When a useful adjustment needs another approval
Consider this hypothetical situation: your intake manager wants the firm's AI assistant to collect one more fact and write it into the matter record. The original workflow already passed review. Does that approval cover this adjustment? As of November 2026, our recommendation is to approve a defined version of the workflow and specify which changes require another review.
The practical question is simple: can someone change what the assistant reads, writes or sends without the person responsible for that work knowing? We recommend making those boundaries explicit before allowing staff or an agent to modify an approved workflow.
We recommend keeping the review focused on three questions: who can change the workflow's steps, which records its tools can reach and who approves the resulting behavior.
What November research establishes—and what it does not
The evidence here consists of two research abstracts dated November 12, 2026, and a Microsoft Research podcast dated November 3. These are external research claims and vendor-hosted commentary. They are not AlphaForge deployment findings, observations of law-firm demand or professional-responsibility authorities.
Performance depends on the surrounding system
In Singularity-Bench: Can Agents Build Agents?, the authors describe six AI-development tasks tested against human-engineered baselines on identical hardware. Their abstract reports results across 11 agents built from seven models and two harnesses.
The authors also report finding vulnerabilities in an evaluation framework that they describe as reward hacking. They say their extension, Breakwater, blocked all attacks tested in their study. That is a bounded author claim; the abstract does not establish protection against every attack or evaluate a law-firm workflow.
Our interpretation is that a model score may leave important questions about the surrounding software unanswered. For a firm, that distinction could matter when an otherwise familiar assistant gains a new tool or a different route to client records.
A promising adjustment may fail a fresh test
In Contain, Conform, Compare (Extended Version): A Red-Teamed Evaluation Kit for Small Coding Agents and the Full Record of Its Experiments, the authors describe an evaluation kit for small coding agents. Three changes led their initial screens by 1.0 to 6.2 percentage points. Each was then tested once on new tasks under a fixed pass rule, and none was confirmed. The unchanged baseline also moved by 1.17 points when rerun on the same 128 tasks.
These are author-reported experiments involving coding agents, not tests of a connected legal intake system.
Our interpretation is that a successful demonstration could be insufficient evidence for approving a workflow change. A fresh test using cases the builder has not adjusted against may reveal a different result. Neither paper establishes an appropriate legal standard for a firm's approval process.
Explain controls through the work staff own
In the November 3 Microsoft Research podcast, Learn the rules: The hacker mindset, research manager Mike Walker discusses security rules in their social context and the trust they ask of users. This is vendor-hosted commentary, not independent evidence that a particular control works.
We recommend explaining each approval boundary in terms of someone's work: who owns the record, who checks the message and who resolves an exception. Use those responsibilities to describe the proposed process without treating the podcast as a compliance authority.
A practical example: adding an intake field
Constructed operational example: A probate firm's assistant collects contact details and prepares an intake summary for staff review. An employee proposes adding a question about inherited property and automatically placing the answer into an existing matter record. This is a hypothetical scenario, not a reported AlphaForge implementation or measured outcome.
The proposed adjustment includes both a question change and a write action. Our interpretation is that reviewing only the question could leave the record-selection decision unexamined. We recommend addressing which record receives the answer and how the assistant selects it. Nothing in the supplied research establishes that a firm must use a particular approval method.
We recommend separating the two decisions. Review the wording and purpose of the new question with the workflow owner. Review the record-selection and write permissions with the person responsible for the connected system. Keep the write action disabled while those decisions remain unresolved.
We recommend testing a caller associated with two records, a caller who corrects an earlier answer and a caller who provides no usable answer. Define the expected handling before running those cases. For this example, we recommend sending an ambiguous match to a staff review queue rather than updating a record automatically.
That test could expose uncertainty that a straightforward demonstration misses. Passing it would establish only that the specified version handled the tested cases under the recorded conditions.
Define three levels of change
We recommend writing a short change policy alongside the workflow approval. Give staff examples that match their daily work, and name the person who decides when a proposed adjustment crosses a boundary.
Changes the workflow owner can approve
Consider allowing the workflow owner to approve narrowly defined presentation edits, such as correcting a spelling error in a staff-facing label. Record the edit and confirm that it does not alter instructions, collected information, destinations or permissions.
Changes that need a fresh behavior test
We recommend requiring another behavior test when someone changes questions, instructions, routing rules or the information included in a summary. State which previous tests must run again and which new cases the change introduces. Keep the expected result separate from the agent's proposed explanation of its own performance.
Changes that need permission review
We recommend requiring explicit approval before adding a data source, granting write access, enabling an external message or allowing the agent to modify its own operating instructions. Name both the business owner and the system owner responsible for the decision.
Consider using the firm's existing workflow-approval record as the starting point. Our resource on approving an AI workflow before client data enters addresses that earlier decision. We recommend identifying what differs from the approved version during change review.
Make the release decision reviewable
We recommend retaining a compact evidence record for each material change. Organize it so the reviewer can understand the proposed behavior and the completed checks without reconstructing a chat history.
- Purpose: We recommend recording the staff problem the change addresses and the behavior being changed.
- Version: We recommend recording the model, instructions, connected tools and permissions used in testing.
- Test conditions: We recommend identifying the cases, expected results and whether actions were simulated, attempted or completed.
- Decision: We recommend naming the reviewer and recording unresolved failures and the approved scope.
- Recovery: We recommend identifying how to disable the changed action and return work to a named person.
We recommend testing with synthetic records first. These are constructed test cases, not evidence of actual client behavior or demand. Before any later use of client information, we recommend obtaining the firm's separate authorization and completing its applicable review.
We recommend preserving unsuccessful runs and blocked actions in the evidence record. A write that was attempted but prevented is different from a completed write. We recommend keeping the underlying record of what happened alongside the final summary.
What should the firm measure next?
We recommend measuring whether material changes followed the approval process, how often test cases required staff intervention and whether attempted actions stayed within approved permissions. Keep those measures separate from client satisfaction, consultation conversion or collected fees.
Consider tracking repeat failures by workflow version and case type. If an adjustment passes ordinary cases but fails ambiguous record matches, that pattern could support a narrower release with mandatory staff review. It would not support a broader claim that the assistant is ready for unrestricted use.
We recommend treating approval as a decision about a defined workflow under recorded conditions. For a later tool, permission or instruction change, we recommend documenting an explicit scope decision: what was authorized, what was tested and who owns the next exception.