Approve a defined action on a defined version
An instruction to work on a matter can cover several very different activities: reading documents, suggesting next steps, drafting, communicating with a client, or submitting a filing. A useful legal AI workflow makes those boundaries explicit.
The design below is a proposed example for an internal research-and-drafting workflow. Its checkpoints are practical design choices, not a claim that one approval procedure satisfies every jurisdiction’s rules or every client’s requirements.
Start with an internal assignment
Suppose a lawyer wants an internal memo and proposed response letter based on an approved set of public or fictional documents. The agent may read that set, identify missing information, and create drafts in a review folder. The assignment does not authorize sending a letter, changing a deadline, filing a document, or committing the client to a position.
A reviewer should be able to see what was requested, which sources were used, what the agent inferred, and what remains unresolved. “Looks good” is too ambiguous if the system can interpret it as permission to take several downstream actions.
Put approval where consequences change
| Checkpoint | What the agent may prepare | What a person decides |
|---|---|---|
| Before information enters a tool | A list of needed documents and proposed access. | Whether the tool, information, permissions, and any required consent fit the matter. |
| Before the plan expands | A proposed task list with sources, assumptions, and dependencies. | Whether the task is within scope; whether new research, access, or cost is authorized. |
| Before a draft is accepted | The draft, supporting passages, unresolved questions, and validation notes. | Whether the facts, authorities, reasoning, and presentation support the intended use. |
| Before an external action | An exact recipient, document version, attachments, and action summary. | Whether this particular communication, submission, or other action may proceed. |
| Before information is reused elsewhere | A proposed reusable template or general lesson. | Whether it can be separated from matter-specific information and shared in that destination. |
For this example, start with read access and a place to save drafts. Give the agent a way to request further permission without granting the underlying action automatically. A human approval message is useful only if the connected tools actually enforce the boundary.
Example: a draft summons with an unresolved fact
In Jonathan’s agentic-lawyer demonstration, the validation notes identify uncertainty about prior summons issuance and service. That makes a concrete review task visible: check the court record and file before deciding what document or action is appropriate.
The reviewer should resolve the uncertainty, amend the draft if needed, and inspect the resulting version. A note admitting uncertainty helps direct the review. It does not establish that the remaining contents are accurate or that the agent identified every missing fact.
Separate acceptance from permission to send
Consider a fictional response to an insurer. A lawyer may accept the factual summary while rejecting the proposed tone or legal position. The application should allow approval of the internal memo without authorizing delivery of the letter.
For an external action, the approval record should identify the exact recipient, approved document version, attachments, purpose, and approving person. If the draft changes after approval, return it for review. If an action times out, first determine whether it succeeded; repeating a send or submission blindly can create a second action.
Treat deadlines as their own controlled records. A date extracted from a letter or calculated by a model can be a candidate for review. In this proposed workflow, it does not replace the matter’s verified deadline until an authorized person checks the source, applicable rule, and calculation.
Keep reusable knowledge separate from matter facts
Firm templates, formatting preferences, and approved general research can be useful across matters. Client facts, strategy, and confidential material need different handling. A feature called global memory does not answer who can access its contents or which information may be reused.
Before promoting a lesson out of a matter, identify the proposed text, its source, the destination, and who can read it. Approve only the material suitable for that destination. Test the boundary with fictional information before introducing client material.
Make the handoff inspectable
An approval record can be short: matter reference; requested action; source set; draft version; reviewer; decision; unresolved conditions; and time. Keep it with the approved document so the next person can tell what was reviewed and what was authorized.
If a required source is unavailable, an assumption remains material, or the draft changes, the agent should stop at the relevant checkpoint and explain what is needed. The fallback should be a person completing or revising the task, not an automatic increase in the agent’s authority.
Ground the workflow in professional responsibility
The Illinois Supreme Court’s AI policy announcement keeps responsibility for final work with its human users. ABA Formal Opinion 512 discusses competence, confidentiality, communication, supervision, and fees under the ABA Model Rules. Check the rules and orders that apply to the actual matter when implementing a workflow.
Approval also consumes time. Include it when measuring the completed assignment, and give reviewers the source connections illustrated in the case-brief verification method.
