Every application read, checked and summarised before an analyst opens it. Analysts spend their time on judgement, not copying, and a person approves or declines every customer.
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 retail customer applies in the app and uploads a passport and a utility bill. A small-business customer sends a certificate of incorporation, a shareholder list, two directors' IDs and three months of bank statements. Before anyone can decide, an analyst has to read every document, check it against the application, run sanctions and PEP screening, look the company up at Companies House, work out who really owns it, and write a risk assessment that will stand up to an audit.
In a typical mid-sized lender, that work takes 20 to 40 minutes for a simple retail case and two to four hours for a small business with layered ownership. Most of it is gathering and copying, not judgement. Screening alerts are mostly false matches on common names, but each one still needs clearing and a note. Periodic reviews pile up behind new business, so higher-risk customers go longer between checks than the policy says.
The rules have also moved. Amendments to the Money Laundering Regulations took effect on 30 June 2026. The OFSI Consolidated List closed on 28 January 2026, so screening has to run against the FCDO's UK Sanctions List. Companies House now requires directors and people with significant control to verify their identity. Any change to onboarding has to keep the firm's risk-based approach clear, evidenced and owned by named people.
Agree these, and measure the baseline, before anything is built. Example targets are typical, not promises.
How long a customer waits from complete application to an answer, split by retail and small business.
ExampleRetail same day; small business within 3 days
Time moved from gathering and copying to judgement.
ExampleDown by half
False matches closed with a recorded reason, without lowering the screening threshold.
ExampleDoubled, threshold unchanged
How often analysts change the risk rating or the summary. Very high means it isn't helping; near zero can mean drafts are being signed unread.
ExampleStable, and sampled weekly by QA
Reviews completed by the date the customer's risk rating requires.
ExampleAbove 98%, high risk first
Errors found by first- and second-line QA, and by internal audit, in AI-assisted files against manual ones.
ExampleNo worse than manual, then falling
Cross-sector rules are on the UK regulation map.
Customer due diligence, beneficial ownership, enhanced due diligence and ongoing monitoring stay the firm's duty. The 2026 amendments narrowed mandatory country-based EDD to the FATF 'black list' and made EDD apply to unusually complex or unusually large transactions. Rules like these change, so they live in a versioned rule set owned by financial crime, not in a prompt. Data collected for KYC is used only for anti-money laundering purposes.
The analyst's checklist and the agent's tasks follow the firm's JMLSG-based procedures. The agent drafts against them and cites them. Revised Part I paragraphs were sent to HM Treasury for approval in 2026, so the knowledge base is versioned and re-tested when guidance changes.
Financial crime systems and controls must be proportionate and documented, and the MLRO (SMF17) is accountable. The design gives the MLRO a clear view of what the AI does, what it can't do, and how its quality is measured. The AI never files a suspicious activity report and never drafts anything that could tip off a customer.
Since 28 January 2026, the UK Sanctions List is the only source for UK designations, and new designations carry only a Unique ID. Screening runs in the existing engine against that list. The AI helps clear alerts by setting out the evidence; a person confirms or escalates every possible match.
Directors and PSCs must verify their identity, with the transition for existing companies running to mid-November 2026. The register is a useful source, but the firm can't rely on it alone to verify beneficial ownership. Material discrepancies the analyst confirms are reported to Companies House by a person.
Since February 2026, government guidance says certified services on the DVS register can be used for the identity verification step of customer due diligence. Choose a certified provider, record its certification on each file, and keep the rest of CDD, risk assessment and monitoring in-house.
Large organisations need reasonable procedures to stop associates committing fraud for the firm's benefit. Onboarding targets and incentives, and the suppliers who run parts of the process, belong in that fraud risk assessment.
ID documents and selfie checks involve biometric data. That means a DPIA before the pilot, UK processing, clear retention set by the MLR record-keeping rules, and contracts that stop providers training on your data. Since 5 February 2026, solely automated significant decisions need safeguards, and a person makes every decline or exit decision here.
Closing a payment account after a periodic review now generally needs 90 days' notice with reasons the customer can understand, unless an AML or other legal exception applies. Exit decisions and their wording are made by people, with financial crime and legal input.
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: applications and documentsKeep the original upload unchanged, with a hash, for the MLR record-keeping period. | API Gateway, S3 | API Management, Blob Storage | Kong or your existing API gateway; Ceph or IONOS S3 Object Storage |
| Identity verificationBuy this, don't build it. Check the provider is on the DVS register and record its certification on each file. | A certified DVS provider, called through the API layer | A certified DVS provider, called through the API layer | A certified DVS provider, called through the API layer |
| Reading documentsTest on real uploads: phone photos of payslips, scanned share registers, statements from 30 different banks. | Amazon Textract | Azure Content Understanding (includes Document Intelligence) | Open-source parsing (e.g. Docling) plus an open-weight vision model |
| Reasoning and draftingAWS and Azure both offer a choice of models, including Claude. On-prem and sovereign options run open-weight models. | Amazon Bedrock | Microsoft Foundry | Open-weight models served with vLLM on your own GPUs, or IONOS AI Model Hub |
| Procedures knowledge (retrieval)Your KYC procedures, risk methodology and JMLSG-based standards, versioned, so every draft cites the version it used. | 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 |
| ScreeningThe engine stays the control. Confirm it now sources the UK Sanctions List, not the closed OFSI list. | Your existing screening engine | Your existing screening engine | Your existing screening engine |
| Company dataCache responses with the date fetched, so the file shows what the register said at the time. | Companies House API through the API layer | Companies House API through the API layer | Companies House API through the API layer |
| Risk rulesRisk scoring stays deterministic and owned by financial crime. The model explains it; it doesn't set it. | Your existing rules engine, or rules held as versioned config | Your existing rules engine, or rules held as versioned config | Your existing rules engine, or rules held as versioned config |
| Workflow and events | Step Functions, EventBridge, SQS | Logic Apps or Durable Functions, Event Grid, Service Bus | Temporal or n8n; Apache Kafka or RabbitMQ |
| Human reviewAnalysts review where they already work, not in a new screen. | In the analysts' existing KYC case tool | In the analysts' existing KYC case tool | In the analysts' existing KYC case tool |
| GuardrailsBlock customer-facing wording about holds, suspicion or reasons for referral. | Bedrock Guardrails | Foundry 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 source, draft, edit and decision, so a file can be replayed for QA, audit or the FCA. | CloudWatch, CloudTrail | Azure Monitor, Log Analytics | Prometheus, Grafana, OpenTelemetry |
| UK data residencyIONOS AI Model Hub runs in German data centres: EU, not UK, residency. | 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
Core banking, the screening engine and the KYC case tool stay where they are. Many lenders run these on their own servers or with a hosted core banking supplier, and nothing here needs them to move.
Uses cloud AI
Document reading, the procedures index and the two agents can run in AWS or Azure UK regions, or on your own GPUs with open-weight models if identity documents must stay in your estate.
How they connect
A private link (AWS Direct Connect or Azure ExpressRoute) and the API layer. Only the documents and fields the agent needs are sent, and drafts come back as events to the case tool.
If the link drops
Applications keep arriving and queue on your side. Analysts work cases manually, as they did before, and the backlog is drafted 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.
KYC file agent
Where a person signs offAn analyst signs off every customer. EDD cases go to the financial crime team, and any suspicion goes to the MLRO. The agent's permissions are enforced by the platform's identity and policy controls, not by instructions in a prompt.
Periodic review agent
Where a person signs offA person completes every review. Any restriction, exit or change of risk rating is a person's decision, made with financial crime and, for exits, legal input on notice and reasons.
Core banking, screening and the KYC case tool are reached only through the API layer. If you already run an integration platform, the agents use it like any other system does.
Each application becomes an event with a single ID. If the screening engine, the DVS provider or Companies House is slow or down, the case waits safely on the queue and is retried, and a retry never creates a second check or a second account.
External checks are recorded with what was asked, what came back and when. That evidence trail matters as much as the answer.
The AI can be switched off without stopping onboarding. If it's paused, cases go to analysts exactly as they did before.
Map today's retail and small-business journeys, measure the baseline, check access to screening, Companies House and the DVS provider, and pick the first journey.
Gate · must be true to passBaseline agreed, MLRO and DPO involved, DPIA started, data access confirmed.
Retail onboarding in shadow mode: the AI drafts, analysts work as normal, and QA compares the two.
Gate · must be true to passDraft accuracy, alert-clearing quality and QA results within agreed limits, and MLRO sign-off.
Drafts go live for retail, with monitoring, support and a rollback plan.
Gate · must be true to passMonitoring, support, rollback plan and analyst training in place.
Add small-business onboarding, then periodic reviews, one at a time.
Gate · must be true to passEach new journey passes the same pilot gate, with second-line testing.
| Activity | Head of financial crime / MLRO | Product owner | Solutions architectI can fill this | Technical PMI can fill this | EngineersI can fill this | Compliance & DPO | SecurityI can fill this |
|---|---|---|---|---|---|---|---|
| Business case and benefits | Consulted | Accountable and Responsible | Consulted | Consulted | Informed | Consulted | Informed |
| Target architecture | Consulted | Consulted | Accountable and Responsible | Consulted | Consulted | Consulted | Consulted |
| Risk rules and procedures in the knowledge base | Accountable | Consulted | Consulted | Informed | Responsible | Consulted | Informed |
| DPIA and data protection review | Consulted | Consulted | Consulted | Informed | Informed | Accountable and Responsible | Consulted |
| Build and integration | Informed | Consulted | Accountable | Responsible | Responsible | Informed | Consulted |
| Pilot testing and QA comparison | Accountable | Responsible | Consulted | Responsible | Responsible | Consulted | Informed |
| Go-live decision | Accountable | Responsible | Consulted | Responsible | Informed | Consulted | Consulted |
| Run, monitor and improve | Accountable | 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 target architecture, the integrations with core banking, screening, Companies House and the identity provider, and the gates. I'd work closely with your MLRO and report to your programme board.
Option B
If your partner or your own team already has an architect, I'd take on the document AI, the agents and the screening and Companies House integrations under their lead, and give them a second senior opinion on security, data protection and the evidence trail.
Goes wrong
The AI quietly becomes the decision-maker, because analysts approve whatever it drafts.
How the design avoids it
Rules set the rating, people decide, override and edit rates are watched, and QA samples AI-assisted files at a higher rate.
Goes wrong
Screening alerts get 'auto-cleared' to hit the efficiency target.
How the design avoids it
The agent sets out evidence; only a person clears a match. Thresholds stay owned by financial crime and don't move to make the numbers look good.
Goes wrong
The procedures in the knowledge base are last year's, so drafts cite the wrong standard.
How the design avoids it
Versioned documents with a named owner, citations on every draft, and a re-test whenever procedures or JMLSG guidance change.
Goes wrong
A helpful drafted message tips off a customer about a suspicion.
How the design avoids it
No free text to customers. Messages use approved templates only, and guardrails block wording about holds or reasons.
Goes wrong
The demo used clean passports. Real small-business files are 40-page scans of share registers.
How the design avoids it
Pilot on last quarter's real applications, both journeys, not a curated sample.
Changes already announced or under way that affect this design. More in the UK horizon scan.
Record what the register showed and when, and treat unexpected changes in a business customer's officers or PSCs as a review trigger.
Keep procedures versioned in the knowledge base, so approved changes can be loaded, tested and switched on together.
Check your provider's certification and its uplift plan, and plan to check the register automatically when it becomes available.
Keep the decision log, the explanations and the human review records in a form you could show the ICO as well as the FCA.
Expect your IT, screening and identity suppliers to face new duties, and update supplier assurance and contracts to match.
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.
