Millions of transactions matched every day, the breaks investigated before anyone opens a spreadsheet, and a person approving every posting, with the audit evidence built as the work is done.
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.
Every day, the group's own ledgers have to agree with the outside world. That means bank statements from several banks, settlement reports from Faster Payments, Bacs, CHAPS and the card schemes, files from payment processors and partner firms, and the customer balances in the core banking and e-money platforms. Most items match on amount, date and reference. The rest don't, for ordinary reasons: a reference cut short by a scheme's field length, fees netted from a settlement, a payment split across two days, a return, a duplicate, a currency conversion.
In a typical challenger group, deterministic rules match 92–95% of items. That still leaves thousands of breaks a day. Analysts work them in spreadsheets: export, look up the ledger, find the scheme report, search old emails for a similar case, then raise a journal. Aged breaks hide real problems, such as a missed settlement, a mis-posted customer payment or a shortfall in the money held for customers.
Those shortfalls now carry more weight. Since 7 May 2026, payment and e-money firms must reconcile the funds they safeguard every reconciliation day, send the FCA a monthly safeguarding return, and have an annual safeguarding audit. The investment arm already reconciles client money daily under CASS 7. Senior managers are personally accountable for these controls. Speeding up matching only helps if the evidence of who decided what is better than before, not worse.
Agree these, and measure the baseline, before anything is built. Example targets are typical, not promises.
Items matched by rules, plus suggestions accepted without change.
ExampleFrom 94% to 98–99%
Unresolved items older than five business days, by account and value.
ExampleDown by three quarters
Time from a break appearing to a resolution being approved.
ExampleDown by half
How often analysts accept the model's match or the agent's proposed fix. Very high means it may not be checked; very low means it isn't helping.
ExampleStable, and reviewed weekly
Days with an unresolved shortfall, and how fast each was corrected.
ExampleEvery shortfall resolved the same day
Issues raised by internal audit and the safeguarding and CASS auditors on reconciliation controls.
ExampleFewer each year, none repeated
Cross-sector rules are on the UK regulation map.
The e-money subsidiary must do internal and external safeguarding reconciliations every reconciliation day, keep a resolution pack, send a monthly safeguarding return and, above £100,000 of safeguarded funds, have an annual audit. Non-standard reconciliation methods must be documented. The AI speeds up matching and investigation, but a person signs off each day's reconciliation and decides every top-up or withdrawal of excess. Name one accountable person for safeguarding.
The investment arm runs an internal client money reconciliation each business day, using a standard method or a documented non-standard one, plus regular external reconciliations, and records each one. Deposits the bank holds as a bank aren't client money. The design keeps the reconciliation method itself in rules and code that audit can read; the AI only helps clear the breaks. Its records sit in the resolution pack index so they can be found quickly in a wind-down.
A named senior manager is accountable for finance and for client assets, and must be able to show they took reasonable steps. Every AI-assisted step is visible to them: what was suggested, who approved it, and how often suggestions were overridden. The regime is being simplified (Phase 1 rules from 24 April 2026, Phase 2 out for consultation), but individual accountability stays.
Records must be orderly and complete enough for the regulator to check compliance. Every match, suggestion, approval and posting is stored with its reason and kept in write-once storage for the agreed retention period.
Since 1 May 2025, CHAPS payments between financial institutions carry LEIs and purpose codes, and property payments carry purpose codes. Richer, structured data makes matching easier, so the matching model uses these fields where they exist. The next set of changes, including the end of fully unstructured addresses, was deferred from November 2026.
Statement lines and payment references contain personal data. That means a DPIA, processing in UK regions, a clear retention period, and a contract that stops providers training their models on your data. Reconciliation doesn't make decisions about individuals, but a mis-posted customer payment does affect one, so fixes are always approved by a person.
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: files and messagesBank connectivity (SWIFT, host-to-host, bank APIs) usually stays as it is. Take a copy of each file as it lands. | AWS Transfer Family, Amazon S3 | Blob Storage with SFTP, Logic Apps connectors | Your existing managed file transfer; Ceph or IONOS S3 Object Storage |
| Parse and normaliseTurn camt.053, MT940, scheme reports and partner CSVs into one transaction format, keeping the original line for evidence. | AWS Glue | Data Factory in Microsoft Fabric | Apache Spark on Kubernetes |
| Matching data storeHolds both sides of every match, its status and its history. | Amazon Aurora PostgreSQL | Azure Database for PostgreSQL | PostgreSQL |
| Deterministic matchingKeep the rules auditors already trust. AI starts where the rules stop. | In your existing reconciliation platform, or rules in your own code | In your existing reconciliation platform, or rules in your own code | In your existing reconciliation platform, or rules in your own code |
| Match suggestions (ML)A ranking model trained on your own past matches, giving a confidence and a reason for each suggestion. No language model is needed for this. | Amazon SageMaker AI | Azure Machine Learning | MLflow and Kubeflow on Kubernetes |
| Investigation and draftingReads statement narratives, scheme reports and past cases, then explains a break in plain English. Both clouds offer a choice of models, including Claude. | Amazon Bedrock | Microsoft Foundry | Open-weight models served with vLLM on your own GPUs, or IONOS AI Model Hub |
| Past breaks and procedures (retrieval)Resolved breaks, desk procedures and account notes, so the agent learns from how your team fixed the same problem last time. | 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 |
| WorkflowRuns each day's cycle to the cut-off, and escalates if a safeguarding or client money reconciliation isn't signed off in time. | AWS Step Functions | Logic Apps or Durable Functions | Temporal, or n8n |
| Link to the ledgerApproved postings go in through the ledger's own interface, with its own approval limits. The AI has no write access. | API Gateway, plus your existing integration platform | API Management, plus your existing integration platform | Kong or your existing API gateway |
| Human reviewAnalysts and controllers work where they already work, not in a new screen. | In the existing reconciliation platform or finance workbench | In the existing reconciliation platform or finance workbench | In the existing reconciliation platform or finance workbench |
| Audit evidenceWrite-once storage for every match, suggestion, approval and posting, for the agreed retention period. | Amazon S3 with Object Lock, CloudTrail | Blob Storage immutable storage, Azure Monitor logs | Ceph with S3 Object Lock, OpenTelemetry |
| Guardrails | Bedrock Guardrails | Foundry guardrails (Azure AI Content Safety) | Open-source guardrails (e.g. NeMo Guardrails, Llama Guard) |
| Identity, keys, data protectionSeparate roles for proposing and approving, so nobody approves their own fix. | IAM Identity Center, KMS, Macie | Entra ID, Key Vault, Purview | Keycloak or your existing Active Directory; OpenBao or HashiCorp Vault; an HSM |
| MonitoringTrack match rates, acceptance rates and ageing by account. A sudden change usually means a source file changed format. | CloudWatch | 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
The general ledger, core banking, the e-money platform and bank connectivity stay where they are. The deterministic matching and the daily sign-off also run next to them, because they must happen every reconciliation day, whatever else is down.
Uses cloud AI
Match suggestions, the investigation agent and retrieval over past breaks can run in a UK cloud region, or on your own GPUs with open-weight models if statement data must stay inside your estate.
How they connect
A private link (AWS Direct Connect or Azure ExpressRoute) and your API layer. Only the transactions in each break are sent, and suggestions come back as events onto your queue.
If the link drops
Rules matching and the analysts' queue carry on locally, and the day's reconciliations are signed off as they were before the AI. Breaks are investigated by hand until the link returns, then the backlog is picked up.
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.
Break investigation agent
Where a person signs offAn analyst accepts or rejects every proposal, and a controller approves every posting, with a second approver above an agreed value. Safeguarding and CASS reconciliations are signed off by the oversight team. The agent's permissions are enforced by the platform's identity and policy controls, and by the ledger's own approval limits, not by instructions in a prompt.
The ledger is reached only through its own API or integration layer, never directly. Postings carry an ID, so a retry never posts twice. If you already run an integration platform, the AI uses it like any other system does.
Each file that lands becomes an event. Parsing, matching, suggestions and investigation follow from it, so a late bank statement simply starts its part of the run late. Nothing waits on a fixed batch window that a delayed file can break.
The daily reconciliation never depends on the AI. If the model or agent is down, the rules still match, the breaks still appear in the analysts' queue, and they work them as before. The legal duty to reconcile is met either way.
Map every account, source and cut-off, measure today's match rates and break ageing, and pick two or three accounts with the most manual work.
Gate · must be true to passBaseline agreed, data access confirmed, DPIA started, finance, safeguarding oversight and internal audit involved.
Shadow mode on the chosen accounts: suggestions and proposed fixes are produced, analysts work as normal, and the two are compared.
Gate · must be true to passSuggestion accuracy and acceptance within agreed limits, no wrong posting proposed above tolerance, and internal audit comfortable with the evidence.
Suggestions and the agent go live on the pilot accounts, inside the existing approval limits, with monitoring and a rollback plan.
Gate · must be true to passMonitoring, support model, rollback plan, updated reconciliation method documents and analyst training in place.
Add accounts in waves, then safeguarding and client money accounts once the evidence has been through an audit cycle.
Gate · must be true to passEach new account passes the same shadow-mode gate.
| Activity | Finance director | Product owner | Solutions architectI can fill this | Technical PMI can fill this | EngineersI can fill this | Safeguarding & CASS oversight | Internal audit | SecurityI can fill this |
|---|---|---|---|---|---|---|---|---|
| Business case and benefits | Accountable | Responsible | Consulted | Consulted | Informed | Consulted | Informed | Informed |
| Target architecture and ledger integration | Consulted | Consulted | Accountable and Responsible | Consulted | Consulted | Consulted | Consulted | Consulted |
| Reconciliation method documents | Accountable | Consulted | Consulted | Informed | Consulted | Responsible | Consulted | Informed |
| Build and integration | Informed | Consulted | Accountable | Responsible | Responsible | Informed | Informed | Consulted |
| Accuracy testing on real breaks | Informed | Accountable | Responsible | Consulted | Responsible | Consulted | Consulted | Informed |
| Go-live decision | Accountable | Responsible | Consulted | Responsible | Informed | Consulted | Consulted | Consulted |
| Run, monitor and improve | Accountable | Responsible | Consulted | Consulted | Responsible | Consulted | Informed | 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: how the AI sits beside your existing reconciliation platform, the ledger integration and approval limits, the evidence trail and the gates. I'd work closely with your finance director and safeguarding oversight, and report to your programme board.
Option B
If your partner, your reconciliation vendor or your own team already has an architect, I'd take on the matching model, the agent and the event flows under their lead, and give them a second senior opinion on controls, resilience and the evidence your auditors will ask for.
Goes wrong
The AI is allowed to post fixes, and a wrong posting reaches a customer's balance.
How the design avoids it
No write access to the ledger. Proposals only, approved through the ledger's own limits, with a second approver above an agreed value.
Goes wrong
The model replaces rules that audit already understood.
How the design avoids it
Rules first, unchanged. ML only works on what the rules leave, and the method documents say so.
Goes wrong
Analysts accept suggestions without looking, so a 99% match rate hides real errors.
How the design avoids it
Acceptance rates watched by account, a weekly sample review, and suggestions that show their reason, not just a score.
Goes wrong
A partner changes its file format and match rates quietly collapse.
How the design avoids it
Monitoring on match rates per source, alerts on sudden change, and parsing that keeps the original line.
Goes wrong
Safeguarding is added on day one, before anyone trusts the evidence.
How the design avoids it
Start with operational accounts. Bring in safeguarding and client money accounts after an audit cycle has reviewed the evidence.
Changes already announced or under way that affect this design. More in the UK horizon scan.
Make the reconciliation evidence audit-ready now, and keep the method configurable so the end-state rules change settings, not the build.
Use enhanced CHAPS data in matching where it exists, and plan the parsing changes for the revised dates rather than the old ones.
Don't hard-code business days and cut-offs into the reconciliation run. Treat them as configuration.
Keep the map of who is accountable for each reconciliation and control, whatever the regime calls their role.
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.
