Back to all articles Payments

AI and Open Banking standards: set the evidence boundary

Open Banking Limited's new report identifies useful roles for large language models in standards work. Use a bounded experiment to measure the value while keeping the current Standard and release evidence authoritative.

Kieron James
AI and Open Banking standards: set the evidence boundary

Open Banking Limited published Do LLMs Have Standards? on 17 November 2026. The report examines how large language models could support Open Banking and open finance standards, drawing on six formal responses, wider industry engagement and external research.1

For enterprise payment teams, the immediate decision concerns engineering method. You can test AI assistance in standards mapping, documentation and test preparation today. Approval should depend on evidence from your own work, because OBL found no substantiated, generalisable efficiency savings for Open Banking delivery.

What the OBL report changes

The report gives technology leaders a clearer starting point. Respondents identified several tasks where an LLM may help: mapping legacy data, generating or reviewing code, improving documentation, preparing API data and supporting test work. These tasks share one useful property. A team can compare the output with an authoritative source and a deterministic result.

OBL also describes the limits of the evidence. The report refers to productivity studies from other activities as planning context, while confirming that respondents did not provide generalisable Open Banking savings. A supplier claim or a successful demonstration therefore provides weak support for an enterprise business case.

Measure the task you plan to approve. Record current completion time, review effort, defect rate and rework before introducing an AI tool. Run the same measures during a bounded trial. The result can then support a decision about that task, team and standards version.

Give AI work you can verify

Standards work contains useful candidates for controlled assistance. An LLM can propose a field mapping, explain a schema change, draft a test case or identify differences between two documents. Each output can be traced back to the published specification and checked by existing engineering controls.

That boundary matters in an AIS, PIS or Pay by Bank delivery. A plausible answer can still use an old version, miss an optional field, invent a constraint or overlook a bank-specific behaviour. The person approving the change needs the exact source, version and validation result.

The current public specifications identify v4.0.1 as the latest release.2 Use that published source as the baseline for current work. If your integration depends on an older version or provider profile, record that dependency explicitly. Asima's earlier analysis of standards as delivery infrastructure explains why shared definitions improve delivery only when teams keep their implementation aligned with the source.

Keep one authoritative standards source

OBL describes the current Standard as a strong base and identifies changes that could make it easier for AI systems to consume. These include a single version-controlled source, richer metadata, more examples and machine-readable workflow information.

The report also says OBL is centralising standards on a new platform and exploring services such as model context protocol access and embedded question answering. Treat these as announced direction. Plan current assurance around the public Standard and the controls available now. OBL says live implementations should continue through approved, tested and version-controlled release processes.

The same discipline applies when a model generates a correct answer. Store the standards version and section used for the task. Preserve the proposed change as a reviewable diff. Run the usual static checks, contract tests and journey tests. A controlled API version migration should retain its compatibility, rollout and rollback evidence regardless of how the change was drafted.

Build evidence for every AI-assisted change

An AI-assisted change needs a durable evidence record. This is Asima's recommended minimum:

  • the approved task and the data the tool may receive;
  • the authoritative Standard version, provider profile and supporting documents;
  • the model, tool and configuration used;
  • the generated output and the human-reviewed diff;
  • deterministic validation, journey results, approval and rollback reference.

Keep customer data, payment credentials, production logs and confidential supplier material outside the tool unless the normal security, privacy and contractual reviews approve their use. Use synthetic or sanitised examples for early experiments. Access controls and retention settings should match the sensitivity of the material.

This record supports three decisions. Engineering can reproduce the work. Risk and compliance can see the information boundary. The release owner can confirm that normal acceptance evidence remains complete. Asima's production acceptance strategy provides the wider test structure for the affected API and customer journey.

Need help scoping your open banking proposition?
Book a discovery call with Asima to discuss delivery, compliance and commercial structure.

Separate development assistance from payment authority

The OBL report also considers agentic AI, where software may act for a customer. That question has a different risk boundary from development assistance. A model drafting a schema mapping has no authority to access account data or initiate a payment. A customer-facing agent may interact with consent, authentication, liability and revocation controls.

Keep those workstreams separate in governance and testing. Development tools need information-handling rules, provenance and code assurance. Customer-facing agents need explicit authority, transaction limits, auditability, failure handling and a clear route to withdraw permission. The agentic payments architecture sets out how deterministic services can enforce those controls.

Clear ownership prevents an engineering experiment from expanding into a payment-authority decision. It also helps suppliers answer the right assurance questions. Ask a development-tool supplier how data is used and retained. Ask an agentic-payment supplier how consent, execution and liability are controlled.

Start with measurement

Choose one task with a stable source and a clear expected result. Schema comparison, documentation review or synthetic test generation can work well because reviewers can inspect the full input and output. Use a current backlog item so the trial reflects normal complexity.

Set the success measures before the trial begins. Time saved is useful only alongside review time, escaped defects and rework. Include the cost of preparing safe inputs and preserving evidence. Stop the trial if the tool repeatedly loses source references, produces version drift or adds more review work than it removes.

The trial should end with a decision record. Approve the task for a defined team and information class, change the controls and run another test, or close the experiment. A positive result for one task provides evidence for that task. Wider use deserves its own measurement.

The decision for enterprise teams

OBL's report supports a focused opportunity: use AI where output can be compared with an authoritative Standard and verified through existing release controls. It also gives leaders a firm reason to ask for local evidence before accepting productivity claims.

Start with one measurable task, one current standards source and one accountable reviewer. Preserve the input, output, validation and approval. You will then have evidence that remains useful when the model, supplier or Standard changes.

Footnotes

  1. Open Banking Limited, Do LLMs Have Standards? Findings, Perspectives and Practical Priorities, published 17 November 2026.
  2. Open Banking Limited, Open Banking API Specifications, accessed 19 November 2026.
  3. Open Banking Limited, AI and Open Finance: shaping the future of financial innovation, published 23 September 2025.
Payments