JTJuris Tdesk Pro
Guides › Legal AI

Seven Ways Law Firms Get Generative AI Rollouts Wrong, and the Fix for Each

Most disappointing legal AI projects don't fail at vendor selection. They fail at rollout: idle seats, unchecked output and client terms nobody read. Here are seven common mistakes and the fix for each.

Juris Tdesk Pro editorial team · November 2026 · 9 min read

Why the vendor usually isn't the problem

Most firms that are unhappy with generative AI for law firms didn't buy the wrong product. They bought a reasonable one and then deployed it badly. Licenses went out before anyone agreed what the tool was for, and nobody wrote down how its output should be checked. Six months later the managing partner asks what the spend bought, and nobody can answer.

This piece doesn't compare vendors. It covers seven rollout errors and a fix for each. They apply just as much to a 15-lawyer boutique in Nairobi as to a 40-person in-house team in Frankfurt. If your contract is already signed, most of the fixes still work. If it isn't signed yet, deal with the first two before you negotiate, because they decide how many seats you need and what you should ask the vendor for.

Advertisement

Mistake one: buying seats before naming the use cases

Vendor pricing nudges firms toward licensing everyone at once. Volume discounts kick in at higher seat counts, and a firm-wide deployment looks decisive. Here are typical 2026 market ranges, which vary widely by contract and region. General-purpose assistants bundled into productivity suites usually cost about $20–$40 per user per month. Legal-specific tools for research, drafting and review often cost $100–$400 or more per user per month, frequently with annual minimums. Local and enterprise pricing differs, so confirm current figures with each vendor.

Hypothetical example: a 60-lawyer firm buys 60 legal-specific seats at $150 a month, which comes to $108,000 a year. If only 18 lawyers use the tool regularly, the real cost is about $500 per active user per month, and the firm still can't say which work got faster.

What to do instead:

  • Name three to five specific use cases, such as first-draft NDAs, deposition or witness-statement summaries, or due diligence issue lists. Give each one a partner owner, an estimated monthly volume and a baseline for how long the task takes today.
  • Pilot with 10–20% of fee earners for 60–90 days. Pick them from the practice groups that own those use cases.
  • Buy the rest in waves, each tied to an adoption threshold. Negotiate the right to reallocate or reduce seats at every renewal.

Mistake two: no protocol for verifying output

Courts in several common-law jurisdictions, including the US and England and Wales, have publicly criticized or sanctioned lawyers for filing fabricated citations produced by generative tools. In the reported cases, the process failed, not the technology. Someone assumed someone else had checked.

"Lawyers must review all output" isn't a protocol. A protocol says who checks what, against which source, and where the check is recorded. Scale the depth of review to the risk:

  • Tier 1, internal use only (brainstorming, issue spotting, internal summaries). The user reads the output critically. Nothing leaves the firm without moving up a tier.
  • Tier 2, client-facing (advice notes, contract drafts, correspondence). The drafting lawyer traces every factual statement and legal proposition to a source document or authority, and a supervising lawyer reviews the work.
  • Tier 3, filed or relied upon (court submissions, opinions, regulatory filings). Every citation is pulled from a primary source and pinpoint-checked, and every quotation is verified word for word. A named lawyer records sign-off in the document management system.

Keep that record. If a client or court asks how a document was checked, a logged sign-off is a far stronger answer than someone's memory.

Mistake three: an AI usage policy nobody can apply

First-draft policies tend to become one of two things: a blanket ban that people quietly ignore, or a 20-page document that reads like a regulatory consultation. Neither changes behavior. A workable AI usage policy for a law firm fits on one or two pages and answers the questions a junior associate has at 7 p.m.:

  • Which tools are approved, and which are prohibited, including personal accounts on consumer chatbots?
  • Which kinds of data can go into each approved tool: public information, anonymized facts, client-confidential material, privileged material, special-category personal data?
  • Which uses are off-limits whatever the tool?
  • Which verification tier applies to which kind of output?
  • When must AI use be disclosed to a client or court?
  • Who do I tell if something goes wrong?

Behind the short policy, keep a longer legal AI governance standard for the risk committee and IT. It covers vendor due diligence, logging, retention, monitoring for model changes, and review cycles. Review both documents quarterly in the first year. Tools, client demands and regulatory guidance are all still shifting, and an annual review will lag behind them.

Mistake four: ignoring what clients have already said

A growing number of institutional clients, especially banks, insurers and large corporates, now address generative AI in their outside counsel guidelines. Common terms include:

  • prior written consent before AI is used on their matters
  • bans on entering their data into tools that train on inputs
  • disclosure requirements
  • billing rules, such as no charging for AI subscriptions or for time the tool saved

Roll out firm-wide without checking these terms and you can be in breach on the first day.

What to do instead:

  1. Audit the guidelines and engagement terms of your top 20–30 clients by revenue. Log every AI-related clause in one register.
  2. Add a client-level AI flag at matter intake, so restrictions follow the matter automatically instead of relying on someone's memory.
  3. Update your standard engagement letter to explain how the firm uses AI, what safeguards apply and how billing works.
  4. Settle your billing position before clients raise it. On fixed-fee and capped work, efficiency gains go to the firm. On hourly work, expect clients to ask how the savings are shared.

Have professional responsibility counsel review the engagement-letter changes and billing approach in every jurisdiction where you practice.

Mistake five: treating conduct and data rules as one jurisdiction's problem

Firms with offices or clients in several countries often adopt their head-office regulator's guidance and assume it covers everywhere else. It doesn't. In the US, ABA Formal Opinion 512 and a growing number of state bar opinions address competence, confidentiality, supervision and fees. In England and Wales, the SRA, the Law Society and the Bar Council have each published guidance. In the EU, the AI Act's phased obligations apply on top of GDPR rules on processing and cross-border transfers. India's Digital Personal Data Protection Act, Brazil's LGPD, Saudi Arabia's PDPL and Singapore's PDPA each add requirements of their own. Some courts, including Singapore's, have also issued guidance on how court users may use generative tools.

What to do instead: build a short jurisdiction matrix that covers every office and every major client location. For each one, record:

  • the professional-conduct guidance that applies and any court disclosure rules
  • data protection and transfer requirements
  • whether the vendor offers in-region hosting. This is increasingly available in the EU, the UK, Australia and parts of the Middle East and Asia, at varying cost
  • a named local partner who keeps the entry current

The matrix maps risk. It isn't a legal opinion, so have local regulatory counsel confirm the obligations in each jurisdiction.

Mistake six: one launch webinar instead of real AI training for lawyers

The usual pattern is a vendor demo, a recorded webinar and a login email. Usage typically peaks in the first week or two and then fades. Lawyers who never see the tool work on their own kind of matter go back to the old way.

AI training for lawyers works when it's tailored to each role and keeps going after launch:

  • Partners: supervision duties, verification tiers, client terms, and how to review AI-assisted work product.
  • Associates: hands-on practice with the named use cases, prompting on approved matter types, and spotting fabricated or outdated authority.
  • Business services and support staff: data handling, approved tools and incident reporting.

For the first six months, run a 30–45 minute session for each practice group every month. Keep a shared library of prompts that have been tested. Appoint one champion per group and give them protected time for the role. Make finishing a core module a condition of getting a seat. Enterprise contracts often include vendor onboarding and customer-success support, so get specific training hours written into the contract during negotiation.

Mistake seven: no measurement plan, so renewal becomes a guess

If you don't capture baselines before launch, the renewal decision rests on stories from the loudest users. Decide what you'll measure before the pilot starts, and record baselines in its first two weeks.

MetricHow to measureRed flag at day 90
Active use per seatVendor admin dashboard: weekly active users divided by licensed seatsBelow roughly 40% weekly active
Time per task on named use casesTime entries or sampled self-reports, compared with the pre-launch baselineNo measurable change on any use case
Verification defect rateErrors caught at Tier 2 and Tier 3 review, logged by typeA rising rate, or any error that reaches a client or court
Matter economicsWrite-offs and realization on fixed-fee or capped matters that use the toolFlat margins while subscription costs rise
Policy complianceSpot audits of the data categories entered; incidents involving unapproved toolsRepeated use of personal consumer accounts

Report these figures to the management committee monthly during the pilot and quarterly after that. They're also your strongest leverage at renewal. A vendor negotiates very differently when you can show that only a third of your seats are in use.

A 120-day sequence that avoids all seven

For most firms of 20 to 200 lawyers, getting the order right matters more than moving fast. Settle governance and client terms before you scale, and measure before you expand.

PhaseWeeksKey outputs
Define1–3Named use cases with owners and baselines; client-terms register; jurisdiction matrix
Govern3–5One-page usage policy; legal AI governance standard; verification tiers; updated engagement letter
Pilot6–1610–20% of fee earners; role-based training; monthly metrics report
Decide17Expand, reconfigure or exit, based on the metrics table

The decision rule: don't buy another seat until the pilot group meets all three conditions.

  • Weekly active use is above roughly 60%.
  • At least two named use cases show a measurable time saving.
  • No verification failure has reached a client or court.

If the pilot falls short, rework the use cases or the training before you blame the vendor.

Quick answers

How many seats should a firm buy for its first generative AI deployment?

Start with 10–20% of fee earners, drawn from the practice groups that own your named use cases. Run the pilot for 60–90 days against defined metrics, then expand in waves. Negotiate seat reallocation and reduction rights at the start so a small pilot doesn't lock you into a firm-wide minimum.

Do we have to tell clients we use AI on their matters?

It depends on three things, and each varies by jurisdiction: the client's engagement terms, your professional-conduct regulator and, for court work, local court rules. A growing number of institutional clients require consent or disclosure in their outside counsel guidelines. Audit your key clients' terms, update your engagement letter, and have professional responsibility counsel confirm what applies wherever you practice.

What should an AI usage policy for a law firm include at minimum?

At minimum it should cover approved and prohibited tools, which data categories may go into each tool, uses that are off-limits, verification requirements by output type, disclosure rules and incident reporting. Keep the version lawyers read to one or two pages. Put the detailed vendor, logging and retention controls in a separate governance standard.

How do we know if our legal AI tools are worth renewing?

Compare day-90 and renewal-period data with the baselines you captured before launch. Look at weekly active use per seat, time per task on named use cases, verification defect rates and margins on fixed-fee work. If utilization is low and no use case shows a measurable time saving, either negotiate down to fewer seats or rework the training and use cases before you commit to another term.

Written by the Juris Tdesk Pro editorial team. Prices, features and rules change often - confirm details with the vendor or a licensed professional before you decide.