I spend most of my working week talking to the people building legal AI: founders, CEOs, the engineers and product leads who sit one layer down. I spent the years before that inside the profession those tools are sold to, as a competition lawyer moving between jurisdictions and watching software get bought, praised, and eventually abandoned. With enough time spent observing both sides, a pattern has started to surface.
Everyone has declared victory. And almost nobody is winning yet.
Look at the headline number and the celebration seems earned. The 2026 Legal Industry Report from 8am, drawn from more than 1,300 legal professionals, found that 69% of legal professionals now use general-purpose AI tools for work. This is more than double the 31% recorded for 2025, and a long way from the 27% recorded in 2024. For a profession that treats novelty with suspicion as a matter of professional instinct, that is a genuinely startling rate of change. Three years, and AI went from curiosity to near-default.
But adoption measured in mere logins is a flattering mirror. Look slightly past the headline and the same report shows something less comfortable: 54% of firms provide no training on responsible AI use and have no plans to start, and 43% have no AI policy and no intention of writing one. Individuals have raced ahead, but the institutions employing them have, for the most part, not moved at all.
That is the adoption gap. Not a gap between firms that use AI and firms that don’t (that argument is over) but a gap between using AI and using it well. And it raises a question worth sitting with rather than answering too quickly. When AI underdelivers in a law firm, whose failure is it: the firm that bought the tool, or the tool that was sold to the firm?
The case against the firms
Starting with the uncomfortable half.
Most legal AI disappointment is not a technology problem. When AI pilots fail, they tend to fail in the background, and not because of weaknesses in the models, but because the firm underestimated governance, workflow design, and the unglamorous work of change management. That diagnosis runs well beyond law. Across industries, roughly seven in ten AI deployment failures are structural rather than model-related: integration, process, people. The technology mostly works, however, the organisation around it doesn’t.
There’s a number that makes this concrete, and it should give managing partners pause. While 69% of individual lawyers use AI, only around a third of firms have adopted it at an organisational level. Many firms are, in effect, outsourcing their AI strategy to whichever associate is curious enough to experiment after hours. And the deeper problem runs past governance: the same body of research suggests that two lawyers handed the identical tool can produce wildly different results depending on how each one frames the task. The advantage exists in the person using it and how clearly they think about the problem before they type, as opposed to the software.
As long as that holds, then buying the tool was never the hard part. The hard part is teaching people to use it, building the guardrails so they use it safely, and redesigning the work around it. Most firms have done the easy thing: signed the contract and called it transformation. Unfortunately, they often can’t tell whether it worked. Legal tech spending grew faster in 2025 than in possibly any year on record, yet the overwhelming majority of firms either don’t accurately measure the return on their AI investment or aren’t sure whether they measure it at all. You cannot close a gap you refuse to measure.
So yes, a great deal of the blame sits with the buyer.
The case some lawyers would make back
I’ve conducted enough interviews to know the lawyers have a rebuttal, and a fair-minded piece has to give it room.
Their argument is not that the tools don’t work. It’s that “works” and “trusted” are not the same thing, and that the distance between those two words is where adoption stalls. One prediction for 2026 that has stuck with me put it plainly: the gap between “AI works” and “AI is trusted” is becoming impossible to ignore. A lawyer who hesitates to lean on a tool is not necessarily a technophobe clinging to the past. They may be making a rational professional judgement about a system that hasn’t yet earned their confidence.
And some of that friction is real, documented, and fixable. Users of major legal AI platforms report concrete irritations such as having to manually feed documents into a tool that was sold as seamless, or discovering that bending a platform to a specialised practice area takes extra configuration and paid services. None of that means the software is all bad. It means the last mile into a working lawyer’s day is genuinely difficult, and not every product has closed it.
That’s a big part of the lawyer’s point, fairly stated: meet us where we actually work. A tool that demands a new tab, a fresh login, and a separate mental mode will be reached for when there’s time and ignored when there isn’t. Which, for a busy lawyer, is most of the time.
Here’s what’s encouraging, though, and why this isn’t a criticism of the industry so much as a description of it maturing. The market is already correcting. The newer wave of legal AI is being built specifically around this complaint. Review and drafting are being embedded directly inside Microsoft Word and pre-built templates now produce usable work on day one rather than after a six-month implementation. The vendors who understood that the obstacle was never model quality, but the last mile into the workflow, are the ones pulling ahead. The complaint and the cure are arriving together.
It was never really firms versus tools
This is why I think the question I opened with is the wrong one.
“Firms or tools” implies a culprit. But look closely at the two failure modes and they collapse into the same shape. A tool that doesn’t fit the way lawyers work, and a firm that buys a tool without redesigning the work around it, produce an identical outcome: expensive software, logged in and barely used. The failure isn’t the model and it isn’t the lawyer. It’s the missing layer between them: implementation, workflow fit, the deliberate redesign of how the work actually moves.
The strongest voices in legal tech are already converging here. The most valuable AI of 2026, they argue, won’t be the cleverest abstract system. It will be the tool that works directly alongside a lawyer, inside the process that already exists. Vendors are starting to build for that. Far fewer firms are buying for it.
So here is the takeaway I’d offer both sides. To firms: stop treating AI as an IT purchase and start treating it as a change-management programme, because that is what it is, and the difference between those two framings is roughly the difference between a project that delivers and one that doesn’t. To vendors: the durable advantage was never the model. It’s whether your product dissolves into the working day or sits beside it as one more thing to remember.
Because the adoption gap is not closing this year. If anything it’s hardening into a line between those who have closed the distance between a tool and a working day, and those who still treat buying the software as the finish line. That line is the real debate. And it was never drawn by the model, or by the lawyer. It’s drawn in the space between them, the part everyone assumed would take care of itself.
