AI Agents for Lawyers

Use Cases

AI Due Diligence: From Twenty Associate Hours to Two

AI due diligence now runs as a tabular agent over the data room: 10,000 documents against 100 questions, every cell cited. Numbers, a question matrix, the junior gap.

By Dr. Niklas SchmidtUpdated 13 November 202610 min read
On this page
  1. The numbers behind the headline
  2. How an agentic diligence review runs
  3. Designing the question set
  4. Change-of-control and assignment sweeps
  5. Connectors and the "every cell cited" rule
  6. Red-flag reports partners will read
  7. The training gap it creates
  8. What to do on your next deal

AI due diligence in 2026 means an agent runs a fixed question set across every document in the data room and returns a grid with a citation in every cell, which lawyers then verify and turn into a red-flag report. A Legalweek 2026 panel reported financial-contract diligence falling from 15–20 associate hours to about 2; firms working live deals report 15–20% savings on standard data rooms and up to 75% on unstructured ones.

The numbers behind the headline

FigureSource and dateWhat it actually measures
15–20 associate hours to about 2Legalweek 2026 panel, 19 Mar 2026 (reported, no primary write-up)Financial-contract diligence on a defined document set
15–20% time saved on standard diligence; up to 75% on unstructured data roomsGSK Stockmann, in Harvey's diligence guide, 29 Apr 2026 (Harvey)Live deals; savings scale with how messy the room is
10,000+ diligence red-flag reportsPwC's Harvey partnership, Apr 2026 (Harvey)Volume, not accuracy
10,000 documents x 100 questionsThomson Reuters Tabular Analysis, CoCounsel Legal GA 20 Aug 2026 (Thomson Reuters)Capacity of one tabular run
Up to 100,000 documents in parallelLegora Tabular Review, per Addleshaw Goddard's adoption note, 6 Mar 2025 (Addleshaw Goddard)Capacity, after roughly a year of evaluation
20M+ terms extracted; 445K+ Deep Analysis reportsHarvey Agent Builder launch, 9 Mar 2026 (Harvey)Platform-wide volume across customers
94.8% vs 70.1%Vals AI Legal AI Report, 27 Feb 2025 (Vals AI)Harvey vs lawyers on document Q&A, the core diligence task; lawyers still won EDGAR research

The pattern is consistent: extraction and question-answering over a defined set of documents is the thing these systems do best, by a wide margin over unaided lawyers on speed and by a measurable margin on accuracy. What the numbers do not measure is the judgment layer: whether a flagged clause matters for this buyer, this structure and this price. Slaughter and May's decision to go firm-wide with Harvey in April 2026 named M&A and due diligence first among the use cases, "after a long period of consideration" (Legal IT Insider). The Harvey, Legora and CoCounsel reviews compare the three most common platforms for this work.

How an agentic diligence review runs

  1. Scope the set. Agree with the deal team which folders are in scope (material contracts, employment, IP, real estate, litigation, regulatory) and what "material" means in this deal (thresholds by value, term or counterparty).
  2. Connect, do not export. Point the agent at the data room or DMS through a governed connector. Exporting 4,000 PDFs into a chat tool breaks the audit trail and, on most consumer tiers, the confidentiality analysis.
  3. Load the question set. One row per question, with the answer format fixed (yes/no, date, amount, clause quote, "not found"). See the matrix below.
  4. Run the tabular pass. The agent returns a grid: documents down, questions across, a citation in each cell to the document and passage relied on.
  5. Verify by exception. Lawyers open every cell flagged as a trigger, an exception or "uncertain", plus a random sample of the cells marked clean. Record who checked what.
  6. Draft the red-flag report from the grid. Filter the grid to the flagged rows, group by issue, and write the narrative around it. The grid is the appendix.
  7. Keep the run. Store the grid, the question set and the verification log with the matter. If the buyer's insurer or the seller's counsel asks how the review was done, this is the answer.

Anthropic's open-source corporate-legal plugin describes step 4 as tabular diligence with "every cell cited", ships a "dataroom-watcher" agent that flags new uploads, and is built to detect and quality-check output from Luminance and Kira rather than replace it (Claude for Legal repo, 18 Sep 2026). Harvey's Agent Builder describes agents that "flag moments requiring user confirmation before proceeding". Both are the same design decision: the agent does the reading, and stops before the judgment.

Designing the question set

The question set is the work product, and it is where most diligence reviews go wrong before any AI is involved. Vague questions produce vague cells. A good matrix has four properties: each question has one answer, the answer format is fixed, the question names what counts as a trigger, and "not found" is an allowed answer so the agent is never pushed to guess.

#Question (one per column)Answer formatTrigger for the red-flag report
1Does the contract contain a change-of-control clause? Quote it.Yes/No + clause quote + clause numberYes
2Is the change-of-control trigger direct, indirect (upstream) or both?Direct / Indirect / Both / UnclearIndirect, Both or Unclear
3What is the consequence on a change of control?Termination right / Consent required / Notice only / Other (describe)Termination right or Consent required
4Can the counterparty terminate for convenience, and on what notice?Yes/No + notice period in daysYes with notice under 90 days
5Is assignment restricted, and does the restriction cover deemed assignment by merger?Restricted / Free / Not addressed + merger language Yes/NoRestricted with merger language
6Exclusivity, non-compete or most-favoured-customer terms?Present/Absent + quotePresent
7Governing law and dispute forumLaw + forum + arbitration Yes/NoAny law outside the buyer's approved list
8Term, expiry date and renewal mechanismDates + Auto-renew Yes/No + notice periodExpiry within 12 months of expected closing
9Annual value or minimum commitmentAmount + currencyAbove the materiality threshold
10Personal data processed under the contract; DPA attached?Yes/No + DPA Yes/No + transfer mechanismPersonal data with no DPA, or non-EEA transfer with no mechanism
11IP: ownership of deliverables and any licence-backAssigned to us / Licensed / Retained by counterparty / UnclearRetained or Unclear
12Uncapped or unusual liability, indemnity or guaranteePresent/Absent + quotePresent

Twelve columns like these, applied to 2,000 contracts, produce 24,000 cells. That is the number to keep in mind when the platform quotes capacity: the constraint is not the agent's reading time but the lawyers' verification time on the flagged rows. Question design determines both the flag rate and the false-positive rate, and it is the skill that separates a two-hour review from a twenty-hour one.

Change-of-control and assignment sweeps

These two columns carry most of the value on a share deal or a merger, and they carry most of the errors. Three patterns to watch:

  • Indirect triggers. A clause that bites on a change of control of "the Supplier or any of its holding companies" is a trigger on an upstream sale even though the contracting entity never changes hands. Agents classify these inconsistently; column 2 exists to force the distinction.
  • Deemed assignment. Some contracts treat a merger or a transfer of undertaking as an assignment. The agent finds the assignment clause and may miss the deeming language two sub-clauses down. Column 5 asks for it explicitly.
  • Schedules and side letters. The consent requirement is in a framework agreement; the operative supply is in an order form that incorporates it by reference. Feed the whole family of documents and ask the agent to identify the parent for each order form.

Then verify by opening the document. Vals AI's 2025 finding that lawyers beat every tool on EDGAR research, 70.1% against 55.2% for the best tool, is a reminder that finding the right document in a corpus is still weaker than answering questions about a document already in hand (Vals AI).

Connectors and the "every cell cited" rule

Where the agent reads from matters as much as what it reads. As of September 2026 the practical routes are:

  • Data-room and DMS connectors. Anthropic's Claude for Legal ships 20+ MCP connectors including Datasite, iManage, NetDocuments and Box (12 May 2026). iManage's MCP Server is "permission-bound", fully logged and "respects existing ethical walls" (iManage, 14 May 2026); NetDocuments' runs every call under the user's own credentials so that "documents never leave the platform" (NetDocuments). The MCP setup playbook covers the configuration; the glossary entry explains what the protocol is.
  • Platform vaults. Harvey Vault, Legora's workspace and Lexis Protégé's Vault (up to 100,000 documents) hold the set inside the vendor's environment, with the vendor's retention and training terms applying.
  • Scans. Data rooms still contain scanned leases and signed PDFs. Run a dedicated OCR pass before the tabular run; models read scans by vision but at a higher error rate on figures and dates. The PDF extraction playbook has the workflow.

The rule to write into the engagement: no cell without a citation to document and passage. A grid that answers "Yes" to a change-of-control question without a clause reference cannot be verified by exception, which means it cannot be verified at all in the time available.

Red-flag reports partners will read

The report a partner reads on the way to the negotiation has four parts: a one-page summary of deal-affecting issues ranked by consequence, an issues table (issue, documents affected, recommended action: consent, indemnity, price adjustment, condition precedent), the filtered grid as an appendix, and a methodology note stating the question set, the tool, the model version, and who verified which rows. The methodology note is new. Clients now ask for it: Citi has begun asking firms bidding for its work to document AI savings, and Axiom's July 2026 survey of 528 in-house leaders found 92% expect or are negotiating AI-related rate cuts (Axiom). The billable-hour guide covers the fee conversation; the M&A practice page covers the rest of the deal.

The training gap it creates

Diligence was how associates learned to read contracts. If the agent reads and the associate verifies by exception, the associate sees the flagged 5% and never the clean 95%. Three data points make the gap concrete:

  • LexisNexis' Mentorship Gap report (Feb 2026, nearly 900 UK lawyers): 58% say AI makes them faster, 72% name deep legal reasoning and argumentation as the biggest junior skills gap, 69% cite weak verification and source-checking, and 2% think AI strengthens learning (Legal Cheek, 10 Feb 2026).
  • NALP: firms with 500+ lawyers hired 6,588 graduates from the Class of 2025, down 538 or 7.5%, the first fall since 2011; Citi's survey of 57 large firms found 67% predicting fewer first-year associates by 2028 (Bloomberg Law, Sep 2026).
  • The counter-view from a firm that built its own funds-diligence tool, Fried Frank's Becky Zelenka: "It enables you to do more deals. We do expect a junior would become what we think of now as a mid-level faster" (Above the Law, May 2026).

Wolf Theiss's Chief Digital and Technology Officer Leonhard Wassiq put the risk bluntly in Kurier: firms that hire fewer people now because of AI "will be hit hard", and AI systems "deliver suggestions, but these must be checked" (Kurier, 2026). The practical answer is to move juniors up the pipeline: they draft the question set, they run the calibration on the first 50 documents by hand, they own the verification log, and they present the methodology note. That is more reasoning, not less. The junior lawyers guide goes further on redesigning the training year. Module 2 of the AI Lab for Lawyers course runs a tabular diligence exercise end to end, from question design to verified grid, on a sample data room.

What to do on your next deal

  • Write the question matrix before you open the data room, with fixed answer formats and "not found" allowed.
  • Connect through a governed connector or a platform vault. Do not upload the data room into a consumer chat tier.
  • Hand-review the first 50 documents against the agent's grid to calibrate the flag rate, then verify by exception.
  • Put a methodology note in the report: question set, tool, model version, who verified which rows.
  • Give the juniors the question set and the verification log. The reading was never the point; the reasoning was.

Frequently asked questions

How much time does AI save on M&A due diligence?

It depends on the data room. A Legalweek 2026 panel reported financial-contract diligence falling from 15–20 associate hours to about 2. GSK Stockmann reports 15–20% savings on standard diligence and up to 75% on unstructured data rooms using Harvey, as of April 2026. The saving is largest where documents are messy and the questions are precise.

What is tabular review in due diligence?

Tabular review runs a fixed list of questions across every document in a set and returns a grid, one row per document and one column per question, with each cell citing the passage it relies on. Thomson Reuters' Tabular Analysis handles 10,000 documents against 100 questions; Legora's Tabular Review is quoted at up to 100,000 documents in parallel.

Can AI find change-of-control clauses reliably?

Finding the clause is the reliable part; classifying it is where errors occur. Agents locate change-of-control, assignment and termination triggers well, but they misread indirect triggers, deemed assignments and clauses buried in schedules. Treat the agent's grid as a first pass and have a lawyer open every cell flagged as a trigger or as uncertain.

Does AI due diligence hurt junior lawyer training?

The evidence says yes unless firms redesign the work. LexisNexis' Mentorship Gap report of nearly 900 UK lawyers found 72% see deep legal reasoning as the biggest junior skills gap and 69% cite weak verification, while only 2% think AI strengthens learning. The fix is to make juniors own the question set and the verification, not the reading.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion