Cabinet Office: Assist
A bespoke Generative AI tool to support members of the government communications profession in their roles in tasks such as brainstorming, creating first draft content and reviewing work.
Tier 1 Information
1 - Name
Assist
2 - Description
Assist is a secure and accessible AI tool designed to enhance the efficiency and effectiveness of government communicators. Powered by generative AI, this bespoke conversational tool enables users to brainstorm ideas, create initial drafts, and review their work. By using Assist, Government communicators also benefit from Assist’s ability to draw upon data from Government Communication Service (GCS) policies and best-practice communication standards alongside wider knowledge provided by the underlying large language model.
Assist provides users with more than 50 communications-specific ‘pre-built prompts’, which reflect the typical tasks a government communicator might need to do on a daily basis. These prompts span across all GCS disciplines, ensuring the tool is tailored to every government communicator use case.
3 - Website URL
GCS Blog with further information on Assist:
4 - Contact email
Tier 2 - Owner and Responsibility
1.1 - Organisation or department
Cabinet Office, Government Communications
1.2 - Team
Applied Data and Insight
1.3 - Senior responsible owner
Head of Applied Innovation
1.4 - External supplier involvement
Yes
1.4.1 - External supplier
*Westminster Heritage Limited *Numswork *Mondo Rosso *Happysoft
1.4.2 - Companies House Number
*Westminster Heritage Limited - 08865055 *Numswork - 08508587 *Mondo Rosso - 05321530 *Happysoft - SC695960
1.4.3 - External supplier role
All suppliers have contributed through standard software development practices, building to Cabinet Office objectives and standards.
1.4.4 - Procurement procedure type
Open
1.4.5 - Data access terms
External supplier access was restricted to project-only data on a need to know basis. External suppliers had no access to personal data.
Tier 2 - Description and Rationale
2.1 - Detailed description
Assist is accessed via ‘Connect’, a Laravel application serving communications professionals across government. All frontend code for Assist is held in the code base for Connect. All user requests to Assist are sent to the Assist API, which is a separate web application. The Assist API is responsible for generating responses to the user queries as well as managing Assist data. When generating a response to the user query, the Assist API accesses a Large Language Model (LLM) hosted on AWS Bedrock. All Assist infrastructure is securely deployed within an Amazon Web Services (AWS) environment. All data processing stays within AWS. The original trainer of the LLM (Anthropic) never receives Assist data. This is guaranteed by the ‘escrow account’ setup provided by AWS. Amazon does not receive any user data (prompt inputs, outputs or analytics) for any third-party uses, including training models.
Users will query Assist using plain English. Assist will use its pre-trained model and its RAG architecture to aid it in adding context to the query. It will also consider the system prompt and then call upon the API Large language model to synthesise and respond to the user in plain English. Users also have access to a range of pre-written prompts available for commonly completed tasks, split by communications discipline. This allows users to quickly apply a prompt to gain an output without prompt engineering.
The product uses the following system architecture: *Hybrid-Reasoning: Uses enhanced reasoning capabilities for complex tasks, while also providing transparency into its step-by-step thought process before it delivers its final answer. *Retrieval-Augmented Generation (RAG): Users have the option to turn this feature on/off. If on, Assist will reference an authoritative knowledge base outside of its training data sources (detailed in 2.4.2.8) before generating a response - thus improving the quality of its outputs. *Pre-trained Transformers: Assist will analyse text based queries (prompts) and predict the best possible response based on its pre-trained understanding of language.
2.2 - Scope
Assist is designed to be used by public sector communicators who are members of the government communications profession in a range of tasks, such as producing first-draft content based on GCS frameworks, guidance and best practice.
The tool is not for use outside of the government communications profession. Assist is not designed for automated decision-making.
2.3 - Benefit
Government communications operates in a tight fiscal environment and therefore need to make the most efficient and productive use of available resources.
The impact of Assist was monitored through several data collection techniques including surveys, focus groups, 1-1 interviews and usability sessions. 1365 users completed the initial data collection form, before 334 and 278 users completed the private beta and public beta surveys respectively. A further 2605 pieces of feedback have been submitted and reviewed through the Assist application itself. The team continuously collect and analyse user feedback, data and analytics to inform all aspects of the service from design to communications and support. User research shows that Assist users on average save around 3 hours per week by using the tool. Assist also ensures that outputs automatically adhere to GCS best practice and standards. 98% of users state that Assist is useful in helping them to perform their role. A further benefit of Assist is the upskillling it brings for communicators in their AI capabilities, which will bring further dividends as they utilise AI tools elsewhere in their day-to-day work.
2.4 - Previous process
Assist is not a decision-making tool. Rather it is designed to improve the productivity and effectiveness of government communications.
2.5 - Alternatives considered
During the alpha phase of Assist, we compared different LLM models and their benefits. We compared OpenAI GPT 3.5, GPT 4 and Anthropic Claude 2.1, all available through APIs, using user feedback and a “colosseum style” comparison to drive the decision. The priority was to decide the best model for output quality against cost.
Tier 2 - Decision making Process
3.1 - Process integration
As outlined in the GCS Generative AI Policy and the Assist terms and conditions, Assist is not to be used as a decision-making tool, and is designed to be a ‘co-pilot’ for GCS members in a range of tasks as part of their role.
Assist will provide a response to a user based on the query they have provided and guided by supporting documentation and the system prompt. The user is responsible for the onward use of Assist’s outputs and ensuring it’s used responsibly.
3.2 - Provided information
Assist will provide users with a text based response. It cannot produce responses in any other format. Users will also receive details of any documents it has cited as part of its RAG system architecture.
3.3 - Frequency and scale of usage
Assist is open to be used by all members of the government communications profession. The tool is currently accessible to c. 4,250 members (April 2025).
3.4 - Human decisions and review
As part of the Assist terms of service, users are responsible for the appropriate onward use of the outputs of Assist. If users are unhappy with Assist’s output, they can continue the conversation to reprompt the tool until they receive high quality outputs.
Assist is not to be used as a automated decision making tool.
3.5 - Required training
Before gaining access to Assist, GCS members must complete an ‘Assist onboarding course’. This course outlines the principles of AI and LLMs before reiterating the guidance of using AI generated content in government.
Responsible use has been the guiding principle of Assist’s development, informed by the GCS Framework for Ethical Innovation, the GCS Generative AI Policy and the GCS Innovating with Impact strategy (all publicly available on the GCS website). Before using the tool, all GCS members must complete a bespoke ‘AI for Communicators’ training course, designed to upskill and inform them of the safe use of Assist and AI in the workplace.
3.6 - Appeals and review
Not applicable
Tier 2 - Tool Specification
4.1.1 - System architecture
Assist source code repository: https://github.com/Government-Communication-Service/assist_service
4.1.2 - Phase
Production
4.1.3 - Maintenance
Assist is maintained by the Government Communications team in the Cabinet Office in an Agile based approach. It is regularly reviewed as part of our ongoing evaluation of the service, in line with our development roadmap. We regularly review the latest abilities of AI models to ensure we choose the best model for capabilities and our user needs. As part of this approach, we regularly review our RAG library for the latest versions of guidance and frameworks, as well as user research to understand any in-demand documentation not already included.
4.1.4 - Models
Assist currently uses Claude 3.7 Sonnet (April 2025), which is an LLM model provided by Anthropic PBC. As part of the Retrieval Augmented Generation (see 2.2.1 for more info) the retrieval aspect algorithmically chooses the correct information to cite based on the user query.
Tier 2 - Model Specification
4.2.1 - Model name
Anthropic Claude
4.2.2 - Model version
3.7 Sonnet
4.2.3 - Model task
Claude 3.7 Sonnet is designed to assist with a wide range of tasks including answering questions, analysing information, creative writing, coding, problem-solving, and engaging in conversation across various topics.
4.2.4 - Model input
Users will input a query in natural language text, related to government communications tasks. Assist has capabilities for processing documents and supports PDF, DOCX, TXT, HTML, PPT/PPTX, ODT formats.
4.2.5 - Model output
Assist will output a response in natural language text with markdown formatting, responding to the users with varied lengths or formats based on the complexity of the query.
4.2.6 - Model architecture
Claude 3.7 Sonnet is a large language model (LLM) with a transformer-based architecture, featuring enhanced reasoning capabilities. Further details can be found in Claude 3.7 Sonnet System Card
Claude 3.7 Sonnet is a state-of-the-art LLM.
Assist uses a GCS specific system prompt designed to optimise outputs for government communications.
4.2.7 - Model performance
Specific details can be found in Claude 3.7 Sonnet System Card.
Claude 3.7 Sonnet achieves a high Elo on lmarena.ai (a human-preference leaderboard, delivered as a randomised control trial)
Data Deduplication: Claude 3.7 Sonnet achieves a median F1 score of 70.2%
4.2.8 - Datasets
Details can be found here: https://www.anthropic.com/legal/model-training-notice
GCS has had no involvement in training this model.
Assist has access to an additional corpus of data as part of it’s RAG algorithm if the user elects to include it in their query. The following documents are included as part of the RAG library:
*Modern Communications Operating Model 3.0: https://gcs.civilservice.gov.uk/modern-communications-operating-model-3-0/ *Accessible by default: https://gcs.civilservice.gov.uk/guidance/accessible-communications/making-your-digital-content-accessible/accessible-by-default/ *Inclusive Communications Template: https://gcs.civilservice.gov.uk/publications/inclusive-communications-template/#audience-insight *British Sign Language Act and guidance: https://gcs.civilservice.gov.uk/publications/british-sign-language-act/ *The Principles of Behaviour Change Communications (COM-B): https://gcs.civilservice.gov.uk/publications/the-principles-of-behaviour-change-communications/ *GCS Evaluation Cycle: https://gcs.civilservice.gov.uk/publications/gcs-evaluation-cycle/
4.2.9 - Dataset purposes
Details can be found here: https://www.anthropic.com/legal/model-training-notice
Tier 2 - Data Specification
4.3.1 - Source data name
Assist draws upon data included as part of Anthropic’s training data-set, more information can be found here: https://www.anthropic.com/legal/model-training-notice
GCS has had no involvement in training this model.
Assist can also pull upon specific data included as part of the RAG framework, this data includes GCS frameworks and best practices which are commonly applied across a range of communications tasks.
4.3.2 - Data modality
Text
4.3.3 - Data description
The data included as part of the RAG framework contains GCS best practices, guidance and frameworks from the GCS Modern Communications Operating Mode 3.0. All communicators must apply these standards when creating first drafts. This data is available to ensure Assist fully meets their needs across a range of communications tasks.
4.3.4 - Data quantities
The following documents are included as part of the RAG library:
*Modern Communications Operating Model 3.0: https://gcs.civilservice.gov.uk/modern-communications-operating-model-3-0/ *Accessible by default: https://gcs.civilservice.gov.uk/guidance/accessible-communications/making-your-digital-content-accessible/accessible-by-default/ *Inclusive Communications Template: https://gcs.civilservice.gov.uk/publications/inclusive-communications-template/#audience-insight *British Sign Language Act and guidance: https://gcs.civilservice.gov.uk/publications/british-sign-language-act/ *The Principles of Behaviour Change Communications (COM-B): https://gcs.civilservice.gov.uk/publications/the-principles-of-behaviour-change-communications/ *GCS Evaluation Cycle: https://gcs.civilservice.gov.uk/publications/gcs-evaluation-cycle/
4.3.5 - Sensitive attributes
There are no sensitive attributes contained within the dataset.
4.3.6 - Data completeness and representativeness
The documents have been chunked into Assist to ensure it draws upon the most relevant information for the user. No data has been excluded as part of this process.
4.3.7 - Source data URL
See 2.4.3.4 for URLs
4.3.8 - Data collection
Not applicable
4.3.9 - Data cleaning
Documents in the central RAG system are pre-processed manually. This pre-processing is performed by transforming documents into a set of ‘chunks’ (semantically-complete sections). The pre-processing is performed by expert humans in the government communication profession.
Documents in the File Upload system are pre-processed automatically using the Python library unstructured
. This library is configured to create chunks based on headings, with a minimum size constraint to prevent overly-brief chunks.
4.3.10 - Data sharing agreements
Not applicable
4.3.11 - Data access and storage
The documents used as part of the RAG framework are available to all GCS members and most are publicly available (excluding the Inclusive Communications Template and British Sign Language Act and Guidance). These have been included as part of the RAG framework to allow easy application for the user. There is no sensitive attributes included as part of this dataset.
Tier 2 - Risks, Mitigations and Impact Assessments
5.1 - Impact assessment
A data protection impact assessment has been completed for Assist on 24/06/2024.
Equalities impact is reviewed as part of our ongoing evaluation of the product.
Assist Privacy notice: https://www.gov.uk/government/publications/government-communication-service-assist-privacy-notice/privacy-notice-for-government-communications-service-gcs-assist
5.2 - Risks and mitigations
We consistently monitor and/or mitigate various risks as part of our ongoing risk framework (available on GOV.UK, once published). The risks are split into the following categories: *Quality Assurance: Risks arising due to (people using) average quality or inaccurate outputs. *Task-Tool Mismatch: Risks arising due to the use of tools for purposes for which they weren’t designed or at which it doesn’t perform well. *Perceptions, Emotions and/or Signalling: Risks arising due to emotional responses induced by AI roll out, people’s perceptions and attitudes about AI or the signals sent by UK Government’s adoption/use of AI. *Workflow and/or Organisational Challenges: Risks arising from the work required to embed AI in Government or changes to people’s ways of working. *Ethics: Risks arising from violations or threats to ethical standards and norms or legal rights (e.g. Equality Act 2010). *Human Connection and Technological Overreliance: Risks arising from reductions in, or removal of, humans from roles or functions or the over reliance on technical solutions for complex problem. *Technical Risks: Cyber security and accessibility risks, currently being monitored and mitigated through regular CHECK penetration testing and Digital Accessibility Audits.
In addition to this, bespoke mandatory onboarding training for users in advance of getting access to Assist introduces users to the known risks around AI and how they can materialise in outputs (such as bias and hallucinations). This training also teaches the importance of keeping the human-in-the-loop and how to check AI outputs for accuracy, reliability and bias.