreement or approval?
- If someone objects to AI being used, do I know what alternative process should be followed?
Privacy, security and data protection
Only approved tools should be used and personal or case data should not be entered into publicly available AI tools.
- Is this tool approved by my organisation for this task?
- Am I clear what information can and cannot be entered into the tool?
- Have I followed local data protection, information governance and safeguarding requirements?
This focuses on using AI tools in practice and the AI Risk Management Toolkit provides advice and guidance on how to support and test the privacy and security of AI tools.
Multi-disciplinary or multi-agency settings
Different organisations may have different policies on how and when AI can be used, which may create barriers or uncertainty in multi-disciplinary settings. Practitioners should ensure that they are able to use AI in multi-agency settings before doing so.
- Are other organisations involved in this meeting, assessment, record or plan?
- Have I checked whether partner organisations have different policies on AI use?
- Has AI use been agreed before the meeting or activity takes place?
- Have participants been told what the AI tool will do and how outputs will be used?
- Is there a clear alternative if AI use is not agreed?
Quick final check before saving, sharing or using AI-assisted work
Use this as a final ‘pause point’ before adding AI-assisted material to a record, plan, report, email or document.
- I have checked the output against the original information.
- I have corrected inaccuracies, omissions and unsupported conclusions.
- I have checked for bias, unfairness, inappropriate language and loss of the person’s voice.
- I have added my own professional judgement, reflections, context and observations.
- I have complied with organisational policy, data protection requirements and safeguarding duties.
- I have not used AI to make a decision about risk, need, eligibility, safeguarding, care planning or intervention.
- I am satisfied that I can explain and stand behind the final version as my professional work.
Annex D: an introductory guide to AI for social care practitioners
This annex provides a short, practical introduction to AI for social care practitioners. It explains how AI works, how it may support practice, and the safeguards practitioners should apply when using it.
What AI is and how it works
Across government, the definition of AI adopted by OECD countries is used:
“An AI system is a machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. Different AI systems vary in their levels of autonomy and adaptiveness after deployment”
AI systems use patterns in data to generate outputs such as text, summaries, predictions or recommendations. AI tools do not ‘think’; they predict outputs and responses based on the patterns and trends in data that the models are trained on. Some AI systems are able to act autonomously and interact with other AI systems to complete specific tasks.
In social care, common examples include transcription tools, summarisation tools, search tools and GenAI tools that produce text from a prompt.
AI tools are ‘trained’ on large datasets. Depending on the tool, these may include documents, webpages, numerical data and images. The AI is taught to learn and identify patterns to then make predictions and generate outputs based on these patterns. This is why AI tools are prone to bias and hallucinations.
GenAI works from a prompt: the instruction, question or information provided by a user. The tool looks for patterns in the information it has been trained on, and any information provided by the user (such as transcripts or documents), then generates a likely response. This response can appear fluent and human-like, but it does not understand.
AI use often involves:
- input: a prompt or information, for example audio recordings or transcripts, text or documents, images or datasets
- processing: the AI identifies patterns and predicts a likely response – like autocorrect on your phone, it predicts only the next most-likely word in a sequence and repeats this to create a full output
- output: the user receives a draft, summary or answer to check and improve
AI does not:
- think, have feelings or understand people, emotions, relationships or lived experiences
- understand the full context of an event or person’s life or circumstances
It may help organise information, but practitioners should add professional judgement, context, observations, the person’s voice and understanding that effective social work requires.
AI systems can reflect biases that exist in the data used to train them. Practitioners should:
- consider whether outputs could disadvantage, stereotype or unfairly represent individuals or groups
- apply professional judgement to identify and address potential bias
Public AI tools should not be used for casework or personal information. Practitioners should only use tools approved by their organisation for the specific purpose and should follow local rules relating to:
- information governance
- data protection
- consent
- transparency and record-keeping
- statutory and regulatory frameworks
How AI can support social care
AI should support practice, not replace practitioners’ judgements. Its main value is likely to be in:
- reducing administrative burden
- improving drafts and records
- helping practitioners find or summarise information
- supporting reflection
Using AI in social care: a quick guide for practitioners has more information.
Ensuring safety in AI use in social care
Safe AI use depends on keeping people in control. Practitioners do not need to be technical experts, but they should know:
- what the tool is for
- what information can be entered
- how the output will be checked
- where accountability sits
Ultimately, responsibility and accountability sit with the practitioner for all decisions they make.
The professional standards that apply to social work and social care practice, statutory frameworks and local policies and guidance continue to apply, irrespective of whether AI is used or not.
AI cannot replace professional judgements and the critical thinking that is fundamental to social care as a human-centred profession.
Practitioners should be transparent about their use of AI in line with organisational policy and local guidance.
Practitioners should avoid becoming over-reliant on AI and should continue to develop and use their own assessment, analysis, writing and decision-making skills.
The guidance using AI in social care: a quick guide for practitioners has more information on what to do before you use AI and what you should do after using it.
Resources
Government
- AI Cyber Security Code of Practice
- Develop and use data analytics tools in children’s social care
- Adult Social Care Assessments Improvement Toolkit Self-assessment Tool for digitising assessments processes
- The Mitigating ‘Hidden’ AI Risks Toolkit
- The AI Risk Management Toolkit includes risk-tiering, risk appetite and likelihood, impact assessment and clearer treatment of higher-consequence uses such as safeguarding and statutory decision-making
- Artificial Intelligence Playbook for the UK Government
- AI Knowledge Hub for resources on using AI in the public sector
Non-government
- Guidance on AI and data protection
- West Midlands ADASS Co-producing an AI Playbook for adult social care - Digital Care Hub
- Artificial intelligence: meeting the Public Sector Equality Duty (PSED) from the Equality and Human Rights Commission
- Artificial Intelligence (AI) – Judicial Guidance (October 2025) - Courts and Tribunals Judiciary
- AI Unpacked from the local government association for local authority officers and councillors
- Artificial Intelligence from the Centre for excellence for Children’s Care and Protection (CELCIS)
- AI and social care - Digital Care Hub
- Exploring the use of AI transcription tools in social care: Scribe and prejudice? from the Ada Lovelace Institute
- AI and technology resources from The Royal College of Occupation Therapy (RCOT)
- Understanding the emerging use of artificial intelligence in social work education and practice in England from Social Work England and Research in Practice