Most legal AI pilots do not fail during the demonstration. They fail afterwards, for reasons that have little to do with the tool, and one of them only law firms have. How to tell which one stopped yours, and where to restart.

AI Answer Summary

Most legal AI pilots fail after the demonstration rather than during it. Across industries, Gartner predicted in 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, and later reported that at least 50% were.

In law firms there are four common reasons. Nobody took a baseline, so nobody could prove the pilot worked. The firm already owned AI it had never switched on. Partners never adopted the tool. And on hourly matters, a pilot that succeeds reduces the invoice, which leaves nobody with a reason to keep paying for it.

The fix is to diagnose which of those it was before changing anything, take the baseline that was skipped, and restart on work the firm was already writing off.

Why Legal AI Pilots Fail After the Demo

A legal AI pilot almost never fails in the demonstration, because a demonstration is designed to work. It fails in the months afterwards, when the tool meets the firm's billing, its partners and its lack of measurement.

The general picture is worse than most firms expect. In July 2024, Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs or unclear business value. In January 2026 it reported that at least 50% had been abandoned by then, for the same reasons, based on its analysis of hundreds of implementations. The prediction was an underestimate. Those figures cover every industry, not law firms, but the direction is hard to ignore.

Gartner's causes are general, and three of these four apply to any organisation. The first applies only to firms that bill by the hour.

Four Reasons a Legal AI Pilot Fails

They are ordered here by how rarely they get discussed, starting with the one only law firms face.

The Pilot Worked, and the Invoice Shrank

This is the failure no generic account of AI pilots can describe. ABA Formal Opinion 512 requires hourly work to be billed at the time spent. A pilot that makes billable work faster therefore reduces what the firm collects on that work.

Follow it through. The partner whose matters sped up has lost revenue. The partner who sponsored the pilot now holds a success that nobody wants to pay for. The firm reviews it at renewal and concludes the pilot produced no value, when the value went to the client in the form of a smaller bill.

Gartner has seen the general version of this. Its own analysis notes that rising costs kill projects even when they are technically successful and delivering value to users. In a law firm, the cost that does the killing is lost billing.

Some clients sharpen it further. The Hartford's panel counsel guidelines bar firms from charging the carrier for AI subscriptions or licences, so on that work the tool cost comes out of margin as well. Our article on AI for insurance defense firms covers those rules.

A caveat belongs here. The mechanism follows directly from the billing rules, but no survey we know of measures how often it ends pilots. This is our reading, and a firm should test it against its own experience.

Nobody Took a Baseline Before the Pilot Started

Only 18% of respondents in Thomson Reuters' 2026 AI in Professional Services report said their organisations collect ROI metrics on AI, according to the North Carolina Bar Association's summary of the major surveys.

A pilot without a baseline cannot be shown to have worked. It can only be defended on enthusiasm, and enthusiasm rarely survives a budget review. This is Gartner's "unclear business value," measured as something organisations do, or fail to do, rather than asserted as an abstract cause. It is also the failure every other article in this series warns about, which is why it is worth saying once more: take the baseline first.

The Firm Already Owned AI It Never Switched On

Some failed pilots were new purchases made while existing capability sat unused.

The International Legal Technology Association's 2026 survey of 508 firms, reported by LawSites, found 94% of firms using or exploring generative AI, up from 80% a year earlier. The most-used tool by a wide margin was Microsoft 365 Copilot, at 76% of firms, ahead of every product built specifically for lawyers. Most firms already pay for Microsoft 365. And as our article on AI legal research notes, firms on Westlaw or Lexis now receive research AI inside subscriptions they already hold.

None of that means the tool a firm already owns is the right one. It does mean a failed pilot is a good moment to list what is already licensed before signing anything new.

Partners Never Adopted It

Thomson Reuters Institute's 2026 survey of lawyers identified as stand-outs by their own clients found that more than three-quarters say their firm has an AI strategy, and fewer than half are confident in their practice area's ability to succeed as AI becomes more embedded. Even among heavy users, only about a third had discussed AI with most of their clients. And giving partners more time to experiment was unlikely, on its own, to raise adoption. The report's own framing is the useful one: where partners adopt unevenly, the strategy stays a leadership document and never becomes client experience.

Gartner describes the same pattern across industries. Without change management, even technically excellent tools see minimal adoption, and usage falls over time.

The fair counterweight: sometimes the partners are ready and the firm is not. Bloomberg Law reported in June 2026 that at one large firm, the brake on adoption was the firm's own capacity to vet tools, train people and put governance in place, rather than any reluctance among its lawyers.

How to Tell Which Failure Your Pilot Had

Each failure leaves a different trace, and the fix for one does nothing for the others.

Most failed pilots show more than one of these, and the order in which you fix them matters. A firm that solves partner adoption without a baseline ends up with more users and still no evidence. Start with the baseline, because every other fix needs it in order to prove it worked.

If the last row sounds familiar, our article on law firm AI policy sets out the decisions a policy has to make, using a federal court's account of where one firm's policy fell short.

One practical way to run the diagnosis is to ask the same question of three people: the partner who sponsored the pilot, an associate who used it, and whoever reviews the firm's billing. Their answers rarely agree, and the disagreement usually tells you which row you are in.

What to Do After a Legal AI Pilot Fails

Diagnose before you change anything, then restart smaller rather than bigger. Five steps, in order.

1. Work out which failure it was. Use the table above. Restarting without knowing is how firms run the same pilot again at twice the cost.
2. Do not relaunch at a larger scale. A second, bigger pilot with the same missing baseline fails the same way, more expensively and more visibly. The instinct after a failure is to prove the doubters wrong at scale. Resist it.
3. Take the baseline you skipped. Record hours per task, per matter, for the group that will use the tool, before anything is switched back on. Our guide to AI for mid-market law firms covers what to measure.
4. Check what you already own. Microsoft 365, the document management system, and any research subscription. The capability the pilot was supposed to deliver may already be licensed.
5. Move the pilot to work the firm is already writing off. This is the step that answers the first failure. On work the firm was never billing for, a pilot that succeeds improves the firm's economics instead of shrinking an invoice. The last section below names where that work usually sits.

The sequence is deliberately unglamorous. Each step lowers the chance of a second failure, rather than raising the stakes of the second attempt.

Should You Switch AI Vendors After a Failed Pilot?

Usually not yet.

The first question most firms ask after a pilot fails is which tool to try instead. Three of the four failures in this article have nothing to do with the tool. A new vendor inherits the missing baseline, the partners who never adopted, and the billing economics that made success unaffordable.

Switch when the diagnosis points at the tool itself: the output was wrong for your practice area, it could not reach the systems where your work lives, or its terms fail your clients' or carriers' requirements on confidentiality and data retention.

Otherwise, a switch mostly buys a fresh round of enthusiasm and the same ending. If your diagnosis does point at the tool, our comparison of AI legal assistants sets out how the main categories differ.

Where to Restart a Legal AI Pilot: Work Already Being Written Off

Restart where a success improves the firm's economics and a mistake gets caught before it leaves the building.

The third one suits a restart especially well, for a reason that answers the second failure directly. The baseline already exists. A firm's own history of pre-bill write-downs is the measurement, which means the pilot starts with the evidence most failed pilots never collected.

All three also share a property that makes a second attempt safer than the first. A mistake in triage, intake or a billing narrative gets caught by someone inside the firm, before a client, a carrier or a court sees it. After a pilot has already failed once, that margin for learning matters more than the size of the saving. A firm that restarts somewhere it can afford to get things wrong is far more likely to reach the point where it can show the tool worked.

How Codebridge Works with Mid-Market Law Firms

We build the three workflows in the table above: document review triage, intake and conflicts screening, and billing narrative cleanup. Each one starts by recording the baseline, because that is the failure most pilots never recover from.

One workflow goes live in three weeks, wired into the systems the firm already runs, with the approval checkpoint designed in during the first conversation. Your firm owns the repository, the prompts and the configuration from day one. We do not sell seats, so a restart with us does not add another licence to the list of things the firm is paying for and not using.

The closest reference we can offer, labelled for what it is: Knowledge Cloud, built for a Big Four tax and legal practice, runs an expert review queue with an immutable audit log, so a senior practitioner approves each output before the firm acts on it. A research platform rather than a law firm system, and what it demonstrates is the review pattern.

Our founding team spent more than a decade at KPMG.

If you want to work out which failure your pilot had, book a 20-minute call.