What AI contract review genuinely cannot do yet

Every vendor in automated contract review, this one included, has a commercial incentive to describe limitations in the vaguest possible terms. The result is a market where buyers cannot distinguish between products, because every product claims to do everything and none of them says what it gets wrong.

So here is a specific account of where the technology fails today, written by people who work on it daily and would prefer customers discover these limits during a trial rather than during a dispute.

Why a vendor should publish this

A contract review tool that is trusted uniformly is dangerous. A reviewer who believes the system catches everything will stop reading, and the one clause the model misparsed will go through unexamined. Calibrated trust, where the user knows which outputs to check, produces better results than high but undifferentiated confidence. That calibration requires the vendor to be specific about failure modes.

Cross-references and defined terms

Contracts are not linear documents. A liability clause in section 9 may be modified by an exclusion in schedule 4, which relies on a definition that appears only in an order form referenced by an incorporated policy. Human reviewers navigate this badly too, but they know when they are lost.

Language models degrade in a less visible way. Given a long agreement, a model will confidently analyse clause 9 while having effectively lost track of the schedule that qualifies it, and the output will read exactly as fluently as an analysis that took the schedule into account. This is why serious systems build an explicit structural representation of the document, a defined-term graph and a cross-reference map, and treat the model as a reader operating within that structure rather than as the structure itself. It reduces the failure rate substantially. It does not eliminate it, particularly where incorporation is by reference to an external URL.

Reasoning about what is absent

The hardest question in contract review is not whether a clause is bad. It is whether a clause that should exist is missing. There is no text to flag, no anomaly to detect, only a silence that a specialist recognises and a general reader does not.

Automated systems handle this through checklists: given this contract type, these provisions should be present. That works for the standard ones, such as a data processing agreement without a sub-processor clause. It fails for the situational ones, where a provision is necessary because of something specific about this deal that appears nowhere in the document. A supply agreement that omits a change-of-control right may be fine, or it may be a serious gap because the counterparty is in the middle of a sale process. No amount of reading the contract reveals that.

Jurisdictional enforceability

Whether a clause is enforceable depends on governing law, on mandatory local provisions that override the parties' choice, and on case law that moves. A model trained on a broad corpus will produce commentary about English law that is roughly correct, commentary about Dutch or German law that is correct in outline and unreliable in detail, and commentary about smaller jurisdictions that should not be relied on at all.

The honest engineering response is to scope the claim. State which jurisdictions carry enforceability commentary, verify those against current law with qualified lawyers on a fixed cycle, and mark everything else as reviewed against the playbook only, with enforceability explicitly not assessed. A blank is safer than a plausible sentence.

The failure mode that matters in this category is not the mistake the system makes. It is the mistake it makes without indicating any uncertainty.

Commercial context it cannot see

Whether a term is acceptable is a commercial question wearing legal clothing. A cap of six months of fees may be an obvious rejection for a strategic customer and an obvious acceptance for a supplier you intend to replace in a year. An aggressive audit right may be unacceptable in principle and irrelevant in practice because of who the counterparty is.

None of that is in the document. A playbook can encode some of it through conditional rules based on contract value, region and counterparty tier, and that captures a useful fraction. The residue, which includes the relationship history, the negotiating leverage and the strategic intent behind the deal, exists only in the heads of the people doing the deal, and a review tool that pretends otherwise is overreaching.

The accountability gap

This is the limit that will not be closed by better models. Professional responsibility for legal advice attaches to a person or a regulated firm. Software cannot hold it, cannot be insured for it in the same way, and cannot be disciplined for getting it wrong.

That is not a temporary state of the technology. It is a feature of how liability is allocated in professional services, and it means automated review will remain an input to a human decision rather than a replacement for one, however accurate it becomes. Any vendor implying otherwise is describing a product it is not permitted to sell.

What follows from all this

Three practical conclusions for anyone evaluating tools in this category.

Ask what the system does when it is unsure. If the answer is that it always produces an answer, that is a defect described as a feature.
Test on your worst documents, not your cleanest. A badly scanned PDF of a 2014 agreement with handwritten amendments tells you far more than a well-formed template.
Check whether findings are attributable. If you cannot see which clause and which rule produced a suggestion, you cannot correct the system when it is wrong, and you will end up either trusting it blindly or ignoring it entirely.

Used within those limits, automated review removes an enormous amount of low-value reading and makes the output of a legal team measurably more consistent. Used outside them, it produces confident text that nobody checked, which is worse than the manual process it replaced.

Clauzy publishes field notes for contract reviewers. Nothing here is legal advice; it is practitioner commentary, and your own counsel decides what applies to your agreements.