The State of AI-Native Legal Practice, Late 2026

I get asked some version of "where are things actually at" often enough that it is worth writing down plainly, without the usual hype-cycle framing on one side or the doom framing on the other. This is what I am actually seeing, from inside the build work, not from a survey or a vendor report.

The gap between talking about AI and running on it has narrowed, not closed

A year or two ago, most conversations I had with lawyers about AI were still hypothetical, would this work, could this be trusted, what would it even look like in my practice. Those conversations have mostly shifted to specific and operational, which agent handles intake, how do I anchor drafting output, what does the review step actually need to catch. That is real progress. What has not closed is the gap between practices that talk about AI adoption and practices that have actually rewired a workflow end to end, agent handles the first pass, human reviews, output ships. Most practices I encounter are still somewhere in the middle, using AI as an occasional tool rather than a structural part of how a matter runs. I drew that specific line in AI-native vs AI-augmented, and it is still the most useful diagnostic question I have: is AI touching every matter systematically, or only the ones where someone remembered to use it.

Evidence anchoring has gone from a nice-to-have to a baseline expectation

This is the clearest shift I can point to. Two years ago, explaining why AI output needed to trace back to a source document was a persuasion exercise. Now it is closer to an assumed baseline, because enough public examples of unanchored AI output causing real professional consequences have circulated that the argument mostly makes itself. What has not caught up is implementation. Plenty of practices agree in principle that anchoring matters and have no actual structure enforcing it, which is a gap between belief and practice that tends to close only after either a near-miss or a system that makes anchoring the default rather than an extra step.

Agents are doing more, but supervision has not gotten lighter

The agents themselves have gotten noticeably more capable this year, at intake triage, first-pass drafting, and document analysis specifically. What has not changed, and should not change, is how much supervision that output needs before it leaves the practice. I still see the same failure pattern I described in my top 10 AI mistakes I've watched law firms make: capability improving faster than review discipline, which produces a false sense of safety exactly when the output is getting more convincing rather than more actually correct. Better agents are not a substitute for the review step, they make the review step matter more, because a more fluent wrong answer is harder to catch than an obviously bad one.

Pricing conversations have started catching up to the workflow shift

This is the most operationally interesting change I have watched this year. For a long time, practices adopted AI for speed while keeping the billable hour exactly as it was, which quietly punishes the practice for getting faster. That is starting to shift, more conversations now involve genuine repricing, flat fee or value-based structures that reflect the actual time a matter now takes rather than the time it used to take before agents carried the first pass. It is still a minority of practices making this change deliberately rather than by accident, but the direction is clear, and I think the practices that price for the old workflow while running the new one are going to feel that mismatch acutely within the next year or two.

What is still mostly talk

Full end-to-end matter automation, from intake through to a filed and closed matter with minimal human touch, is still mostly aspirational for anything beyond the most standardised, high-volume matter types. The claim that AI can run a matter with no supervision does not hold up against a real, messy file, and I do not expect that to change soon, because the judgment calls in the middle of a contested matter are not the kind of thing that benefits from removing a human, they are the kind of thing the human exists for. Anyone selling full automation as a near-term reality for anything but the most templated work is selling a demo, not a practice.

The honest summary

AI-native practice in late 2026 is real, operational, and genuinely changing how work gets done in the practices that have committed to it structurally. It is not universal, it is not automatic, and it does not remove the need for a lawyer's judgment at the points where judgment was always the actual job. The practices ahead of the curve are not the ones with the most AI tools installed, they are the ones with the clearest structure for what gets automated, what gets anchored, and what still requires a human to sign off before it leaves the building.