False Negatives, Real Exposure: Contracting for AI Compliance Risk

Artificial intelligence has moved from the periphery of compliance to the centre of it. UK firms use machine-learning tools to screen sanctions lists, score anti-money laundering risk, monitor transactions and prepare regulatory returns. The commercial pitch is familiar: greater coverage, fewer manual hours, faster onboarding. The contractual question is less familiar: when the tool is wrong, who bears the loss?

The answer is not found in the regulator's handbook. The FCA, the Office of Financial Sanctions Implementation and the National Crime Agency will look to the regulated firm, not to its software supplier. A contract can shift economic risk between commercial parties, but it cannot shift regulatory responsibility away from the entity that is supervised. That distinction should shape every accuracy warranty, review obligation and indemnity in the agreement.

A false positive is an inconvenience; a false negative is exposure

Contracts often refer to accuracy in general terms. A supplier may warrant that its tool is accurate, or that it performs in accordance with its documentation. Those phrases do little work when a screening engine misses a designated person or an AML system fails to escalate a suspicious pattern.

Buyers should press for accuracy to be defined by reference to measurable outcomes. That means distinguishing precision from recall, and false positives from false negatives. A tool that flags too many customers creates friction and cost. A tool that flags too few creates regulatory risk. The second is the more serious contractual gap.

A workable clause will identify the test data, the threshold and the consequences of failure. It may specify a minimum recall rate for sanctions screening against a defined list set, a maximum false-positive rate for a defined customer population, and a process for re-testing after material model changes. Without those details, an accuracy warranty becomes a statement of aspiration rather than an enforceable standard.

Human review must be specified, not assumed

The Money Laundering Regulations 2017 require firms to conduct customer due diligence and enhanced due diligence in higher-risk cases. Supervisory expectations also point towards human oversight of automated decisions. But a contract that simply says the buyer will review alerts may not be enough to allocate responsibility fairly.

The agreement should describe what human review means in practice. Who reviews? What training must they have? What is the maximum queue time before an alert is cleared? How are overrides recorded? What happens when the reviewer disagrees with the tool? If the supplier intends to rely on buyer review as a defence against accuracy claims, that reliance should be explicit and matched by service levels that make review feasible.

Suppliers should be equally careful. A supplier that presents its output as a decision, rather than as a risk indicator, may find itself exposed when the buyer's reviewers rubber-stamp alerts under time pressure. Clear labelling, confidence scores and explanation logs help both parties. They also make it harder for either side to claim after the event that the division of responsibility was obvious.

Audit, notification and the problem of the moving model

AI compliance tools are not static. Models are retrained, thresholds are adjusted and list sources change. A contract that grants an audit right once a year may not capture a material change made in the intervening months.

Buyers should require notification of material changes to the model, the data sources or the logic that affects screening outcomes. Materiality should be defined by reference to the buyer's regulatory obligations, not left to the supplier's discretion. A change that alters the false-negative rate in a sanctions screening tool is material even if the supplier regards it as a routine update.

Notification duties should extend beyond model changes. The buyer needs to know promptly if the supplier becomes aware of a defect, a regulatory inquiry or a systemic error affecting the buyer's outputs. The supplier may resist open-ended notification, particularly where other customers are affected. A middle position is to require notice of incidents that affect the buyer's data or the performance of the service, with a separate duty to notify regulatory inquiries that concern the tool.

Audit rights should be practical. A right to inspect the supplier's premises is less useful than a right to receive test results, model documentation and incident reports. Where the supplier relies on a third-party model, the buyer should establish whether audit rights flow through. They often do not.

Indemnities, caps and regulatory fines

Indemnities are where the negotiation becomes most direct. A buyer may seek an indemnity for fines, penalties and third-party claims arising from a tool failure. A supplier will usually resist, pointing to the buyer's own regulatory obligations and the difficulty of pricing unlimited regulatory risk.

There is no single correct answer, but there are familiar boundaries. Regulatory fines are often excluded from indemnities altogether, sometimes on public policy grounds and sometimes because insurers will not cover them. A supplier may instead offer service credits, a fee rebate or a specific indemnity for breach of an accuracy warranty, subject to a cap. A buyer may accept that structure if the cap is proportionate to the risk and if the supplier accepts clear notification and remediation duties.

Both parties should avoid indemnities that assume a clean division of responsibility. If a loss arises partly from a tool error and partly from the buyer's failure to review an alert, a broad indemnity may produce a dispute rather than a recovery. Contributory fault, mitigation and notification failures should be addressed expressly.

Drafting positions for buyers and suppliers

Buyers should focus on five points: a defined accuracy standard tied to test data; human-review obligations with realistic service levels; notification of material model changes and incidents; audit rights that produce evidence rather than impressions; and a remedy structure that includes termination for repeated failures. Step-in rights and exit assistance matter where the tool is embedded in a regulated process.

Suppliers should focus on four: a clear statement that the tool supports rather than replaces the buyer's regulatory judgement; buyer responsibility for review and escalation; a cap and exclusion structure that reflects insurable risk; and a change-control process that prevents routine improvements from becoming contractual breaches.

The underlying point is governance. A contract cannot make an AI tool reliable. It can, however, make the consequences of unreliability predictable. In 2026, with compliance tools under closer scrutiny and regulatory expectations still evolving, that predictability is worth negotiating carefully.