Skip to content
Novandri.

AI for Contract Review: What to Automate and What to Escalate

AI can find clauses and flag deviations, not decide legal risk. Read the split, a worked triage, and why escalation is a feature.

Rachmadwipa Novandri4 min read

Contract review is an attractive target for AI: the documents are long, the language is repetitive, and a first pass takes experienced time. The useful division of labour is clear — let the model find and compare clauses, and let a lawyer decide what the clauses mean for this deal. Blurring that line produces confident summaries that miss the point.

Direct answer: Anthropic's use-case guides cover legal summarisation, and the document-processing article covers the extraction mechanics. The boundary: a model can flag that a clause differs from the standard, but it cannot judge whether that difference is acceptable given the commercial context, which is a legal and business decision.

Automate clause identification, comparison against a template, and extraction of key terms; escalate deviations, unusual language, and anything with material legal consequence to a qualified reviewer.

A flow: identify clauses, compare to template, flag deviations, then escalate material risk.
Conceptual: escalation is a feature. A flag is a prompt to look, not an opinion.

What to automate

Clause identification. Locate the termination, liability, indemnity, and governing-law clauses in a document. This is a retrieval-and-extraction task, as AI for Document Processing: Classification, Extraction, Redaction describes.

Template comparison. Flag where a clause deviates from the approved standard. The comparison is mechanical — does the text match the expected pattern — and the model can do it at scale.

Key-term extraction. Pull parties, dates, values, and renewal terms into a structured field, using a schema as the structured-output article explains.

First-pass summarisation. Produce a short abstract for a reviewer to orient themselves, with each claim linked to the source clause, as the reduce-hallucinations article recommends for grounding.

What to escalate

Any material deviation. A clause that differs from the template may be a negotiable point or a deal-breaker; that judgement needs context.

Unusual or novel language. If the model has not seen a clause form before, its confidence is not evidence of meaning. The benchmark discussion of coverage applies here, as the benchmarks article argues.

Cross-clause interactions. A liability cap and an indemnity clause may interact in a way that is only visible when read together, which the document-processing article notes as a systemic limitation.

High-value contracts. The cost of an error scales with the deal, so the review threshold should scale too, as the risk-assessment article explains.

A worked triage

The following is hypothetical and illustrative.

A team reviews 500 vendor contracts a quarter. The AI pipeline:

  1. Extracts key terms and identifies the standard clauses.
  2. Compares each against the template and flags deviations with the source text.
  3. Produces a one-page summary for the reviewer.

Outcomes on a sample quarter:

  • 350 contracts have no material deviations and go through a light review.
  • 120 have one or two flagged clauses and go to a lawyer with the clauses highlighted.
  • 30 have unusual language or high value and receive full legal review.

Three checks decide whether the triage is safe:

  1. False-negative audit. Have a lawyer review a random sample of the 350 "clean" contracts. If a material deviation was missed, the comparison threshold is too loose, which the evaluation-rubric article explains for false negatives.
  2. Citation accuracy. Check that each flagged clause is quoted correctly, which the reduce-hallucinations article treats as a core check.
  3. Review time. Measure whether the 120 flagged contracts actually take less time with highlights, or whether reviewers re-read everything, which the workflow-automation article covers.

Three conclusions follow:

  • The value is in triage, not in replacing review.
  • False negatives are the dangerous error, so audit the clean bucket.
  • Legal retains the decision on every material clause, as the human-in-the-loop article argues.

What contract AI cannot do

It cannot give legal advice.

It cannot read commercial intent. Whether a liability cap is acceptable depends on the deal's value and alternatives, which are not in the document.

It stays current only if the template does. A comparison against an outdated template produces misleading flags, which the monitoring article covers as a change trigger.

The counterargument — that experienced lawyers will not use such a tool — is an adoption problem, not a capability problem.

A tool that reduces the time to find clauses while keeping the decision with the lawyer is often welcomed, as the contract-review use-case guide suggests.

A decision test

  1. Automate identification and comparison, not judgement.
  2. Cite the source clause for every extraction and flag.
  3. Escalate material deviations and unusual language.
  4. Audit the clean bucket for false negatives.
  5. Scale review thresholds with deal value.
  6. Keep the template current with an owner.
  7. Measure review time and missed-clause rate, as the evaluation article recommends.

For the extraction mechanics, see AI for Document Processing; for the governance, see AI Governance.

References & Further Reading