Contractual overclaim in AI-enabled knowledge work

Find commitments hidden in ordinary prose

Review more than formal warranties and service levels. Mark absolutes such as “always,” “never,” “fully,” “eliminates,” and “assures every result”; broad security or compliance labels; unrestricted performance claims; delivery dates; data-location and deletion statements; integration coverage; output-accuracy promises; and obligations inherited from exhibits or sales material.

For each statement, identify the exact product, version, plan, configuration, geography, use case, customer dependency, measurement method, exception, and remedy. Distinguish current capability from roadmap, typical behavior from a committed threshold, and a control design from evidence that the control operates effectively.

The UK government's guidelines for AI procurement advise buyers to define the challenge, examine data and model limitations, use output-based requirements, and consider ongoing governance. They are not contract language or legal advice, but they support turning broad AI descriptions into testable requirements with known limits.

Sources: [1]

Run the promise stress test

Create a docket with the proposed clause, plain-language promise, product boundary, supporting source, source date, accountable owner, customer dependency, exception, acceptance test, monitoring evidence, incident route, and disposition. Use five dispositions: support as written, narrow, make conditional, route for specialist review, or remove.

Test the sentence against edge conditions. Would it remain true during a provider outage, model update, permission error, stale source, unsupported language, unusual file type, or customer misconfiguration? If the answer depends on a particular setup, name that setup. If a specialist must interpret law, security, privacy, accounting, or another regulated domain, route the language rather than manufacturing certainty.

PPN 02/24 recommends proportionate due diligence where AI-assisted tender content creates uncertainty and points to clarification or supporting documentation. Although limited to UK public procurement, it supports checking the accuracy, robustness, and credibility of claims instead of treating generated bid text as evidence.

Sources: [2]

Worked artifact: a synthetic analytics clause

A synthetic contract draft says an AI analytics assistant “processes every spreadsheet accurately, never retains customer data, and ensures regulatory compliance.” The promise stress test separates the sentence into three claims.

Testing shows that the supported import formats exclude password-protected and macro-enabled files, and accuracy varies with source structure. The first claim becomes a bounded acceptance test using named formats, representative fixtures, calculation tolerances, and documented exclusions. The data owner can support a retention configuration for a particular service plan but finds an operational log with a separate period, so the second claim is rewritten with system and configuration boundaries pending privacy and contract review. No product can ensure a customer's legal compliance, so the third claim is removed and the contract instead assigns responsibilities for configuration, review, and specialist advice.

The docket retains the rejected wording, source evidence, test results, changed clause, approvers, and unresolved issues. This synthetic example is not legal advice and does not establish that a particular vendor or control meets its commitments.

Verify after signature

Before execution, independently sample high-consequence promises and reproduce their acceptance tests. Confirm that cited documentation applies to the contracted offering and that the named operational owner accepts the obligation. Compare the main agreement, order form, security exhibit, data terms, implementation plan, and incorporated sales materials for contradictory scope.

After release, test promises at the agreed trigger and after material changes. Inspect incidents, failed deletions, support escalations, model or provider updates, and exceptions to determine whether the contractual statement remains supportable. Record failure, containment, customer communication, remedy, retest, and any required amendment.

The control does not replace qualified legal, privacy, security, or commercial review. Two Prune can observe how a participant identifies, verifies, narrows, and escalates claims in a bounded assessment. It does not give legal conclusions, certify contractual compliance, or make employment decisions.