Adoption was never the problem.

For two years the question in most firms was whether to use AI. That question has been answered, nearly seven in ten legal professionals now use generative AI for work. The harder question, the one almost no one has answered, is why all that usage has produced so little value.

The honest picture inside most firms is this: partners and associates are using AI on their own, in scattered ways, with no method and no training, and the firm is not noticeably better, faster, or cheaper for it. The licenses are bought. The tools are open in a browser tab. And the work product, the margins, and the client experience look about the same as they did before. That gap, between adoption and value, is the actual problem, and it is worth understanding precisely, because the usual explanations are wrong.

This is not unique to law firms

The first thing to understand is that the value gap is everywhere, not just in legal. MIT's NANDA initiative studied 300 public AI deployments and surveyed and interviewed hundreds of leaders, and found that roughly 95% of enterprise generative-AI pilots produced no measurable return. Companies of every kind have announced AI initiatives and quietly seen nothing arrive at the bottom line.

What matters is MIT's explanation for why. The failure was not the quality of the models. It was integration. Generic tools work beautifully for an individual precisely because they are flexible and unopinionated, but that same generality is why they stall inside an organization: they don't learn the organization's work, and they aren't adapted to its actual workflows. The tools are powerful and the pilots still fail, because nothing connects the tool to how the business really operates.

That is the universal version of the problem. The legal version is the same in shape, and worse in one specific way.

Inside law firms, the use is ad hoc

Individual adoption among legal professionals more than doubled in a single year, reaching roughly 69%. But firms as institutions have not kept pace, and the gap between individual use and firm-level structure is where the trouble lives. Only about a third of firms have adopted a legal-specific AI tool. More than half, 54%, provide no training at all on responsible AI use, and have no plans to. Only around 9% have a written, actively enforced AI policy.

Put those together and the result is predictable: most AI use in law firms is informal, untrained, and ungoverned. Lawyers reach for whatever tool is at hand, with no shared method for what to use it on, how to check it, or what never to put into it. That is how you get inconsistent work product, confidentiality and privilege exposure, and effort spent without efficiency gained. Data security, ethics, and privilege are the barriers firms cite most, and ad hoc use is exactly what turns those risks live.

The deeper problem: the wrong process, not just an underused one

Even where firms do capture some efficiency, there is a second, subtler problem that is specific to professional work. A general-purpose tool, used out of the box, shapes the work to its own generic defaults, not to how a particular firm has decided to practice. The output is competent and anonymous. It looks like everyone's, because it is everyone's. The firm's own way of working, the judgment, the standards, the structure that make it worth hiring, never makes it into the system.

Worse, AI used on the wrong task actively degrades expert work. In a Harvard Business School experiment with Boston Consulting Group consultants, AI improved output quality by roughly 40% on tasks involving breadth and option generation, but on tasks requiring judgment and synthesis, consultants using AI were 19% less likely to reach the correct answer than those working without it. Studies of legal work show the same split: large gains on straightforward analysis, no benefit and recurring failures on complex reasoning, confident prose masking superficial analysis, missed issues, jumping to conclusions. One study found that the strongest lawyers produced worse revised work product when they let AI edit it.

The lesson is not that AI is dangerous. It is that value depends entirely on which cognitive work is handed to the tool and which stays with the lawyer, and on shaping the tool to the firm's actual practice rather than accepting the vendor's defaults. That is a question of method and design. It is not something a license decides for you, and it is not something ad hoc use will ever stumble into.

Structure is what converts use into value

The firms that get value are not the ones with the best tool. They are the ones that approached AI deliberately. Thomson Reuters found that firms with a visible AI strategy are nearly four times (3.9×) more likely to be seeing real benefits than firms without one. And the prize is not small: each lawyer is expected to save on the order of 190 hours a year through AI, roughly $20 billion of time across the US legal market, time that can be redirected to higher-value work, better client service, and growth.

Read those two findings together and the conclusion is hard to avoid. The benefit is real and large, but it accrues almost entirely to firms that did the structural work first: deciding how the firm practices, what the tool is for, where judgment is non-negotiable, and how quality gets measured. The difference between wasted AI and valuable AI is not the model. It is the method around it.

Which is the whole point

The gap between adoption and value closes in one place: by codifying how the firm actually practices, then fitting and governing the tool to that, rather than letting a general-purpose product quietly decide how the firm works. That is the work. It is the part no license includes, and it is the reason Codified Counsel exists.

← Back to insights See how we close the gap →