When Legal AI starts taking action, governance has to catch up

For the last few years, most Legal AI use cases have followed a familiar pattern:

Ask a question → get an answer → review the output.

But the market has already moved beyond that model.

Three months ago, Anthropic introduced Claude Tag, allowing teams to bring Claude directly into Slack, give it access to selected tools and context, and delegate work to it much like they would to another team member.

Last week, the same shift became even more visible in Legal AI: Legora launched Intake, and Harvey introduced its own intake workflow for in-house legal teams.

The interesting part is not another AI feature. It is what this signals for legal teams:

And once AI starts taking action, another question becomes much more important.

If an AI agent is part of a legal workflow, the team needs to decide:

• What information can it access?
• Which contracts, policies or business data can it read?
• What actions can it take autonomously?
• Which tasks require human approval?
• Can it communicate with the business on Legal’s behalf?
• Who remains accountable for the final output?

Legora’s own approach reflects this shift: how far the Agent goes is determined by the legal team, request type by request type, with lawyer review as the default.

This is where AI governance becomes part of workflow design.

For an AI agent to perform legal work effectively, it needs context:

• Role — what it is responsible for.
• Knowledge — the business and legal information it needs.
• Memory — relevant previous decisions and work.
• Skills — the process it should follow.

But access to context should not mean access to everything. As legal teams give AI agents more capabilities, they also need to define roles, permissions, approval points and accountability around those agents.

This is exactly what we will explore in our next workshop:

We’ll walk through how to configure data access, manage permissions, and establish boundaries for AI agents working with sensitive business information.

The goal is not simply to give AI more autonomy.

It is to make that autonomy controlled, appropriate and accountable.

One question to ask your legal team

If an AI agent joined one of your legal workflows tomorrow:

As Legal AI moves from assistance to action, those questions become just as important as choosing the model itself.

Join us at our upcoming workshop to explore these questions in practice and see how roles, access levels and approval boundaries can be configured in Claude.