First drafts of submission documents built from approved source data in days, not weeks, with every statement traced to its source. Authors and regulatory leads review and sign off, and the AI never submits anything.
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.
In a typical mid-sized UK pharma or biotech company, every submission document is built by hand. A medical writer or regulatory scientist opens the clinical study reports, stability reports, specifications, validation reports and safety summaries, copies the numbers across, writes the narrative around them, and checks every cross-reference. A single Module 2 summary can draw on hundreds of source documents, each with its own version history.
The work comes in peaks. A new application, a cluster of variations after a manufacturing site change, a set of questions from the MHRA with the clock running, a safety change that has to flow into every SmPC, leaflet and label. Under that pressure small errors creep in: a number transcribed wrongly, a superseded stability table, an SmPC statement that no longer matches the clinical summary. Each one costs a validation query, a request for information or a delay. Since July 2026, applicants using the MHRA's International Recognition Procedure get a single chance to respond to validation correction requests.
The documents are regulated records. The marketing authorisation holder is responsible for every word sent to the MHRA, whoever or whatever drafted it, and every statement has to trace back to approved data. A tool that drafts quickly but can't show where each statement came from just moves the effort from writing to checking. The aim here is the opposite: drafts that are quicker to check than to write.
Agree these, and measure the baseline, before anything is built. Example targets are typical, not promises.
From request to a draft ready for author review, by document type.
ExampleA typical variation package from 3 weeks to 3 days
Share of statements and numbers in a draft with a working link to an approved source.
ExampleEvery statement, checked on an audit sample
Statements that reach approval without a source behind them. Found by QA's sample review, not self-reported.
ExampleNone
How much authors change the drafts. Very high means it isn't helping; near zero can mean drafts are being approved unread.
ExampleStable, and reviewed monthly
Validation queries and requests for information traced to a transcription or consistency error.
ExampleFalling, measured against last year
Differences between SmPC, leaflet, labels and the dossier found after submission.
ExampleNone in the audit sample
Cross-sector rules are on the UK regulation map.
The marketing authorisation holder is responsible for the accuracy of everything submitted. Submissions go in eCTD format through the MHRA's submissions portal, compiled and sent by the publishing team. The AI drafts content only: it never compiles a sequence for dispatch, submits, or replies to the MHRA. Deciding whether a change needs a variation, and of which type, stays with regulatory affairs.
Drafting follows the CTD structure, with a template and a source map for each section. Module 1 is UK-specific; Modules 2 to 5 are shared with other markets, so one approved source library can serve several submissions. The quality section is being revised (M4Q(R2)), so keep templates configurable.
Records must be attributable, legible, contemporaneous, original and accurate, and also complete, consistent, enduring and available. Every prompt, source passage, draft, edit and approval is kept with an audit trail. The signed document in the validated document management system stays the record; the AI's output is a draft until a person approves it.
Where drafts feed GxP records, treat the tool as a computerised system: a written intended use, a risk assessment, validation in proportion to that risk, and change control for the model, prompts, templates and source library. A model version change is a change, re-tested against a fixed set of past documents before it goes live.
This matters for UK firms that also submit in the EU. The applicant stays responsible for AI used in preparing its submissions, and AI-drafted product information needs a quality review that checks it is factually and grammatically correct. The EU AI Act doesn't list regulatory drafting as a high-risk use.
Clinical study reports and safety narratives can contain pseudonymised patient data, and documents name investigators and staff. That means a DPIA before the pilot, processing in UK regions, a contract that stops providers training their models on your data, and keeping patient-level listings out of the retrieval index unless a document genuinely needs them.
Showing AWSAzuresovereign / on-prem services. Switch with the toggle; every diagram and table on this page follows it.
01Sources
02Intake
03Intelligence
04Orchestration
05Systems of record
06People
| Layer | AWS | Azure | Sovereign / on-prem |
|---|---|---|---|
| Intake: events and filesApproved documents arrive as events from the document management system, not as uploads by hand. | API Gateway, S3 | API Management, Blob Storage | Kong or your existing API gateway; Ceph or IONOS S3 Object Storage |
| Reading source documentsBedrock Data Automation processes London requests in London (checked 9 November 2026; re-check at build time). Textract plus an in-region model is the alternative. Test on scanned legacy reports and complex tables, not just born-digital PDFs. | Amazon Textract, or Bedrock Data Automation | Azure Content Understanding (includes Document Intelligence) | Open-source parsing (e.g. Docling) plus an open-weight vision model |
| Drafting and reasoningBoth clouds offer a choice of models, including Claude. Check which run in-region in London or UK South, and pin the model version: validation depends on it. | Amazon Bedrock | Microsoft Foundry | Open-weight models served with vLLM on your own GPUs, or IONOS AI Model Hub |
| Source knowledge (retrieval)Only approved, effective versions, with document ID, version and status in the metadata. Superseded documents are removed from the index, not just ranked lower. | Bedrock Knowledge Bases | Foundry IQ, built on Azure AI Search | OpenSearch, or PostgreSQL with pgvector |
| Numbers and tablesNumbers are copied by code and checked against the source. The model writes the words around them, never the figures. | AWS Lambda reading structured sources | Azure Functions reading structured sources | Python services on Kubernetes |
| 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 | 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 document management and RIMAgree in discovery what the suppliers allow: read access to approved documents and writing drafts back as new draft versions. | Through the supplier's published API, plus your existing integration platform if you have one | Through the supplier's published API, plus your existing integration platform if you have one | Through the supplier's published API, plus your existing integration platform if you have one |
| Human reviewAuthors and approvers work where they already do. Electronic signatures stay in the validated system. | In your existing document management system's review and approval workflow | In your existing document management system's review and approval workflow | In your existing document management system's review and approval workflow |
| GuardrailsBlocks drafts that contain statements with no source link, and flags claims beyond what the approved data supports. | Bedrock Guardrails | Foundry guardrails (Azure AI Content Safety) | Open-source guardrails (e.g. NeMo Guardrails, Llama Guard) |
| Identity, keys, data protectionThe assistant sees only the products and documents the requesting author is allowed to see. | 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 prompt, source passage, draft, edit and approval, with model and document versions. This is your validation and inspection evidence. | CloudWatch, CloudTrail | Azure Monitor, Log Analytics | Prometheus, Grafana, OpenTelemetry |
| UK data residencyIONOS AI Model Hub runs in German data centres: EU, not UK, residency. Unpublished dossier data is commercially sensitive, so many firms will want UK-only processing. | 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 validated document management system, the RIM system, the publishing tool and the source data systems stay where they are, whether that is your own servers or a validated hosted service. Nothing in this design moves them or re-validates them.
Uses cloud AI
Document reading, the source index, drafting and consistency checks run in a UK cloud region, or on your own GPUs with open-weight models if unpublished dossier data must not leave your estate.
How they connect
A private link (AWS Direct Connect or Azure ExpressRoute) and the API layer. The AI reads approved documents only, and writes drafts back as new draft versions through the supplier's API. It has no route to the publishing tool or the MHRA portal.
If the link drops
Authors write and review exactly as they did before. Nothing in the document management system changes, and queued drafting requests run 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.
Drafting assistant
Where a person signs offThe author checks every statement against its source and edits the draft. The regulatory lead approves, and QA reviews where the SOP requires, with electronic signatures in the validated document management system. The agent's permissions are enforced by the platform's identity and policy controls, not by instructions in a prompt.
Consistency checker
Where a person signs offThe document owner decides on every finding and records why. Variation classification stays with regulatory affairs.
The document management system stays the record. The AI reaches it only through the supplier's published API and an API layer, never directly. It reads documents whose status is approved and effective, and writes drafts back as new draft versions with the AI's involvement recorded in the document history.
Events keep the index current. When a source document is approved, it is read and indexed. When it is superseded or withdrawn, it is removed straight away. Every event carries an ID, so a retry never indexes a document twice, and the index can be rebuilt from the document management system at any time.
Publishing stays separate. The AI produces content, not eCTD sequences. The publishing team compiles, validates and submits through the MHRA's portal exactly as now. If the AI is switched off, authors carry on with the same templates and the same process.
Pick the first document type (CMC variations or SmPC updates are good starts), inventory its sources, write the intended use and risk assessment, and confirm API access to the document management system.
Gate · must be true to passBaseline measured, intended use and risk assessment agreed by QA, DPIA started, supplier access confirmed.
One document type in shadow mode: the AI drafts past and live documents, authors work as normal, and the two are compared. Validation testing runs alongside.
Gate · must be true to passTraceability and accuracy within agreed limits, no unsupported statements in QA's sample, validation report approved.
Validated release for the first document type, with monitoring, support and a rollback plan.
Gate · must be true to passSOP effective, authors trained, monitoring and periodic review in place.
Add document types one at a time: responses to questions, labelling updates, then Module 2 summaries.
Gate · must be true to passEach new document type passes the same pilot and validation gate.
| Activity | Head of regulatory affairs | Product owner | Solutions architectI can fill this | Technical PMI can fill this | EngineersI can fill this | QA & validation | DPO | SecurityI can fill this |
|---|---|---|---|---|---|---|---|---|
| Business case and benefits | Accountable | Responsible | Consulted | Consulted | Informed | Consulted | Informed | Informed |
| Intended use and risk assessment | Accountable | Responsible | Consulted | Informed | Consulted | Responsible | Consulted | Consulted |
| Target architecture and supplier integration | Informed | Consulted | Accountable and Responsible | Consulted | Consulted | Consulted | Consulted | Consulted |
| Validation plan and report | Consulted | Consulted | Consulted | Responsible | Responsible | Accountable | Informed | Consulted |
| Build and integration | Informed | Consulted | Accountable | Responsible | Responsible | Consulted | Informed | Consulted |
| Accuracy testing on real documents | Consulted | Accountable | Responsible | Consulted | Responsible | Responsible | Informed | Informed |
| Go-live decision | Accountable | Responsible | Consulted | Responsible | Informed | Responsible | 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 source library and retrieval, the integration with your document management and RIM systems, the audit trail and the gates, working with your QA and validation team so the evidence is built as we go. I'd report to your programme board and your head of regulatory affairs.
Option B
If your partner or your own team already has an architect, I'd take on the retrieval, drafting and consistency checks and the document management integration under their lead, and give them a second senior opinion on security, data residency and the validation approach.
Goes wrong
The demo used a clean, public dossier. Real sources are scanned legacy reports with complex tables.
How the design avoids it
Pilot on your own documents for one product, including the awkward ones.
Goes wrong
Drafts quote a superseded stability table or specification.
How the design avoids it
Index only approved, effective versions, and remove superseded ones on the event, not overnight.
Goes wrong
The model rounds or invents a number.
How the design avoids it
Numbers copied by code and checked against the source; the model never writes figures.
Goes wrong
The provider updates the model and the validated state quietly lapses.
How the design avoids it
Pin model versions, treat any change as a change, and re-run the fixed test set before it goes live.
Goes wrong
Authors approve drafts without checking, so checking moves from people to nobody.
How the design avoids it
Source links on every statement, an edit-rate measure, and QA's monthly sample review.
Goes wrong
QA and validation see it a month before go-live.
How the design avoids it
Intended use and risk assessment agreed with QA at the discovery gate.
Changes already announced or under way that affect this design. More in the UK horizon scan.
Keep templates and source maps configurable per section, so Module 2.3 and Module 3 can change without a rebuild.
Record now where AI was used in each document and how it was checked, so you can answer whatever the plan asks.
Keep authored content separate from the publishing format, so the switch is a publishing change, not an authoring one.
Keep generative drafting away from GMP decisions, with a person approving every document.
Author SmPC and leaflet text as structured, reusable content now, so a labelling change flows to every format.
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.
