Business case and benefits
- Finance director
- A · Accountable
- Head of reimbursement
- R · Responsible
- Solutions architect
- C · Consulted
- Technical PM
- C · Consulted
- Engineers
- I · Informed
- IG & DPO
- I · Informed
- Security
- I · Informed
Every branch's NHS payment checked every month, not a sample. Missed service claims caught while they can still be made, margin explained item by item, and a cash forecast finance can plan with.
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.
Each branch is its own NHS contractor and gets its own monthly Schedule of Payments from the NHSBSA: items, fees, Drug Tariff prices, the discount deduction, service payments and adjustments. A group with 450 branches receives 450 of them a month. The reimbursement team checks a sample by hand. Most statements are never compared line by line with what the branch actually dispensed and delivered.
Money leaks in small amounts. A Pharmacy First consultation is recorded in the clinical system but never reaches Manage Your Service, and since June 2025 the window to claim closes at the end of the following month. A branch falls just short of the monthly activity threshold for the fixed payment. Consultations go over the monthly cap and aren't paid. A prescription bundle goes in late. Each one is minor. Across hundreds of branches and twelve months, it's a real sum.
Margin is harder still. Category M prices move every quarter to deliver the agreed retained margin, set at £1.1 billion for 2026/27. Category H, added in March 2026, moves quarterly too. Price concessions are granted month by month, often late in the month: 230 for May 2026 alone. Buyers can't see which lines lose money until weeks later, and finance forecasts branch cash by rule of thumb.
Agree these, and measure the baseline, before anything is built. Example targets are typical, not promises.
The share of recorded but unclaimed consultations flagged before the claim window closes.
ExampleOver 90%
Claims missed, paid late or not paid, by branch and by service.
ExampleDown by half
The share of the gap between expected and actual payment that has a clear cause.
ExampleOver 95% of the value
How long the reimbursement team takes to reconcile all branches.
ExampleFrom weeks to days
Forecast against actual NHS payments, per branch and for the group, three months ahead.
ExampleWithin a few percent for the group
How often a flagged gap turns out not to be one. High means branches will stop reading the flags.
ExampleLow, and reviewed monthly
Cross-sector rules are on the UK regulation map.
Category M prices change each quarter, Category H prices change each quarter on a different cycle, and price concessions apply only to the month they're granted. Prices are held as dated reference data, and every check uses the price that applied in the month of dispensing, including any discount deduction or exemption.
Consultations flow from an approved clinical system, through an API, into Manage Your Service. Since 1 June 2025, a claim must be made by the end of the month after the consultation, and later claims aren't paid unless a proven IT fault caused the delay. Monthly caps apply by band, and the fixed payment depends on minimum activity and on signing up to linked services. So the design compares the clinical system with MYS every day, not at month end.
Each branch declares its prescriptions through MYS by the 5th of the month, or the 6th when a bank holiday falls in the first five days, to get the earlier advance payment. Claims must be accurate: claiming for something not supplied or not delivered is fraud. The AI finds gaps and explains them. It never creates, submits or changes a claim.
Matching services to claims touches patient-level health data. Match on pseudonymised references where possible, leave identifiable data in the source systems, process in UK regions, and fit the platform inside each branch's Data Security and Protection Toolkit assurance.
Wholesaler invoices and credit notes feed both margin and VAT. Keep extracted invoice data linked to the original document, and leave VAT treatment and returns to the existing finance system.
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: statements, invoices, reference pricesMost wholesaler invoices already arrive by EDI. Reuse that feed and read documents only where there's no data. | AWS Transfer Family, S3 | Logic Apps connectors, Blob Storage | An SFTP server or your EDI gateway, Ceph or IONOS S3 Object Storage |
| Reading documentsFor payment schedules and short-line supplier invoices that only come as PDFs. | Amazon Textract | Azure Content Understanding (includes Document Intelligence) | Open-source parsing (e.g. Docling) plus an open-weight vision model |
| Reconciliation storeHolds dated Drug Tariff prices, concessions, claims, payments and purchase prices. | Amazon Redshift | Microsoft Fabric | PostgreSQL, or your existing data warehouse |
| The calculationExpected payment is arithmetic: price, quantity, fees, deductions. That's code with tests, not a language model. | Tested code in the reconciliation store | Tested code in the reconciliation store | Tested code in the reconciliation store |
| ForecastingCash per branch, and the chance a line gets a concession. Amazon Forecast is closed to new customers, so new builds use SageMaker AI. | Amazon SageMaker AI | Azure Machine Learning | MLflow and Kubeflow on Kubernetes |
| Reasoning and explanationsBoth 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 |
| Tariff and service rules (retrieval)Drug Tariff parts and service specifications, so every explanation cites the rule. | 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 |
| WorkflowA daily services check, and a month-end run when the schedules arrive. | AWS Step Functions | Logic Apps or Durable Functions | Temporal, or n8n |
| Human reviewFixes happen in MYS or the clinical system, by people, as they do today. | Tasks in the branch's existing task tool; variances in finance's reporting | Tasks in the branch's existing task tool; variances in finance's reporting | Tasks in the branch's existing task tool; variances in finance's reporting |
| Guardrails | Bedrock Guardrails | Azure AI Content Safety | Open-source guardrails (e.g. NeMo Guardrails, Llama Guard) |
| 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, every explanation and what the person did about it. | CloudWatch, CloudTrail | Azure Monitor, Log Analytics | Prometheus, Grafana, OpenTelemetry |
| UK data residency | 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 dispensing system, the clinical services systems, the finance system and patient-level service records stay where they run today. Matching to claims uses pseudonymised references created on-prem.
Uses cloud AI
Reading payment schedules and PDF invoices, explaining variances against the Drug Tariff, and the cash forecast. The payment calculation itself is ordinary tested code, so it can run on either side.
How they connect
A private link, such as AWS Direct Connect or Azure ExpressRoute, or a VPN, carries pseudonymised extracts out on a schedule and tasks and explanations back. Azure Arc can manage on-prem parts alongside cloud ones on Azure. On the sovereign option, the whole stack runs in your own data centre or with a UK or European provider.
If the link drops
Nothing in the NHS payment process waits for it. Branches claim in MYS as they do today, and the checks catch up when the extracts arrive. The claim window is the deadline that matters, so a link down for more than a day alerts the reimbursement team to check unclaimed services from the clinical system's own reports.
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.
Reconciliation agent
Where a person signs offA named person at the branch or in the reimbursement team makes every claim and every correction, in MYS or the clinical system, as they do today. Finance signs off the cash forecast. The agent has read-only access to every system of record, enforced by the platform's identity and policy controls, not by instructions in a prompt.
Four sources feed the reconciliation, and none was built for it: the dispensing system, the clinical services systems (often more than one supplier across a large estate), the NHSBSA's schedules and MYS reports, and wholesaler invoices. Reuse the EDI feeds you already have for purchasing. Don't build new ones for this.
MYS is a portal. In discovery, confirm what can be downloaded, in what format, and whether automated retrieval is allowed. If it isn't, a named person downloading the month's files into a watched folder is a perfectly good answer.
Nothing in the NHS payment process depends on the AI. If a feed is late or the platform is down, the reconciliation runs later and the branches claim exactly as they do now. Switching it off returns the team to today's sampling.
Take ten branches' last six months, measure what went unclaimed or unexplained, and confirm access to the dispensing system, clinical systems, MYS outputs and invoices.
Gate · must be true to passA baseline value agreed with finance, data access confirmed, DPIA started, and information governance involved.
Thirty branches: back-test on six past months, then run live through two month-ends, with the daily services check sending real tasks.
Gate · must be true to passFlags accurate within agreed limits, variance explained above the target, and finance sign-off.
All branches for the services check and the monthly reconciliation, with the cash forecast going to finance.
Gate · must be true to passMonitoring, support model, reference data ownership and branch training in place.
Add new services as they launch, link purchase prices to buying decisions, and prepare the invoice feed for e-invoicing.
Gate · must be true to passEach addition passes the same accuracy and false-flag checks.
| Activity | Finance director | Head of reimbursement | Solutions architectI can fill this | Technical PMI can fill this | EngineersI can fill this | IG & DPO | SecurityI can fill this |
|---|---|---|---|---|---|---|---|
| Business case and benefits | Accountable | Responsible | Consulted | Consulted | Informed | Informed | Informed |
| Target architecture | Informed | Consulted | Accountable and Responsible | Consulted | Consulted | Consulted | Consulted |
| DPIA and data access | Informed | Consulted | Consulted | Informed | Informed | Accountable and Responsible | Consulted |
| Build and integration | Informed | Consulted | Accountable | Responsible | Responsible | Informed | Consulted |
| Reconciliation accuracy testing | Informed | Accountable | Responsible | Consulted | Responsible | Informed | Informed |
| Go-live decision | Accountable | Responsible | Consulted | Responsible | Informed | Consulted | Consulted |
| Run, monitor and improve | Informed | Accountable | 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 data model, the integration with your dispensing, clinical and finance systems, the forecasts, and the gates. I'd report to your programme board and work alongside your head of reimbursement.
Option B
If your partner or your own team already has an architect, I'd take on the reconciliation engine, the document reading and the agent under their lead, and give them a second senior opinion on the rest.
Goes wrong
The reconciliation finds lost money after the claim window has closed.
How the design avoids it
Check services daily against MYS, and send each branch a task with the deadline on it.
Goes wrong
A language model does the sums, and occasionally gets one wrong.
How the design avoids it
Tested code calculates every expected payment from dated reference data. The model reads documents and explains.
Goes wrong
Concessions arrive late, so every month-end forecast is wrong.
How the design avoids it
Forecast with a range, mark concession lines as provisional, and update when the price is confirmed.
Goes wrong
Branches get hundreds of flags and stop reading them.
How the design avoids it
Rank by deadline and value, group by cause, and measure how often flags are wrong.
Goes wrong
Nobody can get the NHSBSA data out automatically, and it's found out late.
How the design avoids it
Confirm formats and access in discovery. A named person and a watched folder is fine.
Changes already announced or under way that affect this design. More in the UK horizon scan.
Legally available from 29 October 2026, starting branch by branch once the ICB approves each one. New consultation types, a one-off £500 set-up payment and £525 a month for infrastructure, with prescribing consultations counting towards the Pharmacy First thresholds. Add them to the reconciliation before the first claims, not after the first missing payment.
Nineteen products had moved by September 2026, priced from lagged market data. Keep each product's Drug Tariff category as dated reference data, so a move changes the forecast without code changes.
Community Pharmacy England accepted the 2026/27 settlement on condition of joint work with government on alternative contract, funding and reimbursement models. Keep fees, caps and margin rules as configuration, not code.
HMRC published its roadmap on 28 October 2026: a Peppol-based model, a technical consultation in spring 2027 and readiness testing from summer 2028. Wholesaler invoices will arrive as structured data, so design the purchase-price feed to take e-invoices, with document reading as the fallback.
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.