A year ago, "does your firm use AI" still told you something. Today it tells you almost nothing. Early in 2026, LawNext's 8am Report found that AI adoption among legal professionals had more than doubled in a year. The models are good. They're getting better. And now almost everyone has them.

If every firm works from the same engine, the engine can't be what sets one apart from another. Yet plenty of firms still compete as though having it is the edge.

Start with what the model does on its own. Dennis Kennedy has described the way we've treated generative AI for two years as a vending machine: put in a prompt, take out a product. The trouble with a vending machine is that it gives everyone the same snack. Ask the same models the same question and the drafts drift toward the same place, a competent, defensible market average that looks a lot like the one the firm across the street just produced. Your clients are standing at the same machine. As a floor, that's fine. As an edge, it's nothing. 

The obvious answer is that the best tools no longer leave you at the average. A banking partner told us recently that the assistant his team uses is genuinely impressive, better than he'd expected, but still not quite up to the job he most needs it for. The easy read is that the model just isn't powerful enough yet. But the models are so good now they're rarely the limit. What the model has to work from is the more likely source of discontent. 

So what is it working from? A firm's store is not a library of its best work. It's a mix. The perfect clause and a compromised one. The final version and the abandoned draft. The position the firm would defend to the wall and the concession it gave away on a bad Friday. Feed a model all of it and it draws on all of it, because no one has told it which is which. Above the Law's Joe Patrice put it well earlier this year: an AI answer can simultaneously be accurate and still not right - correct on what it happened to read and blind to what it missed. Unmarked, your best work and your worst look identical to a machine. A strong generic draft isn't your firm's best, tailored position, the one a partner would put their name to. That's the question the good tools skip over: which documents?

So the edge was never "your files." It's a subset of them: the approved, standard-bearing work you actually stand behind, the reference a draft should be measured against. And that subset isn't labelled. That market-leading financial instrument your firm drafted is sitting in a closed matter or on a partner's laptop, indistinguishable to any system from everything around it. Legal IT Insider spent much of this year making the point that a firm's know-how is only worth something when it is captured and reachable across the firm. In most firms at best, it’s a team of knowledge lawyers trying their best to keep up with the ever expanding content – at worst,  this know-how is scattered. It's dormant. An AI model that reads everything still can't lift the standard out of the noise on its own. 

Lifting it out is real work. It's work someone in the firm already does. Turning a firm's scattered best thinking into one version the whole team can draft from is a great deal of what good knowledge teams are for. Anyone who's tried it will tell you it's slow, careful work, which is a large part of why it so often doesn't get done. 

This isn't an argument against AI. I use it every day. It's an argument about what has to sit underneath it. The move isn't to feed the machine more of your files. The good tools already read them. It's to decide which of your work is the standard, to approve it, to make that the thing every draft starts from. Left to itself the machine produces confident mediocrity (at best) when the inputs are wrong. More than seven hundred fake, AI-invented cases have already reached court filings. The human sign-off isn't overhead. It's the point. Use the assistants for the reach and the open questions. Use your own approved work for the answer you put your name to. Both, not either. 

This isn't new ground for us. Capturing a firm's best work and making it reusable is what knowledge teams have done by hand for years. We've spent the last decade helping them scale it. By hand, a standard sits with a few people and slowly goes out of date. Scaled, it reaches every lawyer and stays current as the firm's positions move. Automation was the first way to do that. AI and automation together is the next. 

Either way, the machine is only as good as the work you've decided is worth building on. 

This is what the industry keeps skipping past on its way to the next model. Not the intelligence, which everyone can now rent, but the infrastructure that decides which of your own work the intelligence is allowed to build on. Call it drafting infrastructure. It's the difference between a firm that drafts like the market and a firm that drafts like itself. 

The question stopped being whether you have AI. In a year everyone will. It's not even whether your AI can read your own work. Soon all of them will. It's whether you've decided which of your work is worth reading. Same models. Same access to your files. The firm that pulls ahead won't have a better model, or more of its files loaded into one. It will have chosen and curated, the right ones.

David Howorth

Co-founder, Avvoka