Standards for Legal AI That Preserve Human Judgment

The line we haven't drawn: why "AI-assisted" doesn't solve governance

Everyone agrees AI belongs in an assisting role. Organizations are writing policies around that principle, and almost none of them say where the line actually sits. That missing definition is the governance gap.

Someone runs a contract through an AI tool and gets a summary back. That is AI-assisted, technically. It also tells you nothing about whether the output is good enough to skip the attorney read, what the tool assumed in the redline, or what happens when it is wrong.

Start by finding the task that carries the judgment. AI can work almost anywhere, so capability is a poor filter. Contract automation flags an indemnification clause; deciding whether that risk is acceptable for the business is the attorney's call. That decision is the task that needs the attorney. Everything around it is scaffolding.

Most policies say "AI tools approved for document review." A usable one says what the output must contain, names the attorney's decision, and defines the handoff between the two. That turns the tool into information architecture, judged on how well it serves the decision.

So reverse the usual approach. Start from the attorney's irreplaceable role, then build the AI into that workflow to serve it. Legal departments will spend the next eighteen months learning that "AI-assisted" is too vague to hold. The ones that do well will design the architecture for how judgment consumes AI output. That's not a technical problem. It's an organizational design problem.