The short answer
If any record in the dataset came from a client's project, even de-identified, tell them, and ask first where the confidentiality clause is broad. If the dataset holds only your own SOPs, templates and timesheets with names removed, you are not obliged to, but a short note still protects the relationship.
On this page
There are two questions hiding in this one. Do you have to tell clients? And should you? The first is a contract question with a fairly clear answer. The second is about the relationship, and it is the one founders get wrong.
When the contract decides for you
Read the confidentiality clause before anything else. If it defines confidential information as “all information relating to the Client or the Services”, a brief with the client’s name removed may still count. Disclosing it to a buyer without consent would be a breach, de-identified or not.
The IP assignment clause matters too. If it assigns “all materials created in performing the Services” rather than “deliverables”, that account’s revision log may belong to the client outright. In both cases the answer is not “tell them”. It is “ask them first, and leave the records out if they say no”.
When the contract is silent
Plenty of agency contracts say nothing that reaches internal process records. Timesheets, your own QA checklists and SOPs are usually yours. Here you are not obliged to say anything. The question is whether silence is wise.
| What is in the dataset | Obliged to tell? | Wise to tell? |
|---|---|---|
| Only SOPs, templates and onboarding guides | No | Optional |
| Timesheets with staff and client names removed | Rarely | Usually |
| Revision logs and briefs from client projects, de-identified | Often (check the clause) | Yes |
| Anything from one large, recognizable account | Likely | Yes, and ask first |
Why clients find out anyway
Procurement questionnaires are adding AI questions. A client’s security team will ask whether you have shared project information with AI developers, and the honest answer has to match what you did. Buyers’ own processes also create a paper trail: micro1 says companies may review representative samples, and Scale AI says each package is approved before it is shared. Those are approvals your staff will have given, in writing.
The damage in these situations rarely comes from the licensing itself. It comes from the client hearing about it second-hand.
What to put in the note
A good note is short and specific. It should say:
- What record types went (for example: timesheets, internal QA notes, revision logs).
- What did not (deliverables, source files, the client’s own material, anyone’s personal data).
- How the records were de-identified, and by whom. Where anonymizing happened on your own systems before export (Scale AI works that way), mention it.
- That no client name, product or customer is identifiable in what was licensed.
- Who to call with questions.
What it should not do is apologize, or bury the point in a paragraph about innovation. Clients read tone. A plain note reads as confidence; a padded one reads as guilt.
Asking before rather than after
For any account whose work is recognizable even without the name (a single large retainer, a distinctive sector, a famous campaign), ask before you include it. The request is the same note in the future tense. Most clients care far more about their customers’ data and their unreleased plans than about your timesheets, and a request that explicitly excludes both is easy to approve.
If a client says no, leave that account out. The archive of an agency with a few dozen clients does not depend on any one of them, and the goodwill is worth more than the extra rows.
A note on staff
Your people are in these records too: their names on timesheets, their comments on revisions. Remove names, and consider telling the team what you are doing and why. An agency that tells clients but not its own staff has the order wrong.
If you have already licensed without telling
It happens: a founder signs, the records go, and only later does someone reread the client contracts. The fix is the same note, sent now, with one more line: when the license was signed. Check first whether any client’s records went out under a broad confidentiality clause; if so, talk to your lawyer before you write to that client, because the conversation is about a possible breach rather than a courtesy.
Then tighten the process for next time. Keep a simple register of which accounts’ records are in which license, with the clause you relied on. If a client asks in two years, you will want the answer in minutes, not a week of digging.
What clients usually ask back
Expect three questions back: will our name appear anywhere, could a model reproduce our material, and can we opt out of future licenses. Answer the first from your de-identification process, the second from the scope of what went (no deliverables, no client material), and the third with a yes. Offering the opt-out before they ask for it settles most of the rest.
Where to read more
The front page explains how to sell agency data in full, from ownership to buyers’ published terms. The clause-level detail on what you own is on client-owned vs agency-owned.
Clients forgive a lot from an agency that told them first.