Skip to main content
← Back to Blog
AI Governance, Ethics & Compliance
AI-assisted · automated editorial review

What Should Your Firm Verify Before an AI Agent Handles Client Work?

|AlphaForge Editorial|7 min read
AI GovernanceAgent PermissionsHuman OversightLaw Firm Operations

What should you verify before an agent handles client work?

Consider a staff member's proposed workflow: an assistant would retrieve matter documents, prepare a summary and route it for approval. You may need to establish whose files it could open, what it could change and when a lawyer would take over. As of November 2026, our recommended approach is to approve that specific workflow using evidence you can review.

Our recommended answer: approve a defined task with a named owner, limited permissions, tested escalation and a retrievable activity record. Assess broader access or additional actions separately. We recommend evaluating the agent's ability to finish the task alongside whether it stayed within the firm's instructions.

The examples below draw on a research abstract and two OWASP security guides reviewed on November 12, 2026. The study describes a controlled evaluation. The guides describe recommended controls. Our proposed approval checks are for a firm to try with fictional records.

What the November evidence supports—and leaves unresolved

Finding an unapproved agent is a separate task

In their November 9 abstract for Detecting Shadow AI: Unsanctioned LLM Agents on Enterprise Endpoints: A Taxonomy, an Open Detector and a Controlled Evaluation, researchers describe a Windows detector that combines remote-service identity with configuration, installation and execution history. Their controlled evaluation used 12 synthetic agents, three AI components and nine benign workloads, repeated five times on one workstation.

The authors report detecting every agent in the majority of runs, but an intermittent agent appeared in about one scan in three. Attribution for a provider behind a shared content-delivery network depended on a cached provider name lasting one to five minutes. The authors explicitly say the bench was too small to rank detectors.

Our interpretation is that a clean scan could leave intermittent activity undiscovered. This abstract supports asking how detection was performed and what it missed; it does not establish how common unauthorized agents are in law firms or validate a complete monitoring program.

Check the permissions behind each connection

OWASP’s Authorization Cheat Sheet recommends granting only the minimum privileges needed and checking permission on every request.

For a firm considering a document connection, we recommend testing the account and the actions it can take. Include a fictional matter it should reach and one it should be denied. Have the operator check the access record for both.

Verify approval at the point of action

OWASP’s AI Agent Security Cheat Sheet recommends that the component executing an action verify approval outside the agent. It also recommends that approval apply to the exact tool call and be checked and consumed immediately before execution.

We recommend testing routine completion and escalation separately. In a fictional case, compare a routine request unnecessarily sent to a lawyer with a request completed when lawyer review was required. Record which outcome occurred so the person approving the workflow can see where to intervene.

Recommended steps: make approval specific enough to inspect

1. Write down the task and its stopping point

We recommend a short workflow statement before connecting client information. Name the input, permitted work, expected output and person responsible for approval. For example, specify whether the agent prepares an internal draft, changes a matter record or communicates externally. Treat each additional action as a separate decision.

We recommend stating the stopping point in ordinary language: “Prepare the draft and send it to the supervising lawyer's review queue.” List the actions outside scope, including sending the draft to a client or making a substantive decision. Give staff an identifiable person to contact when the request falls outside that description.

2. Check access through the actual connection

Review which account the agent uses, which repositories it reaches and whether its access permits reading, editing or sending. Ask the implementer to demonstrate those permissions with test records. Include a record it should access and one it should be denied, and retain the observed results.

We recommend recording the connected services, account owner and approved configuration together. If a connector changes, check whether the approved boundaries still hold. Keep what works. Change what doesn't. The decision should follow the firm's requirements, whether that means configuring existing software or assessing a scoped replacement.

3. Test refusal and escalation deliberately

We recommend constructing test requests that should succeed, stop or reach a human. Include an unavailable document, an ambiguous matter name, a request outside the approved task and an attempted action without permission. Write the expected response before running each case so reviewers have a consistent basis for judging it.

Review the agent's actions as well as its final answer. Did it attempt to open a prohibited record? Did it stop before sending? Did the receiving lawyer get the necessary context? Our related article on why similar agent scores hide different failures provides a useful companion to separating those outcomes.

4. Preserve a usable review record

We recommend retaining enough information to reconstruct a test: the request, approved configuration, information retrieved, actions attempted, actions completed and human review decision. Distinguish a proposed action from an attempted action and a completed action. Assign responsibility for reviewing exceptions and controlling access to these records.

Check whether an operator can stop the workflow and whether the review queue has an owner. Do not approve unattended operation merely because a demonstration ended successfully. Require evidence for the boundaries relevant to the proposed task, and document unresolved limitations alongside the approval decision.

Illustrative example: an internal probate matter summary

This is a constructed operational example, not an AlphaForge deployment or client result. A probate team might propose using an agent to prepare an internal summary from documents in one designated test matter. In this example, the proposed workflow would end when a supervising lawyer receives the draft and its source references.

We recommend testing four situations: the correct matter, two matters with similar names, a missing document and a request to email the summary externally. Use synthetic records for this exercise. Define success separately for accurate retrieval, source support, access denial and stopping before an unapproved communication.

Consider a run in which the agent prepares a useful summary but attempts to retrieve the similarly named matter. That could indicate an access or selection problem even if the final draft contains no obvious error. Record the attempted retrieval and resolve the boundary before approving the workflow.

We recommend having the supervising lawyer check the summary against its source documents and the operator inspect the action record. A successful exercise would support only the tested workflow under its recorded conditions. It would not establish reliability across every probate matter, document type or future configuration.

What to measure before expanding the workflow

We recommend reporting results by failure type, with counts and denominators. Separate correct completions, correct escalations, unnecessary escalations, missed escalations, prohibited actions attempted and prohibited actions completed. Report missing activity records separately. A small trial should remain visibly small.

Measure the work left for people as well. Record review time, corrections and unresolved exceptions without treating faster completion as proof of sound judgment. If a configuration or permission changes, identify which earlier test results still apply and which need to be repeated.

For unapproved use, we recommend combining a staff inventory with the technical checks available to the firm. Document the systems inspected, observation period and known gaps. Our interpretation is that the shadow-agent abstract's intermittent-detection limit could warrant qualifying a finding of “nothing observed” when reviewing a scan with similar limitations.

We recommend keeping an approval record that answers three questions: What may this agent do? What evidence supports that permission? Who intervenes when it reaches the boundary? Review those answers and the remaining exceptions before expanding the workflow.


AI Governance, Ethics & Compliance

Start with the problem in front of you.

The governance review is a focused entry point into the same Firm Intelligence Audit.

Liked this post?

Get research, operating guides, and implementation notes once a month. No spam.