How to Run an AI-Assisted Due Diligence Review in Six Weeks

Legal due diligence on a mid-market transaction has a familiar shape: four thousand documents, six weeks, a team assembled from whoever is available, and a report that has to be defensible eighteen months later when someone asks why a change-of-control provision was not flagged. This is an operating model for running that review with AI assistance — written as a schedule, because the sequencing is what makes it work.

It assumes a corporate acquisition of a private target with international operations, a virtual data room of three to five thousand documents, and a legal team of four to six. Scale the durations, not the structure.

Week 0 — Before the room opens

The most valuable week is the one before ingestion, and it is routinely skipped.

Agree the materiality thresholds in writing. Contract value, remaining term, counterparty concentration, jurisdiction, and the specific provisions that constitute a red flag for this buyer. These are commercial decisions and they belong to the deal team, not to the reviewers. Written down, they can be applied consistently by a machine. Left implicit, they will be applied differently by six people.

Define the workstreams and their owners. Commercial contracts, employment, IP, real estate, financing, litigation, corporate records. One named owner each, with authority to make calls without escalation below an agreed threshold.

Fix the report format now. Building the review to produce the report is far cheaper than assembling the report from an unstructured review. Agree the columns, the severity scale and the level of summary the deal committee will actually read.

Week 1 — Ingestion and triage

Load the data room with folder structure preserved; the seller’s organisation carries information about what they consider related. Expect and handle three problems immediately.

Duplicates. A typical room contains 15–30% redundancy: the same agreement filed under two workstreams, plus scanned and native versions of the same document. Collapse these into a canonical record with variants attached, or the same finding will appear four times in the report.

Scan quality. Run OCR with confidence scoring and quarantine low-confidence documents for manual handling rather than letting findings be drawn from unreliable text.

Unsupported languages. Route to a separate queue explicitly. The failure that hurts is not the document you knew you could not read; it is the one that was silently skipped.

By the end of week 1 you should have a classified inventory: every document typed, assigned to a workstream, and either in the automated pipeline or in a named exception queue. Report the exception queue size to the deal team — it is a genuine indicator of how the review will go.

Week 2 — Extraction and first-pass findings

Extraction runs against the schema for each document type. For commercial contracts, the buyer-relevant set is stable across transactions: change of control, assignment, exclusivity, most-favoured-nation, termination for convenience, non-compete and non-solicit, IP ownership and licence grants, liability caps and indemnities, term and renewal mechanics, and any restriction on the target’s freedom to operate.

The output is a structured table, not a narrative: one row per contract, one column per provision, with the source passage behind each cell. This is the artefact the rest of the review operates on, and it is worth insisting that it exists in that form before anyone starts writing prose.

Run the materiality rules over it and you have a first-pass exception list — typically 8–15% of documents on a mid-market deal. That list, not the data room, is what the team reviews.

Weeks 3–4 — Human review of exceptions

Reviewers work the exception list, and their job is judgement rather than location: is this finding correct, is it material on these facts, what is the consequence for the transaction, and what should the buyer do about it.

Three practices make this stage work.

Review in the document, not in the report. Every finding opens to its source passage in context. A reviewer who has to search for the clause will eventually stop checking.

Record disagreement. When a reviewer rejects a finding, capture why in a structured field. Clusters of rejections point at a threshold set wrong or an extraction pattern that needs fixing — and fixing it in week 3 improves the remaining 70% of the review.

Sample the non-exceptions. Take a random sample of documents that the pipeline cleared and review them fully. This is your quality control, it takes a day, and it is the answer to the question you will be asked if something is missed. Two per cent is a defensible sample on a room of this size; record the result either way.

Week 5 — Second-order analysis

The findings so far are per-document. The insights that change deal terms are usually cross-document, and they are what a purely manual review most often misses because no individual reviewer sees enough of the estate.

• Concentration. What proportion of revenue sits with counterparties who have a change-of-control consent right? This single number has repriced transactions.
• Clustering by counterparty. Five agreements with the same customer, each individually immaterial, aggregating above threshold.
• Systematic drafting weakness. If the target’s standard terms cap liability at a level the buyer would not accept, that is a post-completion remediation programme, not a contract-by-contract finding.
• Gaps. Customers with revenue in the financial data but no contract in the room. Missing documents are findings.

Week 6 — Report and handover

The report writes itself from the structured data if week 0 was done properly. What still needs human work is the executive summary: the five things that would change the buyer’s view, stated in one line each, with the consequence and the recommendation.

Everything else is appendix, indexed to the source documents. Include the methodology — what was reviewed automatically, what was reviewed manually, what the sample results were, what was excluded and why. This is not defensive padding; it is what makes the report usable by the people who inherit it, and it is what you will want in the file if a warranty claim follows.

What this actually saves — and what it does not

Across the reviews we have supported, first-pass review time falls by roughly two-thirds. The savings are almost entirely in location and extraction. Judgement time is unchanged; second-order analysis time frequently goes up, because for the first time the data supports it.

What does not change: the responsibility. The report is signed by the lawyers on the matter, and the client is buying their judgement. The tooling makes the judgement better informed and more consistently applied. It does not make it optional, and any vendor who suggests otherwise should be politely shown the door.