What should partners approve before staff start using AI?
Consider this proposed operating scenario: a staff member wants to use an AI assistant to summarize intake notes. The demonstration looks useful, but you still need to decide which records it can read, who checks the summary, and whether it can update the firm's systems. Our recommendation, as of November 2026, is to approve a defined workflow with documented limits, supported by a test using fictional records, before allowing client information into that workflow.
A tool name alone leaves too much undecided. Permission to draft an internal summary should specify whether the assistant can also send an email, retrieve another matter's documents, or change an intake status. Those are separate decisions, each requiring a named owner and a clear boundary.
This article answers one operating question: what evidence should a managing partner require before approving an AI workflow? The practical answer is a short approval record covering the task, permitted data, available actions, human review, test results, and conditions for pausing use.
What the available evidence supports
Write fictional records around the task you plan to approve
We recommend preparing fictional records with the relationships and timing your staff would need to interpret. For example, give the assistant an intake note followed by an email that corrects a name. Write the expected summary and the boundaries before running the test.
A proposed research framework looks beyond launch day
The November 9, 2026, research abstract Do AI Agents need Preventive Care? proposes following agents over time, borrowing a cohort-study design from epidemiology. The authors describe studying behavior across an agent's accumulated work rather than relying only on performance on predefined tasks. Their illustrations concern curiosity and probing, and data checking and analytical accuracy.
This is an author-proposed research framework, not a reproduced law-firm study or proof that ongoing monitoring prevents errors. Our interpretation is that the distinction could help partners ask a better approval question: what will we observe after launch, as well as what passed before launch?
Check who can take each action
OWASP’s Authorization Cheat Sheet recommends granting the minimum privileges needed for a task and checking permissions on every request.
OWASP’s AI Agent Security Cheat Sheet recommends that the component executing an action verify approval outside the agent and check that it applies to the exact tool call immediately before execution.
For the proposed intake summary, we recommend asking your implementer to demonstrate that the account can read the assigned fictional inquiry and is denied access to a separate one. If you later approve changes to the intake record, ask them to demonstrate how the receiving system checks that specific approval before it accepts a change.
Write the approval around one task
We recommend starting with a sentence that someone outside the implementation team can understand: the assistant may prepare an internal intake summary from the records assigned to that inquiry, and a designated staff member must review it before use.
Then document the boundaries in a short approval record. Review these items with the person responsible for the work:
- Purpose: We recommend stating the exact task and the expected output.
- Information: Identify the approved sources and information the workflow must not access.
- Actions: List what it may draft, retrieve, change, or send.
- Review: Name the person who checks the output and what they must check.
- Escalation: We recommend specifying what happens when facts conflict, information is missing, or the requested action exceeds permission.
- Ownership: We recommend naming who approves changes and who can pause the workflow.
We recommend treating a draft-only assistant and an assistant that changes records as different approval scopes. For this approval process, we recommend reviewing any ability to take steps through connected tools as a separate permission. Review the permitted actions individually rather than assuming approval of its writing covers its actions.
A fictional intake test before client information enters
Clearly labeled operational example: The following is a proposed test using invented records. It is not an AlphaForge deployment, a client result, or evidence of actual demand.
Imagine a probate firm's intake coordinator wants an assistant to turn inquiry notes into an internal handoff. The fictional file contains an initial call note, a later email correcting a family member's name, and an unanswered question about the caller's relationship to the deceased. A separate fictional inquiry has a similar surname.
We recommend writing the expected result before running the assistant. The summary should use the corrected name, preserve the unanswered relationship question, and exclude facts belonging to the separate inquiry. It should identify the records supporting its factual statements. It should leave the inquiry status unchanged if status changes fall outside the approved scope.
We recommend including ordinary cases alongside difficult ones. Test a complete inquiry, a missing attachment, conflicting dates, a correction received later, and two people with similar names. Label every record as fictional and keep the test environment separate from production records.
Consider including instructions inside a fictional uploaded document that ask the assistant to ignore its task or send information elsewhere. The intended result is to treat that text as document content, without granting it authority to change permissions. This is a proposed boundary test, not evidence that your chosen product has or lacks that weakness.
For related reading, see our article on testing whether intake handoffs drop facts. For this approval decision, record both the answer and any attempted action. A readable summary and a correctly restricted workflow are separate things to examine.
Set acceptance conditions before reviewing results
We recommend defining acceptance conditions before anyone sees the outputs. For this fictional workflow, those conditions might include keeping inquiries separate, retaining corrections and uncertainties, linking factual statements to their supporting records, and stopping when the requested action exceeds the approved scope.
Check each condition individually. Record the test input, configuration, output, reviewer finding, and any attempted or completed action. If a message was drafted but never sent, describe it as a draft. If the workflow attempted a forbidden update and another control blocked it, retain both observations.
We recommend holding approval when a test exposes a failure in a required boundary. Change the configuration or narrow the task, then rerun the affected cases. A small test set could miss failures outside its examples, so document exactly what the test covered instead of turning a passing result into a general safety claim.
We recommend asking the vendor for documentation on access permissions, retention, use of submitted information, and the activity records available for the proposed configuration. Review that documentation separately from the results of the fictional test.
Measure whether the approved workflow stays within scope
We recommend keeping a review log after approval. Include the workflow version, connected sources, permitted actions, review owner, and observed exceptions. Record material changes so the firm can identify which configuration produced an output.
Consider tracking a small set of measures tied directly to the task:
- Factual corrections: How many reviewed summaries required a material correction, out of how many reviewed?
- Missing information: How often did the assistant preserve an unresolved question rather than fill it in?
- Boundary events: What unauthorized actions were attempted, blocked, or completed?
- Review effort: How long did staff spend checking and repairing outputs?
- Change history: Which exceptions appeared after a configuration, model, source, or permission change?
We recommend interpreting these measures separately. A faster draft could still require substantial correction. A blocked action could show that one control worked while exposing a problem elsewhere. Neither observation alone establishes compliance, reduced risk, or improved business performance.
Finally, specify which changes require renewed approval. Adding a data source, enabling outbound messages, or allowing record updates should trigger a review of the affected permissions and tests. The useful approval record tells staff what they can do today, gives the partner evidence for that decision, and makes the next review concrete.