How do you hand work between AI agents and people?

Through the case record, never a side channel. One party owns the case at any moment, and the record says which. When the agent hands over, it attaches a packet: what is done, what is needed, why it stopped, its sources, and the deadline. When the person decides, their decision goes on the record and the agent resumes from there.

Where do hand-overs go wrong?

Three places. Ownership blurs: the agent flags a case and carries on, the person picks it up, and both act. Context is lost: the person receives "needs review" and has to redo the agent's reading to find out why. The way back is missing: the person decides, but the decision sits in an email, and the agent either never resumes or repeats work already done.

All three are design problems, not agent problems. They happen just as often between two people; agents make them more frequent because hand-overs multiply.

What should a hand-over carry?

The "what is needed" field does most of the work. "Please review" is not a request. "Confirm whether the utility bill dated four months ago is acceptable as proof of address; our policy says three" is.

What are the practical steps?

1. Map every hand-over point. From the process document: where the agent hands to a person, and where a person hands back. Each gets a reason code.
2. Use one case record. Both agent and people work in it. No hand-overs by email or chat.
3. Make ownership explicit. A case has one owner at a time. Handing over transfers ownership; the agent stops acting until it is handed back.
4. Standardise the packet. The fields above, filled by the agent at every hand-over.
5. Route to named queues. Each reason code goes to a queue with named people and cover.
6. Record the decision. The person's decision is a structured entry on the case, not a note.
7. Resume from the decision. The agent continues from where the person left it, without re-opening what they decided.
8. Watch ageing. Cases waiting with people beyond their deadline are escalated.

Which work should sit on which side of the line in the first place is covered in deciding what stays with people when an AI agent joins a team.

What does this look like in practice?

Take a bank's account opening team. The agent checks identity documents, proof of address and sanctions screening results for new personal accounts.

An application arrives with a utility bill four months old; the bank's policy accepts bills up to three months old. The agent hands the case to the onboarding queue with the reason "proof of address outside policy age", the bill attached, the policy section cited, and its proposed action: request a newer document. An officer sees the applicant has just returned from living abroad, accepts a bank statement already on file instead, and records the decision with the reason.

The case returns to the agent, which reads the officer's decision, completes the remaining checks, and opens the account. It does not ask for a new utility bill, because the decision on the record has settled that question.

Which controls keep hand-overs safe?

• One owner at a time, shown on the case.
• No action by the agent on a case it has handed over.
• Structured decisions the agent can read, not free-text notes.
• Ageing alerts on cases waiting with people.
• A daily count of open cases by owner, so nothing falls between agent and team.
• Capacity planning for the queues; see how many people you need to supervise AI agents.

Running the mixed team day to day is covered in managing a team of AI agents and people. Design and test hand-over points before go-live; see deploying an AI agent: train, shadow, certify, widen.