AI for Legal: What Law Firms Are Automating Right Now

Four categories. Three of them ship.

Across the engagements gamgi has shipped in legal and the production deployments we’ve audited at other firms, four use cases get pitched as AI for law firms: matter intake and triage, structured document drafting, contract review, and AI-assisted due diligence. Three of those reliably ship in 2026. One still doesn’t, outside narrow specialist pilots. The ROI rank order is roughly the inverse of the vendor-pitch order.

Legal AI automation is in production. The ranking isn’t what the vendor decks say.

The standard pitch for legal AI used to be “this will transform the practice of law in five years.” In 2026 the framing has caught up. The Thomson Reuters Institute’s 2024 Future of Professionals report documents the same shift: legal AI is in production at hundreds of firms, often quietly, often in workflows that nobody photographs for a vendor case study. The shift is structural. Foundation-model quality crossed a threshold that makes document-heavy back-office work tractable in ways generic NLP couldn’t manage in 2022. Compliance-conscious vendors built the auditability infrastructure that regulated practice requires. The technology cleared the bar.

What hasn’t caught up is the ranking. Vendor decks lead with the dramatic categories (contract review at scale, autonomous legal research, generative drafting) because those are the ones that demo well. The legal AI automation that’s actually in production is duller. The single highest-ROI category at the firms gamgi has worked with is matter intake: the structured capture of new client information, conflicts checking, and routing to the right partner. It’s a paperwork problem dressed as a process problem, and current models are exceptionally good at paperwork problems with structured outputs.

This article catalogues the four categories, ranks them by what actually ships, and names where the line falls. The bottom line: most law firms in 2026 are deploying legal AI in the third or fourth category when they should be starting with the first.

From matter intake to AI contract review: the four categories, ranked by ROI

In production order (highest reliability and ROI first, most overhyped last):

1. Matter intake, conflicts checking, and routing. Highest ROI, least glamorous. New client information arrives by email, web form, phone transcript, or referral note. The AI extracts structured fields (parties, matter type, jurisdiction, urgency), runs conflicts against the firm’s existing client database, routes to the partner whose specialism fits, and produces a first-draft engagement letter. Time saved per matter: 30-90 minutes of paralegal work. Multiply by hundreds of new matters per year and the unit economics dominate. Typical project €15-30K, payback in one to two quarters.

2. Structured document drafting. Engagement letters, retainer agreements, standard NDAs, basic wills, standard property conveyancing documents. The AI takes a structured input (matter type, party details, jurisdiction, a few elections) and produces a first draft against the firm’s templates. The lawyer reviews and edits; the lawyer doesn’t write from scratch. Time saved per document: 40-80% of drafting hours. Works because the document space is bounded and the firm’s templates already encode the style. Typical project €20-40K.

3. AI contract review with bounded scope. Peer-reviewed work on contract review benchmarks (CUAD) confirms the pattern: model accuracy on bounded clause-extraction tasks is materially higher than on open-ended contract analysis. AI contract review works when the firm specifies three things: what category of contracts (NDAs, supplier agreements, lease addenda are common starting points; “all contracts” is not), what specific clauses to surface (limitation of liability, IP assignment, change-of-control; not “anything unusual”), and what the human reviewer does with the output. Within those constraints, AI contract review ships and saves time. Outside them (the open-ended “review this 200-page agreement” framing), accuracy degrades and the lawyer ends up re-reading the contract anyway. The firms getting ROI on AI contract review are running it in narrow, repeatable contexts.

4. AI due diligence and autonomous research. Still mostly theatre in 2026. Demos look impressive; production deployments are rare outside a few specialist transaction practices. The failure mode is consistent: the AI surfaces things that look interesting; the lawyer can’t validate them without re-doing the work; trust collapses; the tool gets shelved. The category will ship eventually, probably 2027-2028. It is not where a firm starting its AI work in 2026 should be spending its budget.

Together, these four categories define what AI legal tech actually means in mid-2026 practice. The ranking matters because firms that start with category 4 spend twelve months proving the technology isn’t ready; firms that start with category 1 ship something in eight weeks and learn what their next category should be.

What category-1 AI legal tech looks like in production

Two production references locate where the line falls between categories that ship and categories that don’t.

LexAlert: legislative monitoring as continuous category-1 work. LexAlert isn’t strictly a law-firm tool. It’s a legislative-monitoring product for legal teams that need to track regulatory change across multiple jurisdictions. But the underlying architecture is exactly category-1 paperwork-automation work: official-gazette documents arrive, the system classifies them by client matter, branching logic decides which partner gets alerted with what priority, and every classification decision is auditable for compliance. The system shipped to production because the scope was narrow, the action space was bounded (no autonomous writes outside the alert system), and the auditability requirement was treated as a first-class scoping constraint rather than a bolt-on. Full structural detail in the LexAlert case study.

The lesson generalises directly. The matter-intake and conflicts-checking workflows that dominate category-1 in a law firm have the same architectural shape: documents arrive, get classified, get routed, every step is logged. They ship because the operational requirement is well-defined and the AI’s job is bounded.

The category-3 cases that worked. We’ve seen contract review ship at a small commercial firm that runs a high volume of supplier agreement reviews for SME clients. The firm’s deployment used AI to extract the seven clauses they care about for that contract type and produce a structured report against the firm’s standard playbook. The lawyer reads the structured report, not the contract. Throughput went from six contracts per associate-day to eighteen. The deployment took four months, cost about €35K, and shipped because the scope was narrow on purpose: one contract type, one playbook, one reviewing-lawyer workflow. The firm explicitly declined to extend it to general contract review, because that is where the technology stops being reliable.

A structured audit is what surfaces which category your firm should start with, because the answer depends on the volume mix of your matter types and the structure of your existing templates. The wrong category-rank assumption is the most common reason legal pilots stall, and the pattern matches the wider failure mode in pilots that don’t make it to production. For the engagement-architecture question of how a legal AI project should be staffed and scoped, the audit-first process carries the structural detail.

When AI legal tech is the wrong investment

The framing here pushes toward category-1 and category-2 work. There are cases where the answer is to not do AI work at all, or to wait:

• Solo practitioners with low matter volume. If you bring on six new matters a month, the absolute time saving on intake is two to three hours per month. The infrastructure cost (vendor or build, plus integration with practice management) doesn’t pay back at that volume. Wait until the firm hits the threshold where the unit economics work: typically 30+ new matters per month for category-1, higher for category-2.
• Firms with bespoke practice areas. If 80% of the work is unique and unstructured (high-end M&A, complex commercial disputes), the document space isn’t bounded enough for category-2 to work. Category-1 still applies because intake is universal; skip category-2 and category-3.
• Firms without owned templates. Category-2 drafting works because the firm’s existing templates encode the style and risk posture. A firm whose drafting is partner-by-partner ad hoc will produce inconsistent first drafts, and the AI will multiply the inconsistency. Standardise templates first; deploy AI after.
• Firms that haven’t mapped the workflow. Deploying any AI tool against a workflow you haven’t mapped is a confidence-trick on yourself. The diagnostic framing in do you need an AI audit is the right starting point for a firm that hasn’t done this yet.