AI Intake Human Handoff: A Playbook for PI Firms

A strong handoff is not a notification that says “new lead.” It is a transfer of context and responsibility. The person receiving it should know what arrived, what the system said, what remains unknown, and what decision belongs to them.

That distinction matters more when AI is involved. ABA Formal Opinion 512 says lawyers using generative AI must consider duties that include competence, confidentiality, communication, supervision, and candor. NIST's AI Risk Management Framework similarly emphasizes documented roles, scope, oversight, testing, and ongoing risk management.

1. Draw the boundary in verbs

“AI-assisted intake” is too vague to supervise. Write a list of verbs the system may perform and a list it may not perform. The boundary should be readable by attorneys, intake staff, and the implementation team.

The exact line depends on the firm's jurisdiction, policy, and approved workflow. ABA Model Rule 5.3 addresses measures and supervision intended to keep nonlawyer assistance compatible with a lawyer's professional obligations. Formal Opinion 512 applies established duties to generative AI tools rather than treating AI as an exception.

2. Set handoff triggers before writing prompts

Prompts are implementation details. Escalation rules are operating policy. Define the events that should stop the ordinary automated path and place a person in control.

• the person asks for legal advice, a case-value estimate, or a promise about outcome;
• the conversation suggests immediate danger, medical emergency, or another urgent situation;
• the person requests an attorney or asks to stop automation;
• identity, consent, language, or contact details are unclear;
• the model produces low-confidence, contradictory, or out-of-scope output;
• a message fails, is undelivered, or receives an opt-out request;
• the workflow reaches the firm's maximum number of unanswered attempts.

Each trigger needs a destination, an owner, and a time expectation. “Escalate” without those three fields is only a label.

3. Design a review packet, not a black-box summary

An AI summary can be useful, but it should not replace the source record. The handoff package should let a reviewer verify the summary quickly.

1. Source and timing: channel, received time, timezone, response events, and current state.
2. Identity and contact: the submitted name and contact details, with any uncertainty made visible.
3. Raw conversation: the messages in order, including system and opt-out events.
4. Collected facts: fields mapped to the questions the firm approved.
5. Missing or conflicting facts: unanswered questions and contradictions, not hidden blanks.
6. Machine-generated summary: clearly labeled, with a link back to the source conversation.
7. Requested next action: review, call back, take over, or investigate a failure.

NIST's AI RMF materials emphasize documentation and clearly defined human oversight. In practice, preserving the source record makes the human review real rather than ceremonial.

4. Close the acknowledgement loop

A handoff is incomplete until somebody or some defined queue accepts responsibility. Add an acknowledgement event and record who acted. If nobody acknowledges within the firm's selected window, route the item to a second destination or surface it visibly as overdue.

Do not interpret a successful webhook or CRM write as a human acknowledgement. System delivery and operational ownership are separate states.

5. Test the uncomfortable paths

The happy path is the least informative test. Before launch, use synthetic data to verify:

• an ordinary English intake and an ordinary Spanish intake;
• a request for legal advice;
• a possible emergency or safety concern;
• an opt-out in each supported language;
• an undelivered message and a disconnected integration;
• a staff takeover in the middle of the conversation;
• a missing required field and a contradictory answer;
• the emergency stop and the audit record it creates.

Record the expected result before running each test. If the team cannot agree on the expected result, the policy is not ready to automate.