AI Claims Assurance
Preventing AI-washing through evidence-based review."AI washing" is the use of false, exaggerated, vague or misleading statements about AI. Regulators have already acted against unsupported AI claims. We help you make claims that are specific, understandable, consistently used and supported by retained evidence.Our review provides structured, evidence-based challenge and an assurance memorandum within the agreed scope. It is not a statutory audit, legal opinion or certification unless expressly contracted through appropriately qualified parties.
What you get
From risky claim to defensible statement
Before - unsupported
"Our AI eliminates human error and makes fully automated decisions."
After - evidenced & qualified
"Our system uses machine-learning models to prioritise cases for trained reviewers. Final decisions remain subject to human approval. Performance was measured on a defined dataset during a stated period, with documented limitations."
Specific
Every claim identifies the AI component and its material function - no blanket "AI-powered" language.
Item 1Evidenced
Architecture, model cards, test results, logs, data sources and approval records back each statement.
Item 2Defensible
Conditions, limitations, human involvement and geographic scope are stated where they are needed.
Item 3The claims assurance method
Eight controlled stages, one register
Capture
Record every AI claim in websites, tenders, investor materials, sales decks, product sheets, social media and contracts.
Classify
Capability, performance, automation, proprietary technology, safety, fairness, accuracy, compliance, productivity or customer outcome.
Define
Translate marketing language into a testable proposition - "AI-powered" must identify the AI component and its material function.
Evidence
Collect architecture, model cards, test results, logs, contracts, data sources, user research, limitations and approval records.
Challenge
Is the evidence current, representative, reproducible, statistically meaningful and applicable to the actual product version?
Qualify
Add conditions, limitations, human involvement, confidence ranges, geographic scope and dates where needed.
Approve
Technical owner, legal/compliance and marketing approve the final wording before publication.
Monitor
Set expiry dates and revalidate after model, data, supplier, product or performance changes.
Red flags
High-risk phrases that require challenge
| Claim | Required evidence / correction |
|---|---|
| Fully automated | Prove no material human intervention; otherwise describe the actual human role. |
| Proprietary AI | Identify what is owned versus third-party models, APIs, fine-tuning or orchestration. |
| Eliminates bias | Replace absolute wording; define tested harms, datasets, metrics, thresholds and limitations. |
| 100% accurate | Normally indefensible; use validated performance figures with test conditions and error rates. |
| Compliant with the EU AI Act | Map applicable role and obligations; avoid blanket claims without legal and technical evidence. |
| ISO/IEC 42001 compliant / certified | State exact status: implementing, assessed, certification scope, certification body and certificate validity. |
| Secure or safe | Define security/safety properties, testing, threat model, residual risks and monitoring. |
| Human-like understanding | Avoid anthropomorphic implication unless clearly explained and technically supported. |
The deliverable
A Claims Assurance Register with claim owner, source, exact wording, evidence link, reviewer, approval date, limitations, expiry / retest date and status. The review is not a legal opinion unless delivered or signed by qualified legal counsel.