The scope of the risk depends on the authority the firm gives it. An agent that can search one matter workspace presents a different problem from one that can search every matter a user can access. An agent that can prepare an email presents a different problem from one that can send it. Client confidentiality, ethical walls, and matter-specific restrictions must all hold when someone accesses information through an agent.
At Clarra, we have applied this principle by having our AI assistant inherit the user’s role and matter permissions. If a paralegal cannot open a matter, the assistant working on that paralegal’s behalf cannot open it either. That addresses one boundary, though permission controls are only part of governing what an agent may do with information it is allowed to see.
The instructions an agent encounters can also come from the material it is asked to read. Imagine an agent reviewing an opposing party’s email that contains a direction to search for and forward a client document. The lawyer assigned the agent to review correspondence, but the email is attempting to direct its next action. This is known as indirect prompt injection, a risk identified by the National Institute of Standards and Technology for agents that process outside information. AI controls need to keep instructions found in a document or email from carrying the same authority as instructions from the lawyer.
Supervision becomes harder when the action happens first, and the review comes later. ABA Formal Opinion 512 explains how lawyers’ existing duties, including confidentiality and supervision, apply to using generative AI. As firms give AI tools the ability to take actions across systems, they need to decide where approval is required and how they will detect actions that cross a boundary. Those decisions belong in the workflow, not just the AI-use policy.
Five Decisions to Make Before an Agent Starts Acting
An AI-use policy remains useful, but a policy directed at employees cannot substitute for controls on the agent’s access and actions. Technology, security, information governance, and management committees should make five decisions together:
- 1. What can the agent reach? Apply minimal access. Grant agents access to a specific matter, workspace, or data set, not the entire environment.
- 2. What can the agent do without asking? Decide what the agent may prepare and what requires a person to authorize execution. Sending an external email, filing a document, changing permissions, and deleting records warrant explicit controls at the point of action, not a reminder buried in a prompt.
- 3. Can we see what the agent did? Keep logs that show the task, the agent’s actions, the information it accessed, and the person who approved consequential steps. The record should help people investigate a mistake, not just sit somewhere in the platform.
- 4. Who owns the agent? Assign a named owner to every agent and establish an approval process for deploying new ones. Agents launched by individual users without firm review are the new version of shadow IT.
- 5. How do we stop the agent’s actions? Build and test a way to pause or shut down an agent quickly, from anywhere. Test that process while the stakes are low.
These decisions are easier to enforce when agents work inside defined matter workflows with clear permissions. An open-ended request across inboxes and repositories gives an agent more choices and the firm more activity to monitor.
Give Autonomy a Defined Scope
Law firms can gain real value from agents that gather information and move routine work forward. As those agents gain access to firm systems, each new connection and capability should come with a decision about what they are authorized to do. An agent might be trusted to find the latest agreement and prepare a client briefing without being allowed to send either one.
In the opening example, the agent chose the wrong draft. In a well-governed firm, that mistake lands in the lawyer's review queue, not the client's inbox. Readiness for autonomous AI comes down to building that boundary before the agent needs it. AI doesn’t go rogue to cause harm; it was only trusted with one step too many.