Volume work with checkable output. This is the strongest productivity case in legal practice, and it comes with a real question about what you have verified.
Start a conversation with the AI Adoption Concierge, already scoped to document review & diligence. Pick a starting point, or describe your situation directly.
If there is one unambiguous productivity case in legal AI, this is it. Review and diligence are high volume, well specified and — crucially — the output is checkable: whether a document contains a change-of-control provision is a question with an answer, unlike whether a clause is well drafted. The task also has a long history of technology assistance, which means the profession has already worked through the validation question once with predictive coding, and the framework transfers. The hard question is not whether the tool works but what your validation actually establishes. Reviewing a sample and finding it clean tells you something precise and limited about the rest, and the size of that limit is a number worth knowing rather than assuming.
Volume and specificity at the top; judgement at the bottom.
Change of control, assignment, termination across thousands of documents. Very strong.
Parties, dates, values, terms into a table. Fast, and checkable against the source.
Sorting a large set into what needs a human and what does not. Large saving.
Medical records, transcripts, correspondence. Strong with verification of anything relied on.
Whether a found provision matters to this deal is judgement and stays with the lawyer.
Proving something is not there is harder than finding it, and the failure is silent.
How firms validate the output.
The largest available saving, and the one place where a silent miss is the real risk.
A wrongly flagged document costs a minute of someone's time. A provision the system never surfaced is invisible — nobody knows to look for it, and it is found later by the other side.
Less than the tool's confidence scores suggest and more than a fully manual process, which is a genuinely useful position to be in. The workable pattern is triage rather than replacement: the system sorts the set, humans review everything it flags plus a validated sample of what it did not, with the sample sized to support a real confidence statement. That still removes most of the volume. What does not work is accepting the system's categorisation of the unflagged remainder without sampling it, because that is precisely where the invisible failure lives.
By running it against documents whose answers you already know. Take a set previously reviewed by hand, run the system, and compare — that gives you a measured miss rate on your documents rather than the vendor's. It is the only method that produces a number you can defend, and it is more work than looking at what the tool flagged and finding it accurate, which measures the wrong thing entirely. Retain the test results; if the review is ever challenged, that is your evidence.
The framework is well-established from predictive coding, which is genuinely reassuring here: courts have accepted technology-assisted review for well over a decade where the process was documented and validated. The requirements are recognisably the same — a stated protocol, validation against known answers, sampling that supports the confidence claimed, and a record of all of it. Generative systems differ in that their behaviour is less predictable between runs, which strengthens rather than weakens the argument for documenting your protocol precisely.
It is the same question as any AI use, at a scale that makes it more consequential. Uploading a client's entire document population to a third-party system is a substantial disclosure, and whether it is permissible turns on the engagement terms, the vendor's contractual commitments on retention and training, and any obligations attached to the documents themselves — protective orders and third-party confidentiality provisions frequently bind material in a diligence set. Establish this before the documents move, not after.
Describe the review you are facing. The Institute will help you design the protocol.