CUSTOMER ACCOUNT / ATTRIBUTED EVIDENCE
What this case establishes
This is a commercial publisher’s customer account. Roseram checked the source and wrote the analysis; it did not interview the customer, run the product, audit the figures or establish that AI caused the reported business result. The pilot below is a proposal, not a completed experiment.
Primary material reviewed 2026-10-12. Analysis and proposed evaluation by Roseram Editorial.- 01Authorized account context
- 02Validated structure and facts
- 03Bounded answer or handoff
bunq: the documented workflow
Anthropic describes bunq’s Finn assistant using Claude for banking questions, document work and multilingual support. The account praises structured outputs for backend automation. Valid JSON is a formatting property, however; it does not establish that a transaction explanation, extracted amount or account decision is correct.
The mechanism worth examining
A structured response lets a service validate required fields before using them. That makes malformed output easier to reject, but semantic checks remain necessary. Verify dates, currency, beneficiary identity and the account context separately. A user-facing explanation should distinguish a general possibility from an authoritative status retrieved from the bank’s systems.
The implementation decision
Evaluate informational assistance first with synthetic account fixtures. Define which actions are outside the assistant’s authority. Require a verified tool result for any stated account status and preserve human review for consequential cases. Test multilingual rendering of amounts and exceptions, where a small wording change can alter the meaning.
Where the workflow can fail
The assistant can return perfectly shaped data with a swapped amount or mistaken status. It can also translate an uncertain explanation into a firm instruction. Validate both schema and semantics. Do not let a successful parsing step trigger a transfer or account restriction without the authorized controls intended for that action.
A practical evaluation, designed for this use case
Create known-status, unknown-status and contradictory fixtures. Check field-level extraction and the factual content of replies under native-language review where needed. Record unsupported conclusions and escalation quality. This is an engineering evaluation of assistance, not evidence that an autonomous banking system is safe to operate.
Primary sources
Figures and product details above are dated and attributed; they are not current guarantees.
- Anthropic: bunq primary account ↗The attributed workflow and reported result. Proposed pilots and decisions below are original Roseram analysis.
The decision matrix
Original evaluation criteria for this workflow. These are questions to test, not results measured by Roseram.
| Decision | Evidence to collect | Stop or revise when |
|---|---|---|
| Format | Schema validation passes | Valid structure is called factual accuracy |
| Semantics | Currency, identity and status checked | Fields are plausible but incorrect |
| Action scope | Authorization remains explicit | Parsing success triggers a transaction |
YOUR EVALUATION / NOT A REPORTED RESULT
A small pilot you can inspect
Use these steps as a planning checklist. Checking a box records your progress in this view; it does not run or validate an experiment.
Questions this case answers
Does error-free JSON guarantee safe automation?
No. Values can be wrong even when their structure is valid.
Can this case justify autonomous transfers?
No. Consequential financial actions require the bank’s authorized controls and appropriate human involvement.
Tools & capabilities referenced
Operators studying this story often evaluate adjacent AI products. Explore Roseram guides where available:
PRACTICAL NEXT STEPS
Take the evidence into your own project.
Calculate the whole project cost
Compare subscriptions, usage, launch work and renewals.
Test an existing-project workflow
Check import, source loading, preview and publication.
Verify the output you buy
Require evidence tied to a real customer task.
Wiki brief
bunq is examined here for banking answers with authoritative state. The article answers “What should a banking-assistant evaluation check beyond valid JSON?” using linked primary material and original evaluation design. It is not a customer testimonial for Roseram or an endorsement by bunq.
Operator takeaways
- Validate semantics after schema.
- Review translation of exceptions.
- Keep transaction authority separate.
Related case studies
Plan → build → review with routed models and transparent usage.