Menu

Applied AI · Development direction

AI assistance.
Inspectable decisions.
Human control.

We are exploring how AI can help teams interpret messages, structure information and investigate integration problems. Each proposed use begins with a defined task and a way to check the result.

Discuss an AI use caseExplore the worked example
Current scope
These are proposed areas for assessment and development. This page does not announce a deployed AI product. The worked example is manually prepared using synthetic data; it does not run a model or process submitted information.

Where AI could help

Three practical areas to investigate.

Proposed capability

Message mapping assistance

Help teams interpret unfamiliar fields, compare schemas and prepare candidate mappings for technical review.

A useful output: a mapping proposal with source evidence, assumptions and unresolved fields.

Proposed capability

Documents into structured data

Explore extracting agreed fields from documents, with missing, conflicting or ambiguous values flagged for review.

A useful output: structured draft data linked to the source passages that support it.

Proposed capability

Integration investigation

Help technical teams navigate schema documentation, explain validation failures and prepare troubleshooting notes.

A useful output: a suggested explanation supported by documentation and checks a team can reproduce.

Synthetic worked example

A plausible mapping still needs evidence.

Consider a source system sending an amount, a currency and an internal recipient reference. The question is which values can safely be mapped, and which conclusions the available information does not support.

1 · Source message

{
  "instruction_id": "AI-DEMO-001",
  "amount": "125.00",
  "currency": "GBP",
  "party_ref": "TEST-R17"
}

All values are synthetic. The reference identifies a test recipient inside the source system.

2 · Candidate suggestions

An assistant might suggest mapping the amount and currency directly. It might also mistake the internal recipient reference for a bank account, or infer the recipient’s country from the currency.

Those suggestions need different review outcomes. A convincing explanation cannot replace source evidence.

Review the proposal

Amount and currency → amount fields

Retain the supplied values in a draft, subject to the agreed target schema and business-rule checks.

Internal reference → bank account

Reject this mapping. The message supplies an internal reference, not a verified account identifier. Keep it as a reference and request the required account information separately.

GBP → recipient country

Reject this inference. Currency does not establish the recipient’s country. Leave the country unresolved if the target requires it.

3 · Reviewed draft

{
  "payment": {
    "reference": "AI-DEMO-001",
    "amount": {
      "value": "125.00",
      "currency": "GBP"
    },
    "recipient_reference": "TEST-R17"
  }
}

What the draft establishes

The draft retains the supplied information without converting an internal reference into an account or inventing a country.

It still needs the receiving system’s schema and acceptance rules. It is not a complete payment instruction, an ISO 20022 message, or output from a live AI service.

AI assurance

Define the controls before the pilot.

An AI engagement should agree how the task will be evaluated and who is responsible for reviewing its outputs. These are proposed engagement requirements, not a claim that a particular implementation already provides them.

Data and deployment

Agree which data may be used, where processing occurs, which providers are involved, and the permitted access and retention. Begin with synthetic data where possible.

Evaluation and validation

Define representative test cases, failure cases and acceptance criteria. Check structured outputs against explicit schemas and rules.

Review and traceability

Record source evidence, assumptions and versions. Identify the reviewer, the rejection pathway and what happens when a result is uncertain or incomplete.

Relationship to the translator

AI can assist the mapping process.

Candidate mappings can inform technical design. A production transformation still requires an agreed specification, validation and operational acceptance.

Explore translator development

Start with one useful task.

Describe the information you have, the output you need and how your team would determine whether the result is correct. We can discuss the scope of a feasibility assessment.

Discuss an AI assessment