Stronger checks on every online prescription, inconsistencies and risky patterns put in front of the prescriber, and evidence ready for any inspection. AI strengthens the safeguards; it never replaces them.
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.
The group's online pharmacy prescribes and supplies prescription-only medicines at a distance, mostly through its own pharmacist independent prescribers. Weight-loss medicines are the largest share, as injections and, since 2026, as licensed tablets too. Demand moves quickly, and prescribers can face hundreds of reviews in a shift, each built on a questionnaire, some uploaded evidence and an order history spread across several systems.
The rules have tightened. The GPhC's guidance for pharmacies providing services at a distance, updated in February 2025, says higher-risk medicines, including weight-loss medicines, must not be prescribed on a questionnaire alone. Weight, height or BMI must be independently verified, by video, in person, from clinical records or by contacting the patient's GP. The GPhC's April 2026 review found weight management in 77 inspection reports over two years, and 27% of 1,307 concerns were about prescribing practice, including missing BMI verification. Advertising is watched just as closely: the ASA and MHRA have acted against hundreds of ads for weight-loss prescription-only medicines.
Most people answer honestly. Some don't: a weight rounded up, someone else's photo, a second account at a different address, a re-order before the last pen could have been used, a child using a parent's details, or someone with an eating disorder who knows what to say. One prescriber looking at one order can't see most of this. And when an inspector or a concern asks what was checked for a particular supply, assembling the evidence takes days.
Agree these, and measure the baseline, before anything is built. Example targets are typical, not promises.
Share of weight-loss supplies with independent verification of weight, height or BMI recorded, and the method used.
Example100%, checked every day
Share of orders where every flag was in front of the prescriber before they decided.
ExampleEvery order
How often a flag changed or informed the prescriber's decision. Too low means noise, and noise gets ignored.
ExampleReviewed weekly
People ordering through more than one account, or for someone else, compared with a manual review sample.
ExampleMeasured against the sample
Approval, decline and referral rates before and after go-live. A rise in approvals is a warning sign, not a win.
ExampleNo unexplained change
Time to assemble everything held about one supply for an inspector or a concern.
ExampleFrom days to under an hour
Cross-sector rules are on the UK regulation map.
Higher-risk medicines can't be prescribed on a questionnaire alone, and weight-loss prescribing needs independent verification of weight, height or BMI plus a route for two-way communication. The AI never counts as that verification. It records which verification happened, flags where it didn't, and blocks the order reaching a prescriber as complete until it has.
The pharmacy owner and superintendent pharmacist are responsible for the service, including any third-party prescribers. Weight management is an inspection focus. The evidence pack is designed for an inspector's questions, and the safeguards are owned by the superintendent, not by the AI team.
Online prescribing by doctors and other listed professionals needs CQC registration for treatment of disease, disorder or injury. Pharmacist prescribers working from the registered pharmacy fall under the GPhC. Know which regulator covers each prescriber, and keep evidence that suits both.
Prescription-only medicines can't be advertised to the public, and avoiding the product name doesn't make an ad compliant. The AI never writes marketing, and an assistant checks proposed copy against the rules before a person in compliance approves it.
Health data, profiling and linked-account matching need a DPIA, a clear lawful basis and condition, collecting only what's needed, UK processing and a retention period. Significant decisions based on health data still need meaningful human involvement, so no order is declined or delayed by the AI alone.
A tool that recommends whether to prescribe would be heading towards a medical device. This one shows the prescriber verification status, inconsistencies and patterns, with the evidence, and makes no treatment recommendation. Keep that intended purpose written down.
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: orders, questionnaires, filesQuestionnaires are fixed, versioned and owned by the superintendent. No AI on the customer side. | API Gateway, S3 | API Management, Blob Storage | Kong or your existing API gateway; Ceph or IONOS S3 Object Storage |
| IdentityBuy this, don't build it, whichever platform you choose. Certification is under the UK trust framework, on a statutory footing since the Data (Use and Access) Act 2025. | A certified provider from the government's digital verification services register | A certified provider from the government's digital verification services register | A certified provider from the government's digital verification services register |
| Reading uploaded evidenceReads what was uploaded. It can't tell whether a document or photo is genuine or current. | Amazon Textract | Azure Content Understanding (includes Document Intelligence) | Open-source parsing (e.g. Docling) plus an open-weight vision model |
| Linked accountsTune it carefully. Families share addresses and cards; that isn't misuse. | AWS Entity Resolution | Matching model in Azure Machine Learning | A matching model built with MLflow and Kubeflow on Kubernetes |
| Order-pattern modelTrained on cases reviewed by senior prescribers, not on past approvals. Amazon Fraud Detector closed to new customers in November 2025. | Amazon SageMaker AI | Azure Machine Learning | MLflow and Kubeflow on Kubernetes |
| Summaries and reasoningBoth clouds offer a choice of models, including Claude and others. | Amazon Bedrock | Microsoft Foundry | Open-weight models served with vLLM on your own GPUs, or IONOS AI Model Hub |
| Protocols and guidance (retrieval)The group's prescribing protocols, product licences and GPhC guidance, so flags cite the rule they relate to. | 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 |
| WorkflowRequired checks are rules in the workflow, never model outputs. | AWS Step Functions | Logic Apps or Durable Functions | Temporal, or n8n |
| Guardrails | Bedrock Guardrails | Azure AI Content Safety | Open-source guardrails (e.g. NeMo Guardrails, Llama Guard) |
| Platform | Managed AWS services | Managed Azure services | Kubernetes on-prem (e.g. OpenShift, Rancher), or IONOS Managed Kubernetes |
| Identity, keys, data protection | 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 flag, who saw it, and what the prescriber decided. | CloudWatch, CloudTrail | Azure Monitor, Log Analytics | Prometheus, Grafana, OpenTelemetry |
| UK data residencyThis is health data. If the group also handles NHS data on the same platform, some NHS buyers will require UK residency, so an EU-hosted option may not fit. | eu-west-2 (London) | UK South / UK West | Your own UK data centre, or a UK/EU provider (IONOS AI Model Hub runs in Germany) |
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 online pharmacy platform, the prescribing and dispensing systems, and the patient records stay where they are today. Required-checks rules live in the workflow, close to those systems, so they never depend on a model being available.
Uses cloud AI
Reading uploaded evidence, linked-account matching, the order-pattern model and the flag summaries run in the chosen platform: a UK cloud region, or the group's own GPUs for the sovereign option. The identity check is a call to a certified provider, whichever platform hosts the rest.
How they connect
Through an API gateway, over a private link such as AWS Direct Connect or Azure ExpressRoute, or inside the group's own data centre for the sovereign option. Flags come back as events and are attached to the order in the prescriber's existing queue.
If the link drops
Orders still reach prescribers, and every required check is still enforced by the workflow. The prescriber simply works without the extra flags and the order is marked as not yet pattern-checked. Nothing is approved faster because the AI is unavailable.
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.
Safeguards agent
Where a person signs offThe prescriber decides every order, and only the person or certified service that carried out a check can mark it verified. Safeguarding concerns go to the safeguarding lead. The agent's permissions are enforced by the platform's identity and policy controls, not by instructions in a prompt.
Marketing check
Where a person signs offA named person in compliance approves every piece of marketing for the online pharmacy.
The online pharmacy platform stays the record of the order and the consultation. The identity provider, the evidence store and the AI connect to it through an API layer, and flags appear in the prescriber's existing queue, not in a new screen.
The branch network is a safeguard in its own right. Where weight can't be verified remotely, the workflow offers a check at a nearby branch, recorded by the branch team. The AI can book and chase that check; it can't stand in for it.
If the AI is down, nothing gets faster. Orders still reach prescribers, with the required checks enforced by the workflow, just without the extra flags. There is no fast path that depends on a model.
Map today's checks for each medicine, measure verification completeness, agree which checks are fixed rules, and choose a certified identity provider.
Gate · must be true to passRequired checks signed off by the superintendent, baseline measured, DPIA started, safeguarding lead involved.
Weight-loss medicines only, in shadow mode: flags go to a senior reviewing prescriber, not the live queue, and are compared with a manual review sample.
Gate · must be true to passFlags useful, no decision drift, safeguarding route tested, and superintendent sign-off.
Flags and evidence packs go live for all prescribers, with monitoring, support and a rollback plan.
Gate · must be true to passMonitoring, decision-drift reporting, rollback plan and prescriber training in place.
Extend to other higher-risk medicines, such as those liable to misuse, then add the marketing check.
Gate · must be true to passEach new medicine passes the same pilot gate.
| Activity | Superintendent pharmacist | Lead prescriber | Product owner | Solutions architectI can fill this | Technical PMI can fill this | EngineersI can fill this | Compliance & DPO | SecurityI can fill this |
|---|---|---|---|---|---|---|---|---|
| Required checks for each medicine | Accountable | Responsible | Consulted | Consulted | Informed | Informed | Consulted | Informed |
| Target architecture | Consulted | Consulted | Consulted | Accountable and Responsible | Consulted | Consulted | Consulted | Consulted |
| DPIA and lawful basis | Consulted | Informed | Consulted | Consulted | Informed | Informed | Accountable and Responsible | Consulted |
| Build and integration | Informed | Consulted | Consulted | Accountable | Responsible | Responsible | Informed | Consulted |
| Testing flags on real cases | Accountable | Responsible | Consulted | Consulted | Responsible | Responsible | Consulted | Informed |
| Go-live decision | Accountable | Consulted | Responsible | Consulted | Responsible | Informed | Consulted | Consulted |
| Run, monitor and review decision drift | Accountable | Responsible | Responsible | Consulted | Consulted | Responsible | 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 required-checks workflow, the identity and evidence integrations, the matching and pattern models, and the gates. I'd report to your programme board and your superintendent pharmacist, and I'd say plainly where a check has to stay with a person.
Option B
If your partner or your own team already has an architect, I'd take on the linked-account matching, the safeguards agent and the evidence packs under their lead, and give them a second senior opinion on the rest, especially security and data protection.
Goes wrong
The safeguards tool becomes a speed tool: a low risk score fast-tracks approval.
How the design avoids it
No score ever removes a required check. Required checks are workflow rules, and approval rates are watched for drift.
Goes wrong
A chatbot helps customers fill in the questionnaire and ends up coaching the right answers.
How the design avoids it
No AI on the customer side of the questionnaire. Questions are fixed, reviewed and versioned.
Goes wrong
Someone treats an uploaded photo or document as proof because the AI read it.
How the design avoids it
The AI reads; people and certified services verify. Every record says which is which.
Goes wrong
The model learns from past approvals, including past poor practice.
How the design avoids it
Train on cases reviewed by senior prescribers and known misuse, not on what was approved.
Goes wrong
Too many flags, so prescribers stop reading them.
How the design avoids it
Measure flag usefulness weekly and retire rules that never change a decision.
Changes already announced or under way that affect this design. More in the UK horizon scan.
Set required checks and pattern rules per product, not per drug class, so a new product gets its own rules before launch.
Keep evidence structured by medicine and verification method, so a new standard is a report, not a project.
Plan for NHS and private weight services side by side, so the same person can't be supplied by both without the prescriber knowing.
Choose an identity provider certified against version 1.0, and check its certification automatically once the register allows it.
Keep the intended purpose narrow and documented: flags and evidence for a prescriber, never a treatment recommendation.
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.
