Six predictions for legal AI in 2026 (and one I hope I'm wrong about)

I’m wary of prediction pieces, because most of them are written to be un-falsifiable - vague enough that the author can claim a win whatever happens. So I’ve tried to make these specific enough to be wrong. If most of them haven’t come true by December, you have my permission to hold it against me.

Here are six predictions for legal AI in 2026, and one I hope doesn’t land.

One: the “AI strategy” gives way to the AI portfolio

In 2025, having an “AI strategy” was a status symbol - a deck, a committee, a statement of intent. In 2026 the firms that are actually ahead will stop talking about their AI strategy and start talking about their AI portfolio: the specific, named set of workflows they’ve productised, with owners and adoption numbers attached. The tell that a firm is serious will shift from “do they have a strategy?” to “can they list the live workflows and tell you who uses them?” Strategy was the 2025 flex. Shipped, adopted workflows are the 2026 one.

Two: procurement starts asking about the boundary, not the benchmark

Legal AI buyers spent 2025 asking vendors about accuracy. In 2026 the sophisticated ones start asking a better question: where does your tool stop and a human take over, and how do you make that boundary visible and checkable? The conversation moves from “how often are you right?” to “how do I know when to trust you?” Vendors who only have a benchmark to offer will start losing deals to ones who can answer the trust question - because the buyers have been burned by confident black boxes and learned what actually matters.

Three: at least one “AI-native” challenger does real damage in a narrow category

Somewhere in 2026, a focused, AI-native provider - not a traditional firm, not a horizontal vendor, but a business built from scratch around productised legal work - takes meaningful market share in one narrow, high-volume category. Not law broadly. One category: a type of NDA, a class of routine filing, a specific compliance check. It won’t make the mainstream headlines, but the firms in that category will feel it, and it’ll be the first clear proof that the productised model can win share, not just margin.

Four: the junior-training problem becomes openly discussed

In 2025, the quiet consequence of automating junior work - that you’re also defunding how juniors become seniors - was largely unspoken. In 2026 it becomes an open conversation, at conferences and in management committees, because the first cohort trained in the AI era starts to visibly lack the reps their predecessors got. Firms will start, awkwardly and publicly, trying to design deliberate ways to manufacture judgement now that the grind that used to manufacture it has been automated away.

Five: “human in the loop” gets more honest

The phrase “human in the loop” was used in 2025 to mean almost anything, often as a fig leaf - a human technically present but not really checking. In 2026 it gets more rigorous, partly because the first incidents force it. Firms and regulators start distinguishing between a human who is genuinely able to check the work and one who is just rubber-stamping at a speed that makes checking impossible. “In the loop” stops being a comfort phrase and starts being a design requirement with teeth.

Six: the winning tools get more boring to look at

The legal AI that wins real adoption in 2026 will be less visually impressive than the demos of 2025, because it will have disappeared into existing workflows rather than announcing itself. The flashy standalone copilot with its own interface gives ground to capabilities embedded in the tools people already use, doing their work in the background. The best legal AI of 2026 will be the kind you have to be told is even there.

The one I hope I’m wrong about

Here’s the prediction I’d genuinely rather not make: in 2026, at least one firm or provider will ship an AI tool with the judgement boundary in the wrong place, a human “in the loop” who couldn’t realistically check the output, and it will cause real harm to a real client - a missed deadline, a bad filing, a wrong answer relied upon - in a way that becomes public.

I hope I’m wrong because the harm would be real and would land on someone who didn’t choose it. But I’d be surprised if the year passes without it, because the incentives are pushing exactly that way: pressure to automate, pressure to show savings, and a temptation to call something “supervised” when the supervision is theatre. If it happens, the lesson won’t be “AI is too risky for law.” It’ll be the lesson it always is - that the model was never the hard part, and that putting the boundary in the right place, and meaning it, is the whole job.

Six predictions and a fear. Check back in December and tell me which ones I got wrong - I’d genuinely like to know, especially about the last one.