Source authority
Binding requirements, official guidance and commentary had to remain clearly distinguishable.
Anonymised legal case study
How a Dutch legal services firm used Vera to verify a regulatory policy summary before internal approval
The draft looked credible enough to approve. Vera exposed where guidance had been presented as obligation, where source authority needed clarification and where professional judgement was still required.
No client name or confidential material is disclosed. The case focuses on the workflow, review findings and approval controls.
The challenge
A polished summary is not the same as a defensible policy
A compliance team was preparing an internal policy update from official guidance, legal commentary and internal notes. Before approval, it needed to know whether every material statement accurately reflected the underlying sources.
Regulatory nuance
Qualified source language could not be converted into an unconditional obligation during summarisation.
Confidentiality
The team needed AI assistance without treating sensitive legal material as ordinary public-model input.
“The risk was not an obviously false sentence. It was a plausible sentence that sounded more certain than the source allowed.”
The approval risk
Once a summary becomes policy, its assumptions become operational
The document could influence controls, staff instructions and client-facing decisions. The compliance officer therefore needed an inspectable review process, not another confident answer.
What had to be verified
- Whether every material claim could be traced to a source.
- Whether the source actually supported the wording used.
- Whether older and newer publications had been mixed.
- Whether internal recommendations were presented as law.
- Which passages still required professional judgement.
claim_source_link: required
source_authority: explicit
unsupported_claims: flagged
privacy_verification: fail_closed
final_decision: compliance_officer
The Vera workflow
Independent review before human approval
The draft and its supporting documents moved through a privacy-aware, multi-model verification chain. Each phase had a defined role; no single model was treated as authoritative.
01
Shield sensitive values
02
Analyse the draft
03
Fact-check claims
04
Challenge assumptions
05
Human approval
The compliance officer could inspect the findings, source relationships and disagreements before accepting any change.
Semantic Privacy Shield
Sensitive values were transformed inside the protected workflow before anonymised content was sent to the selected AI models. When privacy verification could not be completed successfully, nothing was sent onward.
Professional judgement retained
Vera surfaced issues and prepared a revised version, but the compliance officer decided which corrections were accepted and whether the policy was ready for approval.
What the review surfaced
Three issues that a fluent summary could hide
The value of the review was not another rewrite. It was the separation of source-backed statements, interpretation and internal recommendation.
Guidance presented as obligation
A paragraph used unconditional language where the underlying guidance allowed contextual interpretation.
Different dates and authority mixed
An older explanatory publication and newer official guidance appeared in one narrative without a clear distinction.
Recommendation attributed to regulation
A sensible internal recommendation was separated from the legal summary because the cited regulation did not state it directly.
Before and after
From plausible text to an inspectable approval record
The workflow changed what the compliance officer could see before making the final decision.
| Review area | Conventional AI summary | Vera-assisted review |
|---|---|---|
| Source checking | Sources may be listed without showing which claim they support. | Material claims are connected to sources or flagged for review. |
| Legal nuance | Qualified guidance can become overly certain during summarisation. | Models independently challenge wording, scope and source authority. |
| Confidential material | Documents may be submitted directly to a public AI workflow. | Sensitive values are processed through the Semantic Privacy Shield first. |
| Approval evidence | The final answer obscures how corrections were reached. | Review steps, disagreements and corrections remain inspectable. |
| Decision authority | The fluent output can appear more authoritative than it is. | The professional retains control over every accepted change. |
The result
A clearer basis for internal approval
The compliance team entered the approval process with a revised summary, visible source relationships and explicit flags for issues that still required human review.
What improved
- Legal requirements, guidance and internal recommendations were separated.
- Material claims were linked to supporting sources.
- Unsupported or uncertain statements were made visible.
- Corrections and model disagreements remained inspectable.
- The final approval decision stayed with the compliance officer.
Responsible use
Verification supports judgement; it does not replace it
Vera does not guarantee truth, eliminate hallucinations or determine an organisation’s legal position. It makes the review process more inspectable so professionals can identify what is supported, challenged or still uncertain.
No automatic legal conclusion
The interpretation and final decision remain the responsibility of the qualified professional.
No hidden consensus
Model disagreement is surfaced rather than collapsed into a falsely certain answer.
No approval without review
The revised text remains subject to the organisation’s own governance and approval process.
Legal research & compliance
Verify regulatory summaries before they become policy
Use Vera to review sensitive legal and compliance documents through a privacy-aware, multi-model verification workflow.
Every important claim should be supported, challenged or clearly marked for human review before approval.