Every letter and email read, sorted and routed on the day it arrives, with a sourced draft reply waiting for an official to edit and sign. People decide what the department says and what happens to each case.
At a glance
An illustrative scenario. The numbers are typical examples, not a client case study. The cloud services and UK rules were checked on 9 November 2026.
A department's postbag is mixed. MPs write on behalf of constituents about a specific case. Members of the public write about policy, often as part of an organised campaign with hundreds of near-identical letters. Some letters are really something else: a Freedom of Information request, a subject access request or a formal complaint, each with its own legal clock. A few contain signs that the writer is at risk. All of it arrives by email, web form and scanned post, and someone has to read each item, log it, work out who should answer, and find the right policy line.
Departments aim to answer MP and peer correspondence within 20 working days, and many set tighter targets of their own. The Cabinet Office publishes every department's performance each year. Across government in 2025, 71% of around 182,000 items from MPs and peers were answered within 20 working days. Much of the time goes before anyone writes a word: logging, chasing the right team, and searching shared drives for the latest agreed line and a good past reply.
The reply matters more than its speed. It goes out in the name of a minister or a senior official, so it must reflect current policy, stay politically impartial, treat the person's own case accurately, and be clear to someone who may be distressed or have a disability. A fast reply with an out-of-date line, or one that gets a constituent's case wrong, does more harm than a late one.
Agree these, and measure the baseline, before anything is built. Example targets are typical, not promises.
The share of MP and peer correspondence answered within 20 working days, or the department's own tighter target.
ExampleUp from the published baseline
How long an item waits before the owning team has it, including misrouted items sent back.
ExampleFrom days to same day
FOI, EIR, subject access requests and complaints found in general correspondence and passed on within a day, checked by audit.
ExampleEvery one in the audit sample
How much officers change the drafts. Very high means it isn't helping; near zero can mean drafts are being approved unread.
ExampleStable, and reviewed weekly
Letters chasing a reply, or saying the reply didn't answer the question. A good measure of quality.
ExampleFalling, measured against last year
Timeliness and follow-up rates by correspondence type and language, including Welsh, and where data allows by group.
ExampleNo group worse off
Cross-sector rules are on the UK regulation map.
ATRS records are mandatory for central government departments and for arm's-length bodies that provide public or frontline services or routinely deal with the public. A tool is in scope if it has a significant influence on a decision-making process with public effect, or interacts directly with the public, once it reaches pilot or production. A system that routes casework and drafts replies is likely to be in scope, so plan to publish a record before the pilot ends. For councils the standard is voluntary, but publishing one is a good way to show how the tool is used.
The playbook's ten principles shape the design. Principle 4, meaningful human control at the right stages, means an official approves every reply. Principle 6, the right tool for the job, means checking first what the government's own shared AI tools for civil servants already do before building. Principle 8 means commercial colleagues are involved from the start. In May 2026 the government said new AI guidance would go to the AI Knowledge Hub rather than into further playbook updates, so check both.
Where the tool forms part of a public-facing service, the Service Standard applies, including understanding users and their needs and making the service accessible. Technology spend goes through spend controls and is assessed against the Technology Code of Practice: use open standards, make things secure, and avoid lock-in. A cloud-neutral design with a sovereign option helps here.
The department must have due regard to eliminating discrimination and advancing equality of opportunity. That means an equality impact assessment before the pilot, checking that drafts are as accurate for letters written in plain or non-standard English as for formal ones, offering replies in accessible formats, and monitoring timeliness and quality by group.
Most correspondence is OFFICIAL, and some is marked OFFICIAL-SENSITIVE, such as letters about named individuals' cases. OFFICIAL information can be processed in UK public cloud regions with the right controls. Sensitive handling caveats travel with the item, and the AI never lowers a marking. Anything above OFFICIAL stays out of this system.
Casework often includes special category data, such as health, ethnicity or religion, and criminal offence data. That means a DPIA before the pilot, a clear lawful basis under public task, processing in UK regions, a retention period that matches the department's records policy, and a contract that stops providers training models on the data. The AI drafts; it never takes a decision about the person's case, so automated decision rules are not triggered, and the design keeps it that way. Since 30 September 2026 the regulator has been the Information Commission.
UK departments serving the public in Wales work to Welsh language schemes, and Welsh public bodies to Welsh Language Standards. Either way, a letter written in Welsh is answered in Welsh and no more slowly. The classifier detects Welsh, the item keeps its normal deadline, and a Welsh-speaking officer or the translation team checks every Welsh reply.
Showing AWSAzuresovereign / on-prem services. Switch with the toggle; every diagram and table on this page follows it.
01Channels
02Intake
03Intelligence
04Orchestration
05Systems of record
06People
| Layer | AWS | Azure | Sovereign / on-prem |
|---|---|---|---|
| Intake: email, forms, scanned postKeep the existing post-scanning route. The AI starts from the scanned image, not a new process. | Amazon SES, API Gateway, S3 | Logic Apps connectors, API Management, Blob Storage | Your mail server, Kong, Ceph or IONOS S3 Object Storage |
| Reading lettersTest on real handwritten letters, enclosures and photocopies, not typed samples. | Amazon Textract | Azure Content Understanding (includes Document Intelligence) | Open-source parsing (e.g. Docling) plus an open-weight vision model |
| Classification and draftingBoth clouds offer a choice of models, including Claude. Test Welsh-language drafting separately. | Amazon Bedrock | Microsoft Foundry | Open-weight models served with vLLM on your own GPUs, or IONOS AI Model Hub |
| Lines to take and past replies (retrieval)Only approved, dated lines and signed replies. Each line has an owner and an expiry date, and withdrawn lines are removed the same day. | Bedrock Knowledge Bases | Foundry IQ, built on Azure AI Search | OpenSearch, or PostgreSQL with pgvector |
| Agent runtimeThe original Bedrock Agents closed to new customers on 30 July 2026; new builds use AgentCore. | Amazon Bedrock AgentCore | Foundry Agent Service | An open-source agent framework (e.g. LangGraph) on Kubernetes |
| Workflow and deadline clockWorking days and bank holidays for England and Wales, Scotland and Northern Ireland, from one shared calendar. | AWS Step Functions | Logic Apps or Durable Functions | Temporal, or n8n |
| Middleware and events | Amazon EventBridge, SQS | Event Grid, Service Bus | Apache Kafka or RabbitMQ |
| Link to the correspondence case systemReuse the integration layer you have. Line-of-business case systems are read only, through the same layer. | API Gateway, plus your existing integration platform | API Management, plus your existing integration platform | Kong or your existing API gateway |
| Human reviewOfficers approve where they already work. The private office sign-off route doesn't change. | In the officers' existing correspondence case system | In the officers' existing correspondence case system | In the officers' existing correspondence case system |
| GuardrailsBlocks drafts that state policy not found in an approved line, or make commitments about an individual's case. | Bedrock Guardrails | Foundry guardrails (Azure AI Content Safety) | Open-source guardrails (e.g. NeMo Guardrails, Llama Guard) |
| Identity, keys, data protectionOfficers see only the cases their team can see today. The agents inherit no wider access. | IAM Identity Center, KMS, Macie | Entra ID, Key Vault, Purview | Keycloak or your existing Active Directory; OpenBao or HashiCorp Vault; an HSM |
| Monitoring and auditLog every classification, every source used, every draft, every edit and every approval. These logs are records too. | CloudWatch, CloudTrail | Azure Monitor, Log Analytics | Prometheus, Grafana, OpenTelemetry |
| UK data residencyIONOS AI Model Hub runs in German data centres: EU, not UK, residency. Check your department's policy before using any non-UK option. | eu-west-2 (London) | UK South / UK West | Your own UK data centre, or a UK or EU provider |
On a phone, the table shows the platform you pick. Sovereign / on-prem means open, portable components you can run in your own data centre or with a UK or European provider; IONOS is shown as one example, and others exist. What sovereign cloud means →
Many organisations keep core systems and sensitive data on their own servers. This is how the same design works across both.
Stays on-prem
The correspondence case system, the records store and the line-of-business case systems stay where they are. Some departments and councils still run these in their own or shared government data centres, and nothing in this design needs them to move.
Uses cloud AI
Reading scanned letters, classification, the lines-to-take index and drafting can run in a UK cloud region for OFFICIAL data, or on your own GPUs with open-weight models if policy keeps the casework on your estate.
How they connect
A private link (AWS Direct Connect or Azure ExpressRoute) and the API layer. Only the letter and the case fields needed for a draft are sent, and drafts come back as events onto your queue.
If the link drops
Post and email keep arriving and are logged in the case system as they are today. Officers triage and draft by hand, the deadline clock keeps running in the case system, and the backlog is classified when the link returns.
What each agent may do through its tools, what it may not, and where a person signs off. The limits are enforced by the platform, not the prompt.
Triage agent
Where a person signs offA correspondence officer checks every routing the agent isn't confident about, and every item flagged as a statutory request or a safeguarding concern. The agent's permissions are enforced by the platform's identity and policy controls, not by instructions in a prompt.
Drafting assistant
Where a person signs offAn official edits and approves every draft, and the minister or senior official signs as they do today. Where no approved line exists, the question goes to the policy team; the assistant never fills the gap.
The correspondence case system stays the record. The AI writes classifications and drafts into it through the API layer, never directly into its database, and officers see them in the screens they already use.
Line-of-business case systems, such as benefits, licensing or grants systems, are read only, through the same API layer and with the officer's own permissions. The assistant can describe where a case stands. It never changes it.
Events decouple the AI from the case system. Each new item becomes a message on a queue with a unique ID, so nothing is lost or logged twice if a system is slow. The deadline clock lives in the case system, not in the AI, so pausing the AI never pauses a deadline.
The AI can be switched off without stopping the unit. If it's paused, items flow to officers exactly as they did before.
Map today's correspondence flow, measure the baseline from published and internal data, inventory the lines to take and their owners, and agree what the case system's API allows.
Gate · must be true to passBaseline agreed, line owners named, DPIA and equality impact assessment started, security classification agreed, information governance involved.
One or two correspondence types, such as MP letters on one policy area, in shadow mode: the AI classifies and drafts, officers work as normal, and the two are compared.
Gate · must be true to passRouting accuracy and edit rate within agreed limits, every hidden statutory request caught in the audit sample, ATRS record drafted, SRO sign-off.
Routing and drafts go live for the pilot types, with monitoring, support, the safeguarding route and a rollback plan.
Gate · must be true to passMonitoring, support model, rollback plan, ATRS record published and officer training in place.
Add policy areas and correspondence types one at a time, then reuse the pattern for complaints and FOI.
Gate · must be true to passEach new type passes the same pilot gate, with its lines to take owned and current.
| Activity | Senior responsible owner | Product owner | Solutions architectI can fill this | Technical PMI can fill this | EngineersI can fill this | Policy leads | IG & DPO | SecurityI can fill this |
|---|---|---|---|---|---|---|---|---|
| Business case and benefits | Accountable | Responsible | Consulted | Consulted | Informed | Consulted | Informed | Informed |
| Target architecture and integration | Informed | Consulted | Accountable and Responsible | Consulted | Consulted | Informed | Consulted | Consulted |
| Lines to take: content and ownership | Accountable | Consulted | Informed | Informed | Informed | Responsible | Informed | Informed |
| DPIA, equality impact assessment and ATRS record | Accountable | Responsible | Consulted | Informed | Informed | Consulted | Responsible | Consulted |
| Build and integration | Informed | Consulted | Accountable | Responsible | Responsible | Informed | Informed | Consulted |
| Accuracy testing on real correspondence | Informed | Accountable | Responsible | Consulted | Responsible | Consulted | Consulted | Informed |
| Go-live decision | Accountable | Responsible | Consulted | Responsible | Informed | Consulted | Consulted | Consulted |
| Run, monitor and improve | Informed | Accountable | Consulted | Consulted | Responsible | Consulted | Consulted | Consulted |
R ResponsibleA AccountableC ConsultedI Informed· not involved
Option A
If there's no architect on the programme, I'd own the design end to end: the target architecture, the integration with your correspondence case system, the controls that keep drafts to approved lines, and the gates, with the ATRS record and DPIA built in from discovery. I'd report to your SRO and programme board.
Option B
If your digital team or delivery partner already has an architect, I'd take on the retrieval, the agents and the case-system integration under their lead, and give them a second senior opinion on security, data protection and service assessment readiness.
Goes wrong
A draft quotes a policy line that was withdrawn last month.
How the design avoids it
Only approved, dated lines in the index, each with an owner and an expiry, withdrawn the day policy changes.
Goes wrong
An FOI request or subject access request buried in a long letter misses its legal deadline.
How the design avoids it
The triage agent flags possible statutory requests, an officer confirms, and the deadline is set from the day the letter arrived.
Goes wrong
Officers approve drafts unread, and a reply gets a constituent's case wrong.
How the design avoids it
Drafts that show their sources, an edit-rate measure, and a weekly sample review against the case facts.
Goes wrong
The drafts read well for formal letters but miss the point of short, angry or non-standard ones.
How the design avoids it
Test on the real postbag, including those letters, and report quality by correspondence type as part of the equality impact assessment.
Goes wrong
The tool goes live without an ATRS record or a service assessment, and has to be paused.
How the design avoids it
Plan the record, DPIA and assessment evidence from discovery, not the week before go-live.
Changes already announced or under way that affect this design. More in the UK horizon scan.
Collect service evidence continuously, such as timeliness, edit rates and user research, so the service is ready whichever form assessment takes.
Keep the classification log, the sources behind each draft and the human approval records in a form you could show the regulator.
Build on open standards and your existing API layer, so shared government components can be adopted later without a rebuild.
Expect your IT and AI suppliers to face new duties. Update contracts and supplier assurance as the rules are published.
Record which encryption the platform uses for casework data and keep it easy to change.
Good answers are specific. Vague answers to any of these are worth following up.
Whether you need someone to lead the design or to support the architect you already have, let's talk it through.
