AI for Legal Due Diligence

Two years ago, AI in legal due diligence meant slightly better keyword search and a few experimental clause-extraction pilots that never quite worked. The honest assessment today is different: in a defined set of high-volume, low-judgement workflows, large-language-model tooling now genuinely outperforms a competent junior associate, at a fraction of the cost and time. In other parts of the legal DD workflow, it remains unsafe and overconfident. The practitioner skill is knowing which is which.

Where AI Now Works Well

The pattern is consistent across the workflows where AI delivers real value. They are high-volume, repetitive, and have well-defined outputs against well-understood concepts. Reviewing four hundred commercial contracts for change-of-control, assignment, exclusivity, and termination provisions is the canonical example. The concepts are well established in case law and standard drafting; the work is mechanical; the volume is what makes it expensive when done by people. Modern tooling reliably extracts these provisions, flags the unusual ones, and produces a structured output that drops directly into an issues list. The same is true of employment-contract review for restrictive covenants, of lease review for assignment and consent provisions, and of IP-assignment chains.

Disclosure Schedules

Drafting disclosure schedules from a populated contract set used to be one of the most thankless tasks in a deal. With AI tooling fed a structured set of contracts and a representations and warranties shell, a credible first draft is now a matter of hours rather than weeks. The lawyer's job shifts from drafting to reviewing and judgement — which is the work the client is actually paying for. The productivity uplift here is genuine and large, and is one of the cleanest ROI stories in legal AI.

Where AI Still Fails

The failure modes are equally consistent. Novel risk identification — the kind that depends on understanding the commercial context, the regulatory environment, and the deal thesis simultaneously — is not something current models do reliably. They will confidently miss the unusual indemnity that is the actual deal risk while flagging twenty boilerplate provisions as material. Privilege analysis, advice on negotiation positions, and final views on whether a contract is acceptable are not jobs to delegate to a model. So is anything where the cost of a single wrong answer exceeds the savings from automating a hundred right ones.

Citations Are Non-Negotiable

Every AI output used in a legal DD workflow should carry a citation back to the source document and the specific clause. This is not optional. A model that returns an answer without showing which contract and which paragraph it came from is a model that cannot be safely relied upon — and, worse, is a model whose errors are invisible until they surface in a closing dispute. The good vendors solved this. The bad ones have not. The choice of tooling matters.

Privilege and Vendor Selection

The legal-engineering question of which vendor to use is now also a privilege question. Data residency, whether the vendor trains on uploaded documents, who has access to the workspace, and whether outputs are shared with non-lawyers all affect the privilege analysis in a given jurisdiction. These conversations belong at engagement-letter stage, not after the workspace has been populated. Firms that have done this work once benefit from a clean repeatable answer; firms that have not are slower and more exposed.

The Honest Net

Used well, AI in legal DD compresses the elapsed time of a mid-market deal by one to three weeks and shifts associate hours from extraction to judgement. Used badly — as an autonomous reviewer or without citations — it produces confident wrong answers that survive precisely until they matter. The leverage is real; the discipline to use it well is the differentiator.