What records should a company keep when employees use AI agents?
A handover brings the recorded AI work and its verifiable receipts into one proof link. Review the evidence behind each claim, check what was and was not observed, then Accept or Flag.
Scope: an operational records framework, not a compliance certification. Regulated duties vary by sector and jurisdiction; your counsel and your auditors set the binding requirements.
Companies discover their record gaps at the worst moments: a client audit, an incident review, an employee departure. The framework below is sized for a company adopting agents now, before the gaps compound. Class one, the work record: per job, the steps, the saved results, and the producer labels, kept in a format a third party can recompute (AER-1 receipts, IETF Internet-draft-zambo-aer1-21, are the Zambo form of this class). Class two, the approval record: consequential outputs carry a named human decision; the handover's acceptance record is the pattern.
Class three, source states: for outputs the business relies on, what was resolved behind them, what was not resolved, and what was never checked. Class four, the spend record: the limits set per agent or team, and every refusal when a limit stopped a call, which is the record that proves the control works. Class five, the access record: which agents hold which connections, reviewed on a schedule, and revoked on departure the same day as the laptop.
Assign each class an owner and a retention period, and write both down. Retention is a policy choice informed by your contracts and your sector; the failure mode this framework prevents is not a wrong period, it is no period, no owner, and a record that lives only inside one employee's chat history.
Framework: the five record classes for employee AI work
Class | Contents | Owner | Review cadence. Work record | Steps, saved results, producer labels per job | Team lead | Monthly sample. Approval record | Named decision, decider, date, per consequential output | Approver of record | Per job. Source states | Resolved / not resolved / not checked, per relied-on output | Job owner | Per job. Spend record | Limits set, refusals logged | Finance or ops | Monthly. Access record | Agent connections and tool grants per employee | IT or ops | Quarterly and at every departure. Retention: set per class from your contracts and sector duties; record the period next to the class.
Worked example
Classes one and two exist as one public object: https://zambo.dev/handover/45434a17-4afc-486a-9646-e689bcee4297 shows a job's steps with saved results and the client decision recorded on the same record. A company adopting the framework can point its first audit at objects shaped exactly like this.
Frequently asked questions
Do we need all five classes from day one?
Start with the work record and the access record; they are the two that cannot be reconstructed later. Add approvals for consequential outputs next, then source states and spend as agent use spreads beyond one team.
Who should be able to read these records?
The record's owner, the approver of record, and whoever runs an audit or incident review. Client-facing records (handovers) are shared by link per job; internal records follow your normal access controls, reviewed with the access record itself.
How long should we keep AI work records?
Match the longest obligation that touches the work: client contract terms, sector rules, and your tax record periods are the usual drivers. The framework's demand is that the period is chosen, written, and owned, not that one number fits every company.
By Brennan Zambo, founder of Zambo.
Last verified: 2026-10-17.
Try a handover
A handover gives your client one proof link with verifiable receipts to review, then Accept or Flag.
try it: say 'use zambo to hand this over to <client>' in your ai