What problem does it solve?
A large share of civil service time goes into writing: replies to letters and emails from the public and from members of parliament, ministerial correspondence, briefings for ministers and senior officials, submissions, meeting notes and summaries of long documents. Much of it follows known patterns. A correspondence officer finds the current approved lines in a briefing pack, adapts them to the question and routes the draft for clearance; a policy official condenses a stack of papers into a two page brief.
The work is slow and uneven. Finding the right, current line takes time, service level targets for replies are missed when volumes spike, and quality depends on who drafts. Generative AI can produce a good first draft in seconds, but in government a fluent draft that states a superseded policy, gets a case fact wrong or sounds careless in a ministerial letter is a real problem, so the design has to keep drafts grounded in approved sources and every word owned by an official.
How does it work?
- Start from the request. The official pastes or forwards the incoming letter, or selects the documents to be summarised or briefed on.
- Retrieve approved content. The assistant searches the department's approved standard lines, briefing packs, policy documents and, for casework, the relevant case record.
- Draft in house style. It produces a first draft in the right template (reply letter, briefing, submission) with the tone set for the audience, and cites the sources it used.
- Edit and clear. The official checks facts against the sources, edits the draft and sends it through the normal clearance and approval route; nothing is sent automatically.
- Learn from edits. Feedback and edits show which lines are missing or outdated, so owners update the approved content rather than the prompt.
- Audience
- Employee facing
- Autonomy
- Copilot
- Adoption
- Emerging
- Channels
- Internal tools, Email, Microsoft Teams
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Users served | Not pooled | 30 to 2000 | 2 | 2 organization |
| Hours saved | Not pooled | about 3 hours | 1 | 1 organization |
Value drivers: Employee productivity, Speed and cycle time, Customer experience.
Indicative value
A ministry that sends 40,000 replies and briefings a year
EUR 133,333 to EUR 1.2 million
Drafting time released per year
How this is calculated
Formula: drafts * minutesSaved / 60 * hourlyCost. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Replies, briefings and summaries drafted per year drafts, documents per year | 20,000 | 60,000 | Editorial assumption. For scale, the UK Department for Education says its correspondence teams handle about 1,000 external queries a month that need a reply. Source |
| Minutes saved per document after human review minutesSaved, minutes per document | 10 | 20 | Editorial assumption, replace with your own. The Department for Education expects its tool to cut drafting from about 30 minutes to about a minute (a calculation made before user testing), and user research for Cabinet Office Assist found users save about 3 hours a week; review, fact checking and clearance still take time. Source |
| Fully loaded cost of an official's hour hourlyCost, EUR per hour | 40 | 60 | Editorial assumption. Replace with your own staff cost. |
What it leaves out: Time released, not cash saved, unless headcount or contractor spend changes. It leaves out the cost of licences and assurance, the value of faster replies to citizens and the risk cost of errors that slip through review.
Market estimates (analyst estimates, not deployments)
- The Alan Turing Institute estimates that AI could support up to 41% of tasks across the UK public sector, according to the UK government. Landmark government trial shows AI could save civil servants nearly 2 weeks a year (2025)
Who already uses it?
5 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Cabinet Office (Government Communication Service)
United Kingdom · Government and public sector · 2026
Assist is a bespoke generative AI tool for members of the UK government communications profession. It offers pre built prompts for typical communications tasks, drafts first versions, helps with brainstorming and review, and can ground its answers in Government Communications policies and standards through retrieval. It is not to be used for decisions, and users remain responsible for every output. It is in production and was open to about 8,500 members of the profession in July 2026. The transparency record reports that users save around 3 hours a week on average and that 98% of users find it useful in their role, both from user research.
- Hours saved: about 3 hours, per user per week, from user research
"User research shows that Assist users on average save around 3 hours per week by using the tool."
Claimed by: organization
Crown Prosecution Service
United Kingdom · Government and public sector · 2025
The Crown Prosecution Service sends large volumes of letters and emails to people involved in prosecutions. Its Correspondence Drafting Tool pulls key information from the case management system into standard CPS templates and uses a large language model to summarise and pre populate the mandatory content, which authors then refine for the specific case. A built in workflow routes every letter through review and approval by trained staff before it is sent, and the tool plays no part in prosecution decisions. It replaced manual drafting in Word, where copying information by hand could introduce errors, and was in beta with 30 users.
- Users served: 30, beta phase
"Currently in beta phase, being used by small group of users (30 users) the number of users will increase as the tool is implemented."
Claimed by: organization
Department for Education
United Kingdom · Government and public sector · 2025
The Department for Education's correspondence teams paste an incoming query into the Correspondence Drafter, which retrieves the relevant passages from the department's core briefing packs of approved standard lines and drafts a reply in email form, with options to adjust tone and audience. Staff edit and quality assure every draft before sending. The department says the new process has been calculated to be 30 times quicker than searching briefing packs and copying standard lines by hand, which took about 30 minutes per reply, and expects the final phase to cover about 800 of the roughly 1,000 external queries a month that need a reply. The record describes a private beta that had not yet entered user testing, so no measured result is published.
No outcome disclosed.
Department for Science, Innovation and Technology (Incubator for Artificial Intelligence)
United Kingdom · Government and public sector · 2025
Redbox was an assistant built by the UK government's Incubator for Artificial Intelligence (i.AI) that let civil servants chat with large language models, with or without their own documents, for work up to OFFICIAL SENSITIVE, and choose between models. It returned summaries, redrafts or translations of material users provided and did not search the internet. In beta in the Cabinet Office, No 10 and DSIT it had about 2,000 users in February 2025, growing by approximately 150 a week. i.AI later decided on a controlled shutdown: tools such as Microsoft Copilot, whose chat became freely available to many departments, offered similar functionality, and the Cabinet Office moved to an enterprise Gemini option, so the service was run only until the end of 2025. In its lessons learned post i.AI reports that around 70% of interactions were general requests such as drafting emails and brainstorming. The code is open source, and the Department for Business and Trade built its own adapted version.
- Users served: about 2000, February 2025, Cabinet Office, No 10 and DSIT
"As of February 2025 Redbox is used by 2,000 Civil Servants across the Cabinet Office, No10 and DSIT."
Claimed by: organization
Government Digital Service
United Kingdom · Government and public sector · 2024
The Government Digital Service ran a trial of Microsoft 365 Copilot with 20,000 employees across UK government from 30 September to 31 December 2024. Copilot in Teams was the most used application throughout, with adoption peaking at 71%, and one participant named summarising meeting notes among the tasks where it helped. Participants estimated an average saving of 26 minutes a day across all tasks; that figure is self reported, covers every Copilot use and is not recorded as a meeting metric. The report also notes weaker results on complex, nuanced or context heavy work, and concerns that Copilot relied on external sources without built in verification.
- Employee adoption: up to 71%, peak share of trial users using Copilot in Teams, October to December 2024
"Teams was the most popular tool for M365 Copilot and remained dominant throughout the experiment with a maximum adoption of 71%."
Claimed by: organization
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Current approved standard lines and briefing packs, each with an owner and review date
- Templates and style guides for letters, briefings and submissions
- For casework replies, read access to the case record under the right permissions
Systems to integrate
- Correspondence management or case management system
- Document management (for example SharePoint) for approved content
- Email and Microsoft Teams
- Identity and access management for security classifications
Complexity: Low
A drafting assistant over approved documents is quick to build. The hard parts are organisational: keeping standard lines current and owned, agreeing what may be pasted in at which security classification, and fitting drafts into existing clearance workflows.
- 1
Start with high volume correspondence on stable lines
Pick a correspondence stream where replies mostly reuse approved lines, as the Department for Education did with its briefing packs, before moving to ministerial submissions.
- 2
Clean up the source of truth
Put every standard line and briefing pack under an owner with a review date, and remove superseded versions; the assistant is only as current as this library.
- 3
Fit the existing clearance route
Keep drafts inside the current approval workflow, as the Crown Prosecution Service did with its multi layer review, rather than creating a side channel.
- 4
Mark AI content and require citations
Show which text is generated and which source each statement came from, so reviewers can check quickly.
- 5
Measure time and quality together
Track drafting time, clearance rework and complaints or corrections after sending, not only user satisfaction.
Guardrails
- Drafts only; nothing is sent, published or cleared without an official's approval
- Answers grounded in approved lines and case data, with refusal when the library has no answer
- Security classification limits enforced on what can be entered and on the model endpoint used
- Personal data in correspondence masked in logs and prompts
- No political or policy positions beyond the approved lines
KPIs to instrument
- Drafting time per document, before and after
- Share of drafts sent with minor edits only
- Replies within service level targets
- Corrections, complaints or follow ups caused by errors in replies
- Active users as a share of those with access
Human in the loop
The official who owns the reply or briefing edits and approves it, and the normal clearance chain applies. Content owners keep the standard lines current, and a sample of sent replies is reviewed for accuracy and tone.
Common failure modes
- Superseded lines in fluent prose
- The assistant confidently repeats an outdated policy position. Prevent it with owned, dated content and by removing old versions from the index.
- Rubber stamp review
- Busy officials approve drafts without reading them properly. Show sources, sample sent replies and keep accountability with the named official.
- Sensitive data in the wrong tool
- Staff paste classified or personal information into a tool not cleared for it. Provide a sanctioned tool at the right classification so they have no reason to use public ones.
- Tone that damages trust
- Replies that sound generic or evasive to a distressed constituent or a member of parliament. Tune templates by audience and review high profile correspondence closely.
- A general chatbot overtaken by enterprise suites
- i.AI withdrew its Redbox assistant at the end of 2025 after tools such as Microsoft Copilot offered similar functionality, and it reports that most interactions were general drafting and chat. Build where a bespoke tool adds value that a suite does not, such as approved lines, case data and clearance workflows.
What are the risks and rules?
EU AI Act
Minimal risk
An internal drafting assistant that an official reviews is not listed in Annex III. Article 50(4) requires disclosure of AI generated text published to inform the public on matters of public interest, unless it has undergone human review and a person holds editorial responsibility, which this design provides. If the tool is used to evaluate eligibility for public assistance benefits or services rather than to draft, Annex III point 5(a) can apply.
Guidance
- Algorithmic Transparency Recording Standard hub (UK government, Europe). UK departments publish transparency records for drafting tools such as Assist, the Department for Education's Correspondence Drafter and the now withdrawn Redbox.
- Public attitudes to the use of AI in DfT consultations and correspondence (Department for Transport, Europe). Research on how the public views AI support for drafting replies to their correspondence.
Controls to put in place
- Approved content library with named owners and review dates
- Transparency record and staff guidance on permitted use
- Security classification controls on inputs and model endpoints
- Audit log of drafts, edits and the approving official
- Periodic quality sampling of sent correspondence
Frequently asked questions
- How much time does AI drafting save civil servants?
- Reported figures vary by task. User research for Cabinet Office Assist found that users save about 3 hours a week on average, and participants in the UK's cross government Copilot trial estimated 26 minutes a day across all tasks (a self reported figure). The Department for Education calculated, before user testing, that its correspondence tool would be 30 times quicker than a manual process of about 30 minutes.
- Can AI send replies to the public on its own?
- It should not in this design. Every tool on this page drafts for an official who edits and approves, and replies go through the normal clearance route.
- Is this high risk under the EU AI Act?
- Not as designed here. Internal drafting assistance is not listed in Annex III. If the tool is used to evaluate eligibility for essential public assistance benefits and services rather than to draft, Annex III point 5(a) can apply. Public bodies should still tell people how AI is used and keep a named official responsible for what is sent.
How to cite this page
Blits.ai AI Use Case Library, "AI drafting copilot for civil servants for correspondence, briefings and ministerial replies", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/civil-servant-drafting-copilot. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published