What Legal AI Is Good For, and What It Is Not
Short answer: In a corporate legal function, AI creates value on work whose output can be verified and whose errors are cheap: summarising documents, producing first drafts, searching large volumes, translation for comprehension and structured data extraction. It does not create value where output cannot be verified or errors are expensive: final legal opinions, strategy decisions, commitments to clients and unverified case citations.
Asking whether AI works in law is the wrong question. The right one is which work it helps with and which it does not. The distinction comes from the nature of the task, not the sophistication of the technology, and it is the only sound basis for placing AI inside a legal function.
Two tests
Can the output be verified?
If you can check within a reasonable time whether the result is right, AI saves time. A contract summary can be verified against the contract. The correctness of a strategy is only revealed by its outcome, which is not verification.
How expensive is an error?
If an uncaught error costs little, the cost of trying is low. A wrong sentence in a draft is corrected. A wrong citation already filed with a court is not. Together these two tests form a matrix.
| Cheap errors | Expensive errors | |
|---|---|---|
| Verifiable | Use AI. Summarising, first drafts, search, translation, data extraction | Verification mandatory. Case citations, deadline calculations, amount calculations |
| Not verifiable | Use with care. Idea generation, listing alternative approaches | Do not use. Final opinions, strategy decisions, client commitments |
Five tasks where it creates value
- Document summarisation. The summary of a long expert report or case file can be verified against its source, and the time saved is directly measurable.
- First drafts. A first version of a filing or contract is a clear speed gain over starting from a blank page. The draft will be reviewed anyway, so errors are cheap.
- Search across volume. Finding where a topic appears in thousands of pages is expensive for a person and cheap for a machine. The located passage is then opened and confirmed.
- Translation. Understanding a contract or decision in another language. Translations requiring legal validity are outside this.
- Structured data extraction. Pulling dates, parties, amounts and terms from contracts into a table. Every extracted field can be checked against its source.
Four tasks where it does not
- Final legal opinion. An opinion given to a client is the lawyer's responsibility and is not a verifiable output. AI can feed into preparing it; it cannot be the opinion.
- Strategy decisions. Which case to bring, whom to settle with: these rest on information not in the file and on commercial context.
- Unverified case citations. No citation should be used without its source opened and full text read. This is a matter of rules, not technology.
- Commitments to clients. Timing, outcome or amount cannot be based on AI output.
Where to start: a twelve-week sequence
The most common mistake when introducing AI is starting with the most visible task. The right order moves from the top-left cell of the matrix rightwards.
| Stage | Scope | Success measure |
|---|---|---|
| 1. Weeks 1-4 | Document summarisation and search across volume | Measurable reduction in file review time |
| 2. Weeks 5-8 | First drafts | Drafting time; review time must not increase |
| 3. Weeks 9-12 | Data extraction and reporting | Reduction in manual data entry |
| 4. Ongoing | Case law research, under the verification rule | Share of citations verified; target one hundred per cent |
Why this order: The first stage carries the lowest risk and the most measurable benefit, which is where team confidence is built. Case law research comes last because starting it before the verification discipline is established produces the most expensive error earliest.
A written usage policy
Individual care is not sufficient at corporate scale. A minimum policy covers five points.
- Which task types may and may not use AI, stated explicitly.
- When client data must be anonymised before it reaches an assistant.
- Verification of case citations is mandatory and the verification is recorded.
- When AI-assisted output is disclosed to the client.
- How often the policy is reviewed, since regulation and professional rules change.
Frequently asked questions
Where does AI create value in a corporate legal function?
On work whose output can be verified and whose errors are cheap: document and matter summarisation, first drafts, search across large volumes, translation for comprehension and structured data extraction. What these share is that the result can be checked against its source within a reasonable time.
Which tasks should not use AI?
Those whose output cannot be verified or whose errors are expensive: final legal opinions, litigation strategy decisions, unverified case citations and commitments to clients on timing, outcome or amount. AI can feed into these tasks but cannot be the decision itself.
Will AI replace lawyers?
The question is framed wrongly. The distinction is drawn per task, not across the profession. On verifiable work with cheap errors AI saves time; on unverifiable work with expensive errors accountability stays with the lawyer, so it cannot take over. In practice this shifts a lawyer's time towards the second group.
Where should a legal function start with AI?
With document summarisation and search across volume. These carry the lowest risk and the most measurable benefit, and team confidence is built there. Case law research should come last, because starting it before the verification discipline is established produces the most expensive error earliest.
Why does a firm need a written AI usage policy?
Individual care does not scale. A minimum policy states which task types may and may not use AI, when client data is anonymised, that case citation verification is mandatory and recorded, when use is disclosed to clients, and how often the policy is reviewed.
Should clients be told that AI was used?
This should be settled in the firm's policy and depends on professional rules and contractual commitments. The important thing is that the policy takes a position rather than leaving it open, and that it is reviewed as rules change.