Why most legal AI will be invisible

The legal AI that gets talked about has a face. It’s a copilot with a name, a chat box, an interface you go to, a brand the firm can point at. That’s the version in the demos, the conference keynotes, the procurement decks. And I think it’s mostly a transitional form - a phase, not a destination. The legal AI that actually wins, in the end, will be invisible: dissolved into the tools lawyers already use, doing its work in the background, with no face at all. And the invisibility is the whole reason it’ll win.

Visible AI is a tax

Start with why the visible, go-to-a-copilot model is structurally weak, even when the underlying technology is excellent. A separate AI tool - its own interface, its own place you navigate to - imposes a tax on every use. The fee-earner has to leave what they’re doing, switch context, go to the tool, do the thing, bring the result back. Every step of that is friction, and busy lawyers don’t pay friction taxes if they can avoid them.

The visible copilot also asks the user to do something unnatural: to think about the AI as a distinct thing they’re using. That’s a cognitive overhead on top of the actual work. The whole interaction is framed as “now I am using the AI,” which is a framing that has to justify itself every single time against just doing the task the way you already know. A lot of the time, it loses that justification, which is why so many beautifully-built copilots sit unopened.

Invisible AI is just better work

Now contrast the invisible version. The AI isn’t a place you go - it’s a capability inside the tool you’re already in. The document you’re drafting just gets smarter. The matter file just assembles itself. The review just surfaces the anomalies as you work. There’s no context switch, no separate interface, no “now I am using the AI” framing. There’s just the work, getting easier, in the place the work already happens.

This is strictly better along every axis that actually governs adoption. No friction tax, because you never leave. No cognitive overhead of managing a separate tool, because there isn’t one. No first-session funnel to fall out of, because there’s no separate thing to onboard into. The AI meets the lawyer exactly where they already are and improves the work they’re already doing. The best version of legal AI is the version you’d have to be told was AI, because from the inside it just feels like the tool got better.

Why the visible phase exists at all

If invisible is better, why is everything currently visible? Because we’re early, and visibility is how a new capability gets sold and understood. When something genuinely new arrives, it has to announce itself - the chat box, the brand, the demo - so people can grasp what it is and buyers can evaluate it. The visible copilot is a teaching tool and a sales tool. It makes the capability legible.

But that’s a phase, not the mature form. Every general-purpose technology starts visible and ends invisible. Electricity started as a marvel you went to see and ended as a thing in the wall you never think about. Computing started as a room you visited and ended as a thing dissolved into every object you own. Legal AI is at the marvel-you-go-to-see stage. The mature stage - the one that actually reshapes the work - is the in-the-wall stage, where it’s everywhere and noticed nowhere.

What this means for firms and builders

For firms, the implication is to be a little sceptical of how you’re measuring AI adoption. If you’re counting logins to a copilot, you’re measuring the visible, transitional form - and you may be missing both the friction that’s quietly killing it and the better future where the AI is embedded and your login metric goes to zero precisely because it succeeded. The right question isn’t “are people using our AI tool?” It’s “is the work getting better, wherever it actually happens?” Those can have opposite answers.

For builders, the implication is sharper, and it’s where I’d put my chips: the durable value is in embedding the capability into the workflow, not in building another destination. The standalone copilot with the great interface is competing to be a place people go, and being a place people go is a losing position against being a capability that’s simply there. The teams that win the next phase of legal AI are the ones building the in-the-wall version - the AI you don’t see, in the tools people already use, making the work better without asking for attention.

The legal AI everyone’s talking about has a face and a brand and a chat box. The legal AI that actually changes the profession will have none of those things. It’ll be invisible - and that’s not a limitation it has to overcome. It’s the entire point.