Agentic AI in legal work: what the numbers show
Adoption, cancellation rates, the end of the hype: what Thomson Reuters, Gartner and Forrester say about agentic AI, and what follows from it for law firms.
The available figures show two movements at once. Firm-wide AI tools are spreading fast across the legal industry, and market watchers forecast that a substantial share of the projects built on agentic systems will be shut down again. Both are plausible, and read together they give a usable frame for the agent narrative.
What makes a system agentic
The term describes a system that does not answer a single query but breaks a task into several steps, plans those steps itself and calls tools along the way: it reads files, starts a search, calculates, checks an interim result and then decides what to do next. The difference from a chat window is not the quality of the model but the handover of control over the sequence.
That handover is the point at which the assessment is decided, in both directions.
Adoption is growing, and measurably so
In its 2026 AI in Professional Services Report, the Thomson Reuters Institute measures the share of law firms and legal departments using a firm-wide GenAI tool. It rose from 14 percent in early 2024 to 43 percent two years later. That is no longer a fringe phenomenon but a movement that has reached a large part of the professional market.
What a firm-wide tool means in daily work is something the number does not say. It counts provision, not use, and the use that runs alongside it shows up in no statistic.
The counter-forecasts are just as specific
On 25 June 2025 Gartner forecast that more than 40 percent of agentic AI projects (ventures involving programs that carry out several work steps on their own) will be abandoned by the end of 2027, because of rising costs, unclear business value or inadequate risk control. The analyst responsible is Anushree Verma; the basis is a January 2025 survey of 3,412 participants in a webinar. That basis is not a representative sample of the industry, which does not make the statement worthless but gives it the character of a measurement of expectations.
On 28 October 2025 Forrester wrote in its 2026 Technology & Security Predictions, under the responsibility of Sharyn Leaver, that the AI hype phase would end in 2026. That too is a forecast, not a finding. What is notable is that it comes from the same environment that supplies the adoption figures.
Why the two do not contradict each other
Adoption figures measure purchase, not effect. A firm-wide rollout is a procurement decision; whether it changes the work is a separate matter. It is in exactly that gap that the projects arise which later end up on the cancellation list.
They rarely fail because of the model. They fail because nobody keeps control over the result, and because the task was never clearly defined. Anyone who cannot say in advance what result an agent is supposed to deliver, and how you would recognise that it is correct, still cannot say it twelve months later. When budgets tighten, ventures like that are the first to be up for discussion.
What follows from this for law firms
Autonomy is not an end in itself. In legal work, traceability is usually worth more than independence, because someone signs off on the result at the end and that responsibility cannot be delegated. A result whose origin nobody can reconstruct is worthless for a sign-off, even when it is correct.
An unspectacular consequence follows. As long as a task is stable, a fixed sequence with clearly defined steps beats the freely planning agent. Deadlines, conflict checks and fee calculations are tasks of that kind. For them, the simplest fitting architecture is regularly the right one.
Where agentic systems do carry the weight
Some tasks do not allow the number of steps to be fixed in advance without the choice becoming arbitrary. Research across a large body of case files is one of them: it is not settled beforehand how many documents need reading, which lead is worth following and when enough material is together. A rigid sequence would have to guess those decisions, while an agentic system makes them as it goes.
The condition stays the same. The system delivers sources and intermediate steps; judging the result remains with the lawyers. Under that condition the agentic design is not a gamble but the right shape for a particular kind of task.
That lets all three figures be placed without taking sides. Adoption says nothing about benefit, the cancellation rate nothing about the technology. What decides both is whether the task was clear before the project started.
The content of this blog was created partly with the support of AI.