Can a marketing agency sell its data to AI companies?

It can, provided every record in the deal is its own or covered by a right to license it. For most agencies that means process records (briefs stripped of client identity, revision histories, timesheets, QA checklists) rather than the finished work, which the client usually owns.

Do we hand over client deliverables?

Usually not. Most MSAs assign deliverables to the client on payment. What buyers want from an agency is how work got made, which is often the agency's own record. See client-owned vs agency-owned.

How big does the agency need to be?

Nyne sets the lowest printed bar, twenty or more full-time W-2 staff in the US after three years in business. It rises to 30 full-time staff at micro1 and a preferred 40 at Scale AI, as each page read on 7 Nov 2026. Labelbox Prism publishes no size rule.

Will the buyer see client names?

They should not. All four promise de-identification, and Scale AI does it on your side of the fence. Even so, take out client names and personal details yourself before a sample goes anywhere; a buyer's clean-up is a backstop.

What does an agency archive sell for?

Nobody publishes agency deals. Buyers publish ranges (from around $50K at Labelbox Prism to $3M+ at the top of micro1's tiers), and reported deals are often far smaller. The worth estimate shows which published ranges your agency's size fits.

Does licensing data mean the buyer owns it?

Not by default. On their own pages micro1 leaves ownership with the company, Labelbox Prism leaves the seller in control, and Nyne defers to the contract. The grant clause in the license is what you actually sign, so read that.

Is it legal to license recruiting records?

Candidate personal data is the hard part: resumes, contact details and interview notes are personal information about people who never agreed to AI training. See recruiting agency data for what can and cannot go.

Sell Agency Data

An agency can license its process records to AI training-data buyers: briefs with the client's identity removed, revision rounds, timesheets and QA notes. It usually cannot license the finished work, which the MSA hands to the client. micro1, Scale AI and Nyne are among the buyers that print their rules and a rough idea of pay.

The demand is for how work gets done. Labs training AI agents to run projects need the messy middle: the brief that changed three times, the round of feedback that sent a layout back, the hours logged against each task. Agencies hold years of exactly that, scattered across project tools and inboxes, and most have never been asked for it.

Before anyone asks for a sample, two questions decide the whole thing: does your client contract assign the work product to the client, and can you remove every personal and client detail? This issue walks through both, then the buyers and their published terms.

Get a quote

What does the agency own, and what does the client own?

Start with the MSA, not the archive. Work made for a client usually belongs to the client once the invoice is paid; that is what the IP assignment clause in most agency contracts is for. The logo, the campaign, the site build and the hiring shortlist are theirs to license, not yours.

The record of how you made them is a different matter. Timesheets, internal QA checklists, your own process documents and briefs stripped of anything that identifies the client are often the agency's. "Often" is doing real work in that sentence: some contracts define confidential information broadly enough to catch the brief itself.

Usually the client's

Final deliverables and source files
Client brand assets, data and research
Anything the contract calls confidential information
Their customers' personal data

Often the agency's (check the contract)

Timesheets and resourcing records
Internal process docs, templates and SOPs
Revision logs and QA notes, client identity removed
Briefs, once nothing in them identifies the client

Licensing starts from a plain test: the records are yours, or you hold a right to license them. Client-owned material is out, and so is anybody's personal data, unless both contract and law say it may go in. The full checklist, clause by clause, is on client-owned vs agency-owned.

Which agency records do AI buyers actually want?

The buyers' own lists are telling. micro1 names process documentation, SOPs, product planning and operations records. Scale AI names Jira and Confluence work records, documents with their metadata, email threads and calendars. Nyne names Asana and Jira project records. None of them asks for a portfolio.

Briefs and change requests: The starting intent and every amendment, dated. Valuable because the changes show judgment in context. Remove the client's name, product and anything a reader could trace back.
Revision rounds: Feedback, the response and the next version, in order. This is the record agencies hold in quantity and almost nobody else does.
Timesheets: Hours by task, role and phase over years. Tabular, structured and usually the agency's own record, though staff names need removing.
QA and sign-off: Checklists, bug tickets and approval steps: how a team decides something is done.
Process documents: Onboarding guides, SOPs, estimating templates. Plainly yours, and directly on micro1's published list.

Who buys agency data, and on what published terms?

Only four buyers print enough detail for a founder to test the agency against. Every figure below is paraphrased from that company's site, last opened 7 November 2026. These are advertised ranges; no agency's actual fee is public.

Published terms, as read 7 November 2026
Buyer	Published range	Who qualifies	De-identification
micro1	Three advertised tiers per company: $100K–$500K, $500K–$3M, and over $3M for rare operating data	At least 30 full-time staff and English-language records; buys mostly from American, British and Canadian firms; no set number of years, though a longer record helps	De-identified by micro1, and you may inspect sample outputs; location of that work unstated
Scale AI	An illustrative $10K–$1M+ per partnership, higher for more frequent deliveries; explicitly not a promise	Prefers 40 or more employees but judges each case; wants mature data systems, real workflows and one internal owner	Strips identities on your side before export, with your sign-off on every package
Nyne	$100K or more described as common, varying with the kind and volume of records	US firms with 20 or more full-time W-2 staff and at least three years trading; rights still checked	De-identified ahead of release; the page does not say by whom
Labelbox Prism	Roughly $50K at the small end and $1M+ for enterprise-scale data, paid as an upfront fee with revenue share on top	No general size or age rule (for support logs it leans toward big consumer-facing firms)	Labelbox does intake, cleanup and PII removal; where is not stated

Our interest, declared: micro1 and Scale AI pay introducers, so a signed deal with either after our introduction can earn us a fee. Rows simply follow the order we read the pages in, and no fee buys a better line in this table.

Mercor buys company data too. Its pages would not load for us on 7 Nov 2026, so we print nothing about its terms. Payment timing, exclusivity and referral terms for each buyer are on buyers compared, with a dated list of changes on the terms changelog.

What is an agency archive worth: published ranges against reported deals?

Two kinds of number circulate, and they should never be confused. Published ranges are marketing. micro1 advertises tiers from $100K past $3M; Nyne calls six figures common; Scale AI attaches the word illustrative to its own $10K to $1M+. Reported deals are what outlets have been able to confirm, and they rarely concern agencies.

Reported deals for business data
What was reported	Outlet
Google offered $10 million for the business records of Spirit Airlines (staff email, Teams chats, spreadsheets and calendars, de-identified before any sale), topping a $7.5 million Mercor offer, with a court still to approve it.	Reuters (Dietrich Knauth), 17 Aug 2026, as republished by The Star
Startups that label training data were said to be handing HVAC contractors around $150,000 apiece for their records.	The Information (Stephanie Palazzolo), 2026
The founder of a wound-down crypto startup netted roughly $25,000 for his code base.	MIT Technology Review (Matthew King), 13 Oct 2026

An HVAC firm and a 40-person agency are not the same seller, and an airline's entire corporate inbox is not a design studio's project tool. The honest reading is that a mid-sized agency's archive sits somewhere well below the top of any published range until a buyer has seen a sample. The worth estimate shows which published ranges your headcount and years put you inside. None of it is an offer, and real licenses tend to come in well under the ranges.

Should you tell clients you licensed process data?

If anything in the dataset came from their project, even with their name removed, the safer course is yes. Clients find out. A procurement questionnaire next year will ask whether you have shared project data with AI developers, and "we didn't think it counted" is a poor answer.

The question a client asks is rarely "was it legal?" It is "why did we hear about this from someone else?"

Read your confidentiality clause first. If it covers "all information relating to the project", even a de-identified brief may need consent. Where the contract is silent, a short note explaining what went (timesheets, revision logs), what did not (deliverables, their data) and how it was de-identified protects the relationship more than silence does. The longer version is in the blog piece on telling clients.

What can a recruiting agency license, and what is off limits?

Recruiting agencies hold the most personal records of any agency type: resumes, interview notes, salary expectations, references. Almost all of that is personal information about candidates who agreed to be considered for a job, not to train a model. Scale AI's page says outright that it excludes employee and customer identities.

What can go is the process with the people taken out: role specifications stripped of the client's identity, pipeline stage timings, your own outreach templates and screening rubrics. Whether even that is safe depends on your candidate privacy notice and the states you recruit in. The detail is on recruiting agency data.

How does a data licensing deal run?

Eligibility. The buyer checks size, years, country and record types against its published rules.
Scoping and samples. You agree which systems and date ranges are in. micro1 lets companies look over representative samples drawn from their records.
De-identification. Scale AI describes stripping identities on your own systems ahead of any export, and asks you to sign off each package. Elsewhere the location goes unstated.
The license. Scope, rights, exclusivity and acceptance criteria. Labelbox Prism licenses are exclusive by default. Have your own lawyer read it.
Acceptance and payment. micro1 promises the full sum inside 30 days once it accepts the data. Labelbox claims its Prism deals can go from first call to money in the bank in under three weeks.

Questions agency founders ask

Can a marketing agency sell its data to AI companies?: It can, provided every record in the deal is its own or covered by a right to license it. For most agencies that means process records (briefs stripped of client identity, revision histories, timesheets, QA checklists) rather than the finished work, which the client usually owns.
Do we hand over client deliverables?: Usually not. Most MSAs assign deliverables to the client on payment. What buyers want from an agency is how work got made, which is often the agency's own record. See client-owned vs agency-owned.
How big does the agency need to be?: Nyne sets the lowest printed bar, twenty or more full-time W-2 staff in the US after three years in business. It rises to 30 full-time staff at micro1 and a preferred 40 at Scale AI, as each page read on 7 Nov 2026. Labelbox Prism publishes no size rule.
Will the buyer see client names?: They should not. All four promise de-identification, and Scale AI does it on your side of the fence. Even so, take out client names and personal details yourself before a sample goes anywhere; a buyer's clean-up is a backstop.
What does an agency archive sell for?: Nobody publishes agency deals. Buyers publish ranges (from around $50K at Labelbox Prism to $3M+ at the top of micro1's tiers), and reported deals are often far smaller. The worth estimate shows which published ranges your agency's size fits.
Does licensing data mean the buyer owns it?: Not by default. On their own pages micro1 leaves ownership with the company, Labelbox Prism leaves the seller in control, and Nyne defers to the contract. The grant clause in the license is what you actually sign, so read that.
Is it legal to license recruiting records?: Candidate personal data is the hard part: resumes, contact details and interview notes are personal information about people who never agreed to AI training. See recruiting agency data for what can and cannot go.

Pages this relies on

micro1 (micro1.ai/data), seen 7 Nov 2026
Scale AI (scale.com/data-partnership), seen 7 Nov 2026
Nyne (nyne.ai/sell-company-data), seen 7 Nov 2026
Labelbox Prism (prism.labelbox.com), seen 7 Nov 2026
Reuters (Dietrich Knauth), 17 Aug 2026, as republished by The Star
The Information (Stephanie Palazzolo), 2026
MIT Technology Review (Matthew King), 13 Oct 2026