EU AI Act transparency duties apply now; high-risk duties from December 2027. Check your exposure
Insights About us Careers
Contact us
Professional Services

AI for Law Firms and Legal Teams

Two thousand and twenty-two recorded court decisions now involve fabricated AI content, and almost all of them share the same missing step.

Tier 1
Our depth here
2,022 cases
As at 5 Sept 2026
Provenance
On every claim

Legal work is the most documented AI failure case in any profession, which is unusually useful: the failure mode is known precisely, and it is avoidable by design rather than by hoping for a better model.

In one paragraph

AI for law firms covers document review and extraction at volume, retrieval across precedent and prior advice with provenance, drafting constrained to firm material, contract analysis, matter and pricing analytics, and the verification workflow required where output is filed or advised under a professional's name.

What the 2,022 cases actually show

The AI Hallucination Cases database recorded 2,022 decisions worldwide as at 5 September 2026 in which a court found that a party relied on fabricated AI content. Read carefully, it is a study of process failure rather than a case against the technology.

  • The distribution is wide. 1,379 in the United States, 217 in Canada, 110 in Australia, 69 in the United Kingdom, with entries from more than forty other jurisdictions.
  • Fabricated case law dominates. Around 1,677 instances of invented authority, 845 of misrepresented content and 547 of false quotation — all failures of open-ended generation with nothing to ground it.
  • Self-represented litigants outnumber lawyers. 1,163 against 805, which is the detail most commentary omits and the one that identifies the real variable: whether anyone checked.
  • Sanctions are mostly procedural, not ruinous. Warnings, fines recorded between $1 and $14,500, bar referrals, struck briefs and adverse costs. The reputational cost exceeds the financial one.
  • The regulatory hook is the old rule. Courts have consistently held that the duty to verify what you file is unchanged by how it was drafted.
Worth knowing

Constrain the generation and the failure mode largely disappears

Every one of these failures involves a model asked to produce authority it was not given. A system that drafts only from documents you supplied, cites only sources it can point to in your own corpus, and marks anything it cannot support, does not invent cases — because it has nowhere to invent them from. Add a verification step that a person actually performs, and you have the productivity without the exposure. This is a design decision, and it should be made before tool selection rather than after an incident.

Document review, where the value is largest and least discussed

Volume review is the strongest case in the profession

Disclosure, due diligence, contract portfolios, regulatory response — large document sets where the task is finding and extracting specific facts. The errors are checkable against the source document, which is precisely what makes it safe.

Recall matters more than precision

A missed privileged document or a missed change-of-control clause is far worse than a false flag a reviewer discards. Tune for recall, review the flagged set, and evaluate on what was missed rather than on aggregate accuracy.

Extraction with per-field confidence

Dates, parties, governing law, termination rights, liability caps and assignment provisions into structured fields, with anything uncertain routed to a reviewer instead of silently averaged into a report.

The economics need deciding before the build

Where review is billed hourly, faster review reduces the fee. Firms that have thought this through have usually moved the affected work to a fixed or capped basis first. See contract analysis.

Retrieval and drafting over the firm's own material

ApplicationFitDesign requirement
Precedent and know-how retrievalStrongProvenance on every result; superseded material marked
Prior advice search across mattersStrongConfidentiality walls between client teams enforced in the index
First-draft generation from precedentStrongConstrained to firm templates; no open-ended authority generation
Contract review against a playbookStrongDeviations flagged for a lawyer, not resolved automatically
Deposition and transcript analysisGoodRetrieval and summarisation with citation to the line
Legal research into authorityModerateOnly over a verified corpus, and never as the final check
Client-facing advice without reviewNot credibleThe liability does not transfer to the model
Worth knowing

Confidentiality walls have to be in the index, not in the policy

A retrieval system over a firm's matter files can surface one client's material to a team acting against them, which is an ethical breach before it is a technical bug. The boundaries must be enforced in the index and the access layer, mirroring the firm's information barriers, rather than stated in an acceptable use policy. This is the single most important design decision in a law firm knowledge system, and it is the one most often deferred to phase two.

Process

How an engagement runs

Boundaries and verification designed first, because neither can be retrofitted.

Weeks 1 to 3

Scope, ethics and client terms

Confidentiality boundaries, client engagement restrictions, and where the verification step will sit.

Weeks 4 to 8

Corpus and index build

Provenance-carrying retrieval with information barriers enforced and superseded material marked.

Weeks 9 to 14

Review or drafting build

Extraction tuned for recall, or drafting constrained to firm precedent with citations to your own corpus.

Weeks 15 to 18

Trial

On live matters against fee-earner judgement, with missed-item rate measured specifically.

Ongoing

Operation

New matters indexed, barriers maintained, and the verification step audited rather than assumed.

Deliverables

What you receive

The firm's knowledge made findable, and generation that cannot invent authority.

01

Document review and extraction

Tuned for recall, with per-field confidence and uncertain items routed to a reviewer.

02

Precedent retrieval

Provenance on every result, superseded material marked, information barriers enforced in the index.

03

Constrained drafting

First drafts from firm templates and supplied documents, citing only your own corpus.

04

Playbook contract review

Deviations from standard positions flagged for a lawyer to resolve.

05

Verification workflow

A checking step built into the process, with an audit trail rather than an assumption.

06

Matter and pricing analytics

Where realisation actually goes, by matter type, client and fee basis.

Fit check

Is this the right starting point?

Worth being direct. There are situations in legal and law firms where custom AI work is the wrong spend, and those are listed rather than buried.

Worth doing if

  • Document review at volume is a material cost and reviewers are the constraint.
  • Fee earners routinely rediscover advice the firm has already given.
  • Precedent exists but only the author can find the right version.
  • You want generation and need it constrained so it cannot invent authority.
  • Client engagement terms restrict AI use and nobody has mapped which clients.

Do something else if

  • Information barriers cannot be represented in an index and there is no route to fixing that.
  • You want client-facing advice produced without professional review.
  • Superseded and current precedent cannot be distinguished in the document system.
  • The commercial model cannot absorb faster review on hourly-billed work and nobody wants that conversation.
Questions

Frequently asked questions

Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.

How do we use generative AI without ending up in the sanctions record?

Constrain what it can draw on. Every one of the 2,022 recorded cases involves a model asked to produce authority it was never given, which is exactly what a system restricted to your own corpus cannot do. Require provenance on every factual claim, mark anything the system cannot support, and put a verification step into the workflow rather than into a training slide. It is worth knowing that self-represented litigants account for more of those cases than lawyers do — 1,163 against 805 — which shows the variable is whether anyone checked, not whether AI was used.

Is legal research a good application?

Only over a verified corpus, and never as the last step. Research is precisely where open-ended generation invents authority, because the task is to produce citations and the model will produce them whether or not they exist. Retrieval over a licensed and verified database with citation to the source is a different and defensible thing. What no configuration removes is the professional obligation to check the authority before it is filed, and courts have been unambiguous that this duty does not change because a machine drafted the brief.

Can we build search across all our matter files?

Yes, and the information barriers must be enforced in the index rather than in a policy. A retrieval system over matter files can surface one client's confidential material to a team acting against them, which is an ethical breach with regulatory consequences, not a bug to fix in the next sprint. Access needs to mirror the firm's existing barriers, and the firm needs to make an explicit decision about which material becomes general know-how. That decision belongs to the general counsel and the risk partner, not to the build.

What is the strongest application in a law firm?

Document review at volume, and it gets far less attention than drafting. Disclosure, due diligence and contract portfolio work involve finding specific facts in large document sets, the errors are checkable against the source, and the throughput gain is large and measurable. Tune for recall rather than precision — a missed privileged document costs far more than a false flag a reviewer discards — and evaluate on what the system missed. It is the safest high-value application in the profession.

Will this reduce our revenue?

On hourly-billed work, yes, and it is better modelled than discovered. Faster review on an hourly basis means a smaller fee, and firms that have thought it through generally move the most-affected work to a fixed or capped basis before deploying, or use the capacity for work that was previously uneconomic to take. What does not work is deploying the efficiency and hoping the commercial model absorbs it quietly; the realisation shows up about two quarters later and by then the pricing conversation is harder.

Tell us what the problem looks like.

Thirty minutes, no charge, no deck. We will tell you whether this is an AI problem, a data problem, or a process problem — and we will say when the honest answer is to buy something rather than build it.