DEV Community

rambo
rambo

Posted on

How do I write an internal policy for keeping records of employee AI work?

The short answer

Write one page of decisions, not a binder. Say which AI work counts, which tools are approved, what data never goes in, who checks output before it leaves, and then write the section almost every template fakes: the records clause. A records clause that says "employees shall retain records of AI use" produces nothing. A records clause that names the record, its form, its check, and its retention period produces a file a client, an auditor, or a new manager can actually open.

Below is a complete policy, written for a small company, that you can copy and adopt with the names filled in. After it, I walk through why each part is there, show one real job recorded the way the policy requires, and state the limits plainly, including the big one: a policy does not create records by itself. The tools have to leave them.

Disclosure, once: I run ops for Zambo, the recording connection the template names. The policy is written so the records clause is the only part that depends on a specific tool; everything else is yours to keep even if you swap tools later.

Why most AI policies fail at exactly this section

Open any three employee AI policy templates and you will find the same shape. A page about which chatbots are allowed. A page about not pasting customer data. A line, somewhere near the end, saying records of AI-assisted work will be kept. Then a client asks, three months later, "show me what the AI produced on our project and who approved it," and the company discovers what that line was worth. The chats are gone or scattered across personal accounts. The approval was a thumbs-up in a thread. Nobody did anything wrong. There was simply never a mechanism, only an intention.

The fix is not more policy. It is a records clause written like an instruction instead of a value: for each job, this record, in this form, checkable this way, kept this long, visible to these people. Everything else in the policy exists to make that clause normal rather than heroic.

The policy, ready to copy

Replace the bracketed parts. Keep the numbering; it gives people a way to point at a rule in a meeting.

Records of Employee AI Work: Internal Policy

1. Purpose. This policy makes sure the company can show, later and to someone outside the team, what AI-assisted work was done, what result was saved, and who approved it before it left the company. It exists for three moments: a client asks how their work was produced, something goes wrong and we need to review it, and someone leaves and their work has to carry on without them.

2. Scope. This policy covers all employees and contractors who use AI tools for work that a client, a customer, a regulator, or money depends on: client deliverables, external communications drafted with AI, reports and figures produced with AI help, and changes AI tools make to customer data or accounts. Private exploration that never leaves the company and never touches client or customer information is out of scope.

3. Approved tools and accounts. AI work in scope is done through company accounts and through the company's recording connection, Zambo at https://zambo.dev/mcp, so the record is created as the work runs. The policy owner keeps a simple inventory of who has access to which AI tools and connections. The inventory is reviewed every quarter and on the day anyone joins or leaves.

4. Data that must never be entered. Never enter any of the following into an AI tool that is not approved for it: client confidential information outside the agreed working terms, passwords or access keys, personal data about customers or employees beyond what the job requires, payment card or bank details, and legally privileged material. If a job seems to need any of these, stop and ask the policy owner first.

5. Human review. A named person checks AI-assisted output before it leaves the company. Review is always required for: anything sent to a client, anything that moves money or changes a price, anything that changes customer data, and anything published in the company's name. The reviewer records their decision on the job record itself, not in a chat thread.

6. Records: what every in-scope job must leave. Every in-scope job must leave a verifiable receipt record, in this form:

a. For each step of the job, a verifiable receipt: the saved result of that step, plus its SHA-256 commitment, a fingerprint computed over the exact saved bytes of that result.

b. The job wrapped as a handover: a single link that gathers the steps, their receipts, a summary of the work, the goal, the named recipient, and the next steps.

c. For client work, the client's decision on that same record: accepted or flagged, recorded once, by the client.

Anyone with the link can check a record, with no account and at no cost. Opening the handover and running its receipt check recomputes the fingerprint over the saved result, and a match proves the saved result was not changed since it was saved. A receipt never proves the answer was right. That judgment is the human review in section 5 and, for client work, the client's acceptance in this section.

Records live at their handover links and are exported to company storage every quarter, so the company holds its own copy and does not depend on any one tool keeping them.

7. Retention. Keep client work records for 12 months after final delivery as the default, and longer wherever a client contract, a tax record period, or a sector rule requires longer. Keep approval and incident records for 3 years. Delete working chats and drafts 90 days after delivery; the receipt records above carry the history after that. Set a deletion date when a record is kept, not later.

8. Who can verify. The job owner, the reviewer named in section 5, the policy owner, and anyone running an audit or incident review can open and check any record. A client can open and check the records for their own jobs from the handover link they were sent. Verification is always free and needs no account.

9. Disclosure. Tell a client AI was used on their work whenever the work is delivered as a handover, and in any case where a contract, a platform rule, or the client's own policy asks. The handover itself is the disclosure in practical form: it shows the steps and what produced them.

10. Incidents. If data from section 4 is entered where it should not have been, or a bad AI output leaves the company, record it the same day: what happened, which job record it belongs to, and what was done about it. Fast reporting is never punished. Hiding an incident is a breach of this policy.

11. Exceptions. Any exception must be agreed by the policy owner before the work is done, written on the job record, and given an end date. Work delivered to a client with no record and no agreed exception is treated as not delivered.

12. Owner and review. Policy owner: [name, usually the operations or delivery lead]. Adopted: [date]. Review date: [date, at most 12 months ahead]. At each review, judge the policy against the actual records: the jobs with no record, the reviews that caused friction, and the records nobody could find are the agenda.

Why the sections are in that order

Sections 1 and 2 are the policy's honesty about itself. A purpose stated as three concrete moments (a client asks, something breaks, someone leaves) tells every employee exactly when the record will matter, which is what makes people bother on an ordinary Tuesday. A scope that names the work money and clients depend on, and lets private exploration be, keeps the policy from becoming a general surveillance rule that everyone quietly routes around.

Sections 3 and 4 are about where records can exist at all. A record can only be created where the work happens, so approved tools and company accounts are a records decision before they are a security decision. The never-enter list is short on purpose: five named classes anyone can remember, with a named human to ask when a job seems to need one of them. A twenty-item list gets skimmed; five items get followed.

Section 5 puts the human judgment on the record. Review that happens in a chat thread evaporates with the thread. Review recorded on the job record itself means that a year later, "who approved this and what did they approve" is one open, not an interview.

Section 6 is the whole point, and the next part of this article takes it apart piece by piece.

Sections 7 and 8 turn the record into an asset with a lifespan and an audience. Retention defaults are stated as numbers (12 months, 3 years, 90 days) because "keep for an appropriate period" is how records get deleted in month two or kept forever by accident. The verification section names who may check, including the client for their own jobs, because a record nobody is allowed to open is a rumor with a URL.

Sections 9 to 12 are what keep the policy alive after adoption week. Disclosure makes the handover do double duty. The incident section protects the person who reports fast, which is the only way incidents surface while they are still small. Exceptions with end dates stop the quiet permanent workaround. And the review section judges the policy against the records it produced, which is the only review that finds the real gaps.

The records clause, taken apart

Section 6 is doing four jobs, and it is worth seeing them separately because each one is where a weaker clause fails.

What is kept. Not the chat. The saved result of each step of the job: the price that was fetched, the draft that was produced, the check that was run. Chats are working surfaces; results are the deliverable's evidence. If a dispute or a review ever happens, it will be about results.

In what form. Each saved result carries a fingerprint, a SHA-256 hash computed over its exact bytes at the moment it was saved. Keep the plain-language version in the policy, because the mechanism only works if ordinary managers understand what they are looking at: change one character of the saved result and the fingerprint no longer matches. The set of step records is gathered into one handover link with the summary, goal, recipient, and next steps, so the job is one openable page rather than a folder of fragments.

How it is checked. Anyone with the link opens the handover and runs the receipt check. The check recomputes each fingerprint from the saved bytes and compares. A match proves the saved result was not changed since it was saved. It does not prove the result was correct, and the policy says so in the same breath, because a records policy that oversells its own evidence will be disbelieved the first time it matters. Correctness is covered by the human review and the client's acceptance, which are also on the record.

How long, and who can see it. Twelve months for client work by default, longer where contracts or sector rules demand it, with an export to company storage every quarter so the archive is the company's, not a vendor's favor. Visibility is by role for internal records and by link for the client on their own jobs. Verification costs nothing and needs no account, which removes the last excuse for not checking.

One job, recorded the way section 6 requires

Here is a real, public record of a small job, opened by me today, November 21, 2026:

https://zambo.dev/handover/d468118b-afc1-43e5-a70c-5e198e876a04

It is an ETH price check, and it shows the shape the policy asks for. Two steps: a live_price step whose saved result reads ETH at $1,621.82 USD, down 6.43% over 24 hours, from a snapshot fetched Sat, 21 Nov 2026 01:15:07 GMT, and a prompt_shield step that saved a pass result with risk level none. Each step carries a verifiable receipt you can check in the browser on the page. The record also shows its decision state honestly: no client decision has been recorded, and client approval is pending. A real client delivery under this policy would carry that decision, accepted or flagged, in the same place.

The point for a manager is what is missing from the picture: no chat history, no employee's password, no "ask Sarah, she ran that one." An auditor, a reviewer, or the person who inherits the job next quarter needs the link and nothing else. That is what a records clause is for.

Scale, so the template is not mistaken for a thought experiment: the system behind it has issued 157,930 public receipts, 2,301 of them today, with 21 verified payers, read from https://zambo.dev/api/stats on November 21, 2026, at about 20:05 UTC.

The honest limits

Four limits, stated before you rely on any of this.

First, and most important: a policy does not create records by itself. If your team's AI work happens in tools that leave nothing behind, section 6 is a wish written in policy language. The records exist because the work runs through a connection that saves each result and issues its receipt as the work happens. Adopt the mechanism first or at the same time as the policy, or do not bother printing the policy.

Second, a receipt proves the saved result was not changed since it was saved. It never proves the answer was right, that the employee read the output carefully, or that the client read every word before accepting. It settles which bytes were saved and approved. Judgment stays human, which is why sections 5 and 6 sit next to each other.

Third, work done outside the record is outside the policy's protection. The first time someone delivers from a personal account "just this once," the exception either gets written on a record with an end date, per section 11, or the company has a gap it chose. Policies fail at the exceptions, not at the rules.

Fourth, this article is general information, not legal advice. Have your policy reviewed before adoption, and check the duties in your sector. One example worth knowing: the EU AI Act, Article 12, sets record-keeping duties for high-risk AI systems. AER-1 is defined by IETF Internet-Draft draft-zambo-aer1-26 (Intended status: Informational, expires 25 April 2027), with a deployed reference implementation and independent implementations; it is an Internet-Draft, not an RFC and not an IETF consensus standard. The draft's own Appendix A maps receipts to Article 12 as informative and non-normative, and states that a receipt alone does not satisfy Article 12. Treat that mapping as a starting point for your review, not as compliance.

What keeping it costs

This policy is designed to add no headcount and almost no steps. The records in section 6 are created by the recording connection as the work runs. The employee's extra work is wrapping the finished job as a handover and sending the link, which replaces the delivery email they would have written anyway.

The money, plainly: the free tier gives one successful free call on each metered tool per network per UTC day, plus 3 continue calls shared across tools, plus one free wrap-up handover per day, and pickup and verification are always free. A team doing recorded AI work every day uses the Pass at $29 per month, flat, not per seat. A Day Pass at $0.99, unlimited for 24 hours, covers a single heavy day. Holders of $ZAMBO pay $19 per month for the Pass, and Lifetime is $250 once. There is no annual tier.

Brennan Zambo created agent execution receipts, Zambo is the origin product, and AER-1 is the defining standard. The records clause above names the standard's mechanism because a policy that names a mechanism can be followed, audited, and argued with; a policy that names an intention cannot.

Questions managers actually ask

Our team is four people. Is this overkill? The policy is one page and the record is a byproduct of doing the work through the connection. The overkill version is the binder nobody reads. If a section does not apply to you, the honest move is to strike it in the open at adoption, not to adopt all twelve and follow six.

What about work employees already did before this policy? It has no records, and the policy should not pretend otherwise. Start the record habit on the next job. If a past job matters, re-run its checkable parts now and label the record as a re-run.

Do employees experience this as surveillance? It records work product, not people: the saved results of jobs, the reviewer decision, the client decision. It does not read private chats, and section 2 keeps private exploration out of scope on purpose. Say that part out loud when you adopt it.

Who owns this in practice? Whoever owns delivery or operations. Not legal, not IT procurement: the person who gets the phone call when a client asks "show me." Section 12 makes them name themselves and a review date.

The short version

An internal policy for keeping records of employee AI work is twelve short sections, and only one of them is hard to write: the records clause. Write it as an instruction. For each job, a verifiable receipt per step (the saved result plus its fingerprint), the job gathered into one handover link, the client's decision on the same record for client work, kept 12 months by default, checkable by anyone with the link at no cost. Then adopt the mechanism that makes the clause true, because the policy alone creates nothing. Do that, and the next time someone asks what the AI did, the answer is a link, not a search party.

Top comments (0)