The Crown Prosecution Service: Correspondence Drafting Tool
The 'Correspondence Drafting Tool' aids CPS teams by utilising key information from the CPS case management system to streamline the correspondence drafting process according to CPS guidelines.
Tier 1 Information
Name
Correspondence Drafting Tool
Description
The Crown Prosecution Service (CPS) handles extensive correspondence with various stakeholders involved in prosecution proceedings. This includes sending out numerous emails and letters each day. To improve this process, a’ Correspondence Drafting Tool’ has been developed. This tool utilises algorithms and large language models (LLMs) to automatically summarise and pre-populate mandatory information, which is then refined by CPS staff based on the specific context of each communication. The automated workflow ensures that all letters are reviewed, amended, and approved by the appropriate parties, resulting in consistent and high-quality outputs.
Website URL
NA - This is an internal tool
Contact email
CPSCyberSecurityTeam@cps.gov.uk
Tier 2 - Owner and Responsibility
1.1 - Organisation or department
Crown Prosecution Service
1.2 - Team
Digital Information Directorate
1.3 - Senior responsible owner
Deputy Director of Digital Change
1.4 - External supplier involvement
Yes
1.4.1 - External supplier
NTT Data UK Limited
1.4.2 - Companies House Number
3085018
1.4.3 - External supplier role
NTT Data UK Limited, as CPS digital delivery partner, played the pivotal role in the development of the ‘Correspondence Drafting Tool’. NTT Data provided the required skills and expertise in AI model Integration, natural language processing and data management, as well as End to End delivery and integration within existing CPS case working journey and tools. NTT Data together with CPS formed a cross functional delivery team consisting of User Researcher, Designer, Business Analysts, AI experts, AI engineers to jointly develop a ‘Correspondence Drafting Tool’.
1.4.4 - Procurement procedure type
Pre-established framework run as a competitive procedure
1.4.5 - Data access terms
Access to any internal data is governed by contractual agreements between the CPS and NTT Data and is compliant with the relevant data protection regulations. Suppliers working with CPS are subject to the relevant security clearance checks. NTT Data employees have access to secure CPS devices with their own login credentials.
Tier 2 - Description and Rationale
2.1 - Detailed description
The ‘Correspondence Drafting Tool’ enables users to draft correspondence such as letters and emails based on specified templates. The tool pre-populates templates with a draft content in a text editor for the user to compose the correspondence as per it’s intended purpose.
After the first draft is generated by the tool, it is reviewed and finalised by a human. There is a multi-layer review and approval process applied in accordance with CPS guidelines to ensure letter quality and factual correctness. The context and specific details of each case are reviewed and approved by trained CPS Staff, ensuring accuracy and quality before sending out the correspondence.
2.2 - Scope
The purpose of the tool is to aid CPS teams (authors, reviewers, and approvers) with a digital tool for accurate and efficient correspondence drafting process, driven by the goal of improving the quality of correspondence. The tool is designed for external correspondence such as emails and letters in accordance with CPS standard correspondence templates. The tool does not make decisions in the prosecution process, it is used to aid correspondence drafting.
2.3 - Benefit
Key benefits of the ‘Correspondence Drafting Tool’:
- Improved quality of correspondence
- Optimising the user journey for CPS teams by providing a digital tool with built-in templates, mandatory information fields, and an automated workflow to manage the review and approval process.
- Increasing quality and reducing time spent on manual process, increasing efficiency.
2.4 - Previous process
Correspondence was drafted manually using MS Word, which could have led to errors through the manual transfer of information.
2.5 - Alternatives considered
The team also looked at alternatives, such as programmatic solutions, however, due to the contextual nature of the correspondence, it was determined that Large Language Models are more suitable for correspondence drafting. Multiple LLM models were considered, however, OpenAI service was identified as the most appropriate meeting CPS security guidelines and standards.
Tier 2 - Decision making Process
3.1 - Process integration
The tool will not impact on the decision-making process. The tool aids users to draft better correspondence.
3.2 - Provided information
The tool does not impact decision making nor provide output which is used for decision making.
3.3 - Frequency and scale of usage
Currently in beta phase, being used by small group of users (30 users) the number of users will increase as the tool is implemented.
3.4 - Human decisions and review
The tool is used to draft the initial version of correspondence this is then adjusted via full and detailed human review and approval process ensuring accuracy and quality. Further details in the detailed description.
3.5 - Required training
End-users have been engaged throughout iterative development process and extensive user research. In addition, end-user training guide which includes roles & responsibilities has been developed to aid users in understanding how the tool and its features work alongside their role in the review and approval process. Extended beta phase is being used to further train the relevant users in the usage of the ‘Correspondence Drafting Tool’. Operations, maintenance and oversight of the tool is completed.
3.6 - Appeals and review
Not used for decision making. Existing CPS policies and procedures for feedback and complaints apply to the correspondence issued by CPS.
Tier 2 - Tool Specification
4.1.1 - System architecture
4.1.2 - Phase
Beta/Pilot
4.1.3 - Maintenance
The tool is being iteratively developed. We release a new version to production regularly. Users can report feedback through established feedback channel. Maintenance and oversight of the tool is ongoing.
4.1.4 - Models
Microsoft OpenAI GPT family models available in the UK South Standard deployment. At the time of development GPT4. In addition, rule based logic is used to map the relevant templates and prompts to generate content.
Tier 2 - Model Specification
4.2.1 - Model name
GPT
4.2.2 - Model version
GPT 4 - version 0125
Different GPT family models may be used as the product is iterated, and new models become available.
4.2.3 - Model task
GPT-4 version 0125 selects the correct template to use when a logic-based approach is incapable of doing so.
GPT-4 version 0125 simplifies complex legal language into clear, accessible text.
4.2.4 - Model input
Inputs to the model come from the CPS Case Management System (CMS). Staff review the drafts, consider content, add additional information and ensuring quality and accuracy.
4.2.5 - Model output
Clear, jargon-free correspondence that is accurate and of high quality.
4.2.6 - Model architecture
Large Language Model - GPT 4 https://cdn.openai.com/papers/gpt-4.pdf
Specifically, Microsoft OpenAI - GPT 4 version 0125 https://learn.microsoft.com/en-us/azure/ai-services/openai/concepts/models?tabs=global-standard%2Cstandard-chat-completions#gpt-4
4.2.7 - Model performance
Model performance for this specific use case was measured both quantitative and qualitative. For the quantitative measurement, the model performance was compared against the current manual processes where users generate comms manually in a word document template. From a qualitative perspective, the model performance was improved through iterative user feedback and testing. Users provided the feedback in regard to what good looks like for the content and appropriate language. During the quality assurance process, anonymised test information and correspondence were created to simulate the model output.
4.2.8 - Datasets
The LLM software is out of the box software, CPS data was not used to develop the LLM.
4.2.9 - Dataset purposes
Out of the box model was used (GPT4); no further development, training, fine tuning of the model has been done at this stage.
Tier 2 - Data Specification
4.3.1 - Source data name
Structured data from case details has been used from the CPS case management system.
4.3.2 - Data modality
Text
4.3.3 - Data description
Content from the CPS Case Management System
4.3.4 - Data quantities
Currently 230 Total requests, 444.19K Total token count, 424.35K Prompt token count, 19.85K Completion token count, all on gpt-4o mode
4.3.5 - Sensitive attributes
The tool will access only the minimum necessary personal data and sensitive personal information required for its operation. This information will be sourced from the CPS case management system.
4.3.6 - Data completeness and representative-ness
The tool has access to the required information to perform its function. The data is drawn from relevant information so is both complete and appropriate to the recipient
4.3.7 - Source data URL
N/A
4.3.8 - Data collection
All data used by the tool has been collected for law enforcement purposes. The LLM was not developed by CPS, we have developed prompts. No permament retention of data has been implemented. Once the correspondence is finalised, data is deleted from the tool.
4.3.9 - Data cleaning
Data is taken directly from source - N/A
4.3.10 - Data sharing agreements
4.3.11 - Data access and storage
Data is contained within the CPS environment. The data is stored within the tool until the correspondence is finalised, then the information is deleted. Access to any data within the tool is in line with the access and permissions control framework within the CPS Case Management System.
Tier 2 - Risks, Mitigations and Impact Assessments
5.1 - Impact assessment
The Data Protection Impact Assessment has been approved, and penetration testing is complete. We have undertaken ethics and harms workshops to evaluate the tool’s potential benefits and harms. Additionally, an AI Ethics and Security assessment for AI-based tools has been completed and approved. Our CPS digital delivery governance aligns with GDS standards as part of the business operations.
5.2 - Risks and mitigations
Main risks are: 1. Hallucination or errors introduced by the AI tool - The risk of hallucination has been mitigated by human accuracy and quality checks. The LLM chooses relevant templates and topics, and then generates a draft template, which is reviewed by human users. Comprehensive training and guidance is provided to all users, ensuring that only authorised personnel can access the tool. Additionally, a thorough usage policy is in place, and regular audits and management checks are conducted to ensure quality and compliance with processes.
2.Security - The integration of Generative AI in CPS operations requires a comprehensive approach to cybersecurity to maintain the confidentiality, integrity, and availability of sensitive information and to mitigate risks associated with AI technologies. This has been addressed by selecting the right technological solution and ensuring security compliance. IT health check by external supplier was completed in January 2025 and any identified vulnerabilities resolved before the beta phase.