The Legal Stack
Independent LegalTech Analysis
← Research Briefings
Research BriefingNo. 086 · August 13, 2026 · 10 min read
Legal AI · Research Report

The Legal AI Embedded Clause Benchmark Report 2026: How Often AI Contract Review Tools Miss Risk Buried in Schedules, Exhibits, and Incorporated-by-Reference Documents — and Whether Any Platform Has Solved the Multi-Document Problem

Whether AI review catches risk in schedules, SOWs and incorporated terms is unmeasured. Why cross-document analysis is hard, and how to test tools.

Research briefing — a review of public sources. The Legal Stack did not conduct a survey, benchmark or interviews for this piece.

The question, and the state of the evidence

Commercial deals rarely live in one document. Risk often sits in schedules, statements of work (SOWs), data processing addenda (DPAs), order forms and terms incorporated by reference. In these places it can quietly change a liability cap, a cure period or a renewal right set out in the master agreement. The obvious question for buyers is how often AI contract review tools miss that risk, and whether any platform has solved the multi-document problem.

The honest answer is that no public, independent benchmark measures this. We found no published study comparing AI tools on risk embedded outside the main agreement, or on detecting conflicts between a master agreement and its attachments. Any "detection rate by tier" figures for named or anonymised platforms should be treated as unsupported. The Legal Stack did not test any platform.

What is documented is useful. The main public benchmarks test narrower tasks. Research on long-context models explains why cross-document synthesis is hard. Vendor documentation shows how multi-document features are actually built. And contract law sets clear rules on when outside documents bind the parties at all.

What existing benchmarks measure

  • CUAD (2021). The Contract Understanding Atticus Dataset, by Hendrycks and colleagues and published for the NeurIPS Datasets and Benchmarks track, contains 510 contracts and more than 13,000 expert annotations across 41 clause categories that lawyers typically review. It is the standard academic test for clause extraction. Its task is to find clauses in a contract, not to reconcile one contract against another.
  • LawGeex (2018). The widely cited LawGeex study compared an AI tool with 20 experienced lawyers on five NDAs and reported about 94% accuracy for the AI against about 85% for the lawyers. NDAs are short documents that stand alone.
  • Vals Legal AI Report (February 2025). Vals tested seven tasks, including document extraction, document Q&A, summarisation and redlining, on products from Harvey, Thomson Reuters (CoCounsel), vLex (Vincent AI) and Vecflow (Oliver). According to LawSites, Harvey scored 94.8% on document Q&A and CoCounsel 89.6%, and the AI tools collectively beat the lawyer baseline on data extraction. LexisNexis withdrew from all tasks except legal research. The published coverage does not describe any test of conflicts across multiple documents.
  • Stanford RegLab/HAI (May 2024). On more than 200 legal research queries, Lexis+ AI and Ask Practical Law AI hallucinated more than 17% of the time, and Westlaw AI-Assisted Research more than 34%. This is research rather than contract review, but it shows that well-known legal AI products make material errors on structured tasks.

The pattern: published evaluations reward accurate answers about one document at a time. None that we found isolates the case where an exhibit quietly overrides the master agreement.

Why cross-document synthesis is technically hard

Two documented points explain why buyers should test this directly rather than assume it works.

Long inputs degrade attention. In Lost in the Middle (Liu et al., arXiv 2023, later published in TACL), the authors found that language models perform best when relevant information sits at the beginning or end of a long input. Performance dropped significantly when it sat in the middle, even for models designed for long contexts. A deal package joined into one long input puts schedules and exhibits exactly where this research says models are weakest.

Per-document extraction is not the same as synthesis. Harvey's public documentation for Vault, its bulk-review product, separates two query types:

  • A Review Query returns an answer from each document into a table.
  • An Assistant Thread returns one aggregated answer across the selected documents.

Vault supports up to 10,000 files per project. Harvey says it achieves 94% to 99% accuracy on objective review tasks with 10 to 50 questions. That is a vendor claim, and it relates to per-document extraction. A table showing "liability cap" for an MSA and, separately, for an SOW is useful. Deciding which cap governs requires reading the precedence clause and applying it, which is a different task. Buyers should find out which mode a vendor's "multi-document" claim actually refers to.

The incorporated-by-reference problem is partly a legal one

Whether a URL-linked policy or a named standard is part of the contract at all depends on the law, and tools cannot settle that by retrieval alone. A March 2025 survey of New York law on Mondaq sets out the main tests:

  • The incorporated document must be identified so that it is identifiable "beyond all reasonable doubt."
  • Terms bind where they are clearly and conspicuously presented and made available for review.

The cases go both ways:

  • In Eshaghpour v. Zepsa Industries (1st Dep't 2019), an oblique reference to an unidentified terms page that was never provided was not enough.
  • In Matter of County of Rockland (3d Dep't 2024), a plan description was not incorporated because the agreements neither referred to it explicitly nor identified it beyond reasonable doubt.
  • In Madison Industries v. Garden Ridge (Sup. Ct. N.Y. Co. 2011), a letter directing the customer to website terms did incorporate them, including a forum-selection clause.
  • Outside New York law, Specht v. Netscape (2d Cir. 2002) held that terms hidden below a download button did not bind users.

The practical point is that a good tool should do two things. It should flag every external reference. And it should report whether it reviewed the referenced document or could not get it. Flagging a gap in coverage is the realistic minimum. Whether the reference is legally effective remains a question for a lawyer.

An evaluation protocol for legal departments (analysis)

Because no public benchmark covers this failure mode, buyers need to build their own test. The protocol below is The Legal Stack's suggested approach.

  1. Assemble a realistic test set from your own past deals, anonymised. Include packages where a schedule, SOW, DPA or order form changes a term in the main agreement, and at least one package that incorporates an external document by URL or by name.
  2. Create an answer key first. A senior lawyer records every material change made outside the main agreement, and which document prevails under the precedence clause.
  3. Upload the package as separate files, not one joined PDF, so you test the tool's native multi-document handling.
  4. Ask synthesis questions explicitly. For example: "Does any attachment change the liability cap, cure period, renewal or termination rights, or indemnity scope in the MSA? Which document governs?" Score the answers against the key.
  5. Test reference awareness. Ask whether the tool reviewed each incorporated document. Treat a confident answer about a document it never had as a serious failure.
  6. Repeat runs. Generative tools can give different answers to the same question, so check consistency.
  7. Set pass marks in advance, and write any multi-document capability you rely on into the contract with a defined performance standard.

Bottom line

AI contract review tools are well tested on finding clauses within a single document. Independent public evidence that any platform reliably reconciles risk across a master agreement and its attachments does not exist yet. Until it does, legal departments should treat cross-document synthesis as unproven, test it on their own deal packages, and keep human review of precedence and incorporated terms in any high-value deal workflow.

Sources

Filed under Legal AI → · The Legal Stack accepts no vendor funding for its research.

More Research

View all →
No. 113
10 min
What Published USPTO Guidance, Federal Circuit Opinions, and PTAB Decisions Through October 2026 Actually Say About AI-Assisted Patent Prosecution, Inventorship Disclosure Obligations, and the Duty of Candor When AI Tools Generate Claim Language
10 min
No. 112
10 min
What Published NLRB Guidance, ALJ Decisions, and Board Orders Through October 2026 Actually Say About the Use of AI-Assisted Monitoring, Discipline, and Work Allocation Tools in Unionized and Union-Eligible Workplaces
10 min
No. 111
10 min
What Published Delaware Court of Chancery Opinions, Docket Orders, and Judicial Commentary Through October 2026 Actually Say About AI-Assisted Legal Filings, Discovery Productions, and Counsel Disclosure Obligations in Corporate Litigation
10 min
© 2026 The Legal Stack — Independent LegalTech Analysis