"AI-powered" now describes a grammar checker and a multi-step contract review with equal enthusiasm. "Automation" is applied to processes that still require a lawyer at every step. The result is procurement decided on vocabulary, and deployments that disappoint for reasons nobody can name afterwards.

The distinction that matters is not technical sophistication. It is whether the system holds state between human touchpoints.

Assistance

In assistance, a lawyer asks and a system answers. Research, summarisation, first-draft generation, clause lookup. The lawyer supplies the context, judges the output, and carries it to wherever it is going next. The system is present for one exchange and remembers nothing that matters.

This is genuinely valuable and it is where most firms should start. Its benefit is also linear: each lawyer gets somewhat faster at the tasks they already do, and the gain stops there.

Automation

In automation, the process is the thing that exists. A review runs: provisions are extracted, deviations from the firm's standard positions are flagged, a redline is drafted, and the file arrives at a named partner in a state that says what was done and what needs deciding. The lawyer's touchpoints are defined in advance rather than improvised, and between them the system is doing substantive work that persists.

The benefit is not linear, because the constraint it relieves is not typing speed. It is how many matters a given team can carry without dropping one.

The firms gaining the most from legal AI have stopped asking what the AI can do and started asking which of their workflows it should run.

Why the distinction decides the return

An assistance tool is bought per seat and justified by hours saved, which is a number nobody can measure honestly. An automation platform is justified by matters handled, cycle time and realisation — numbers a managing partner already tracks and already trusts.

That difference also explains why assistance pilots so often stall at renewal. The tool worked; there was simply no way to show what it changed. Automation is easier to defend precisely because it is harder to install.

What automation demands that assistance does not

A tool that answers one question needs a good model and a text box. A system that carries a matter needs four things a text box does not have.

  • State: a matter that knows which phase it is in, what has cleared and what is blocking
  • Gates: points where work stops until a named person releases it — enforced, not encouraged
  • A record: what was produced, from what context, approved by whom, against which version
  • Isolation: a boundary the process cannot cross even when it is convenient

Those are infrastructure properties, and they are why the two categories are not points on one spectrum. You cannot arrive at automation by adding features to assistance; the state model has to exist first.

How to tell which you are being sold

Ask what happens when the lawyer closes the laptop. In an assistance product the answer is nothing — the conversation ends. In an automation product the answer is that the matter is in a state, someone owns the next step, and the record already reflects what happened.

Both answers are legitimate. They are answers to different problems, and firms that pay automation prices for assistance products are the ones who did not ask.