Anonymised under NDA
AI confidence reports for compliance checks
Compliance · A compliance company in the UK
Information submitted to a UK compliance company is cross-checked against open sources, and AI produces a weighted confidence report on every claim.
The problem
- Submissions to the provider (identities, addresses, employment, company roles, declared salaries) were largely accepted at face value, with manual spot-checks that were slow and inconsistent.
- Analysts re-ran the same open-source searches by hand for every case: web search, LinkedIn, Companies House, postcode lookups, and job boards.
- There was no consistent way to express how confident the provider was in a submission, so downstream decisions treated weak and strong evidence alike.
The architecture
- Each submission is decomposed into a structured set of claims: who the person or company says they are, where they are, what role they hold, and what they earn.
- Source agents fan out per claim across web search, LinkedIn profile matching, Companies House officer and filing records, postcode and address validation, and job-board salary benchmarks.
- A weighting layer scores every claim on source reliability, recency, and corroboration across sources. An LLM then composes the confidence report with citations and a per-claim rationale.
- Submissions below the confidence threshold route to an analyst. Every source hit, weight, and model output is stored for audit.
ClaudeLangGraphPostgreSQLCompanies House APIPostcode lookup APIsWeb search APIs
The outcome
- Every submission now arrives with a weighted confidence score and an evidence trail rather than a manual spot-check.
- Analysts spend their time on the low-confidence cases that genuinely need judgement, with a consistent rationale across the team.
Services drawn on
A similar problem on your roadmap?
A Discovery Sprint is the fastest path to a scoped, costed proposal.