How AI Data Licensing Works: LOI → De-identification → Payment (Timeline & Agreement Terms)

AI data licensing lets a company keep its records while granting a buyer rights to use a de-identified copy for AI training. In current programs the sequence is: an estimate, a nonbinding letter of intent, a signed agreement and approved scope, export and de-identification over a few weeks, then payment, typically within one to three months.

What is AI data licensing?

A data license is a contract that grants someone the right to use data you own, on defined terms, without transferring the originals. In AI deals, the buyer uses the copy to train or evaluate models, or passes it to developers who do.

Two very different markets use the same phrase:

• Content licensing. Publishers and platforms license articles, posts, images or archives to AI developers. Reddit, the AP, News Corp and Shutterstock have announced such deals; see our deals tracker.
• Operational data licensing. Ordinary companies license internal records, such as email, chat, tickets, CRM and billing history, that show how work gets done. This is newer, runs mostly through intermediaries, and is what this guide covers.

In the operational market, the intermediary is usually the counterparty. Mode, for example, says it "buys an agreed copy of company-owned operating records for AI research and training" and is "the direct buyer and payer" (Mode FAQ; application). micro1 says "companies retain ownership of their underlying data" (micro1). It is a data-license payment, not a loan or equity, which is why it gets discussed as non-dilutive capital.

Who is involved in a data licensing deal?

The agent question matters more than it looks. In the Spirit Airlines bankruptcy sale, Google's winning bid provided for a third-party deidentification process at the buyer's cost, and Mercor offered more money if it could use its own tools; the estate picked the third-party route (Stretto). Mode built its own tools after other vendors "would not let it audit how personal information was removed" (Business Wire).

What are the steps, and how long does each take?

The most detailed public timeline comes from Mode's launch release. Mode has since said the whole process takes about three months.

Sources: Business Wire, Mode FAQ, micro1, Mercor.

Step 1: Estimate

You supply headcount, systems and years of history. The output is a range, not an offer. Mode's Terms of Service (§3) say site figures "are not offers, quotes or commitments" and that any transaction happens "only under a definitive written agreement signed by both parties." For how estimates are built, see how much company data is worth.

Step 2: Nonbinding letter of intent

The LOI holds the estimate while the buyer inspects your data. "Nonbinding" cuts both ways: you can walk away, and the buyer can lower the price after review. Mode's release says the extension is available "at Mode's request," not yours. Use these two to three weeks to get counsel involved, inventory what is off-limits and tell the buyer early, because exclusions found late tend to reduce the price.

A well-drafted LOI, even a nonbinding one, should state at least:

• the estimated price or range, and what could move it
• the systems and date ranges the buyer will inspect
• how long the hold lasts, and who may extend it
• whether you may talk to other buyers during the hold
• confidentiality of anything the buyer sees during review

If the LOI is silent on exclusivity during the hold, ask. A short no-shop period is common in many deal types, but you should know whether you are agreeing to one.

Step 3: Final offer and definitive agreement

This is where the deal is actually made. Mode's FAQ says you approve "the value, terms, systems, and date ranges before any export," and that the "agreement defines its scope and any exclusivity." Every clause in the negotiation table below lives here.

The final price usually moves from the estimate once the buyer has seen the data. Mode says it reviews "the systems and dates in scope, years of history, record volume and quality, and ownership and access terms before making a fixed offer" (Mode FAQ). micro1 publishes tiers: $100K–500K for "qualified enterprise data partnerships," $500K–2M for "large-scale operational datasets or ongoing participation across multiple teams," and $2M-plus for "highly unique, proprietary operational data" (micro1). The practical implication is that exclusions cost money. A firm that removes client matters, HR files and finance systems is selling a smaller data set, and the offer should be expected to reflect that.

Step 4: Export and de-identification

Your team signs in to the approved systems; Mode says "no passwords are handed over." The buyer copies the data into a controlled workspace and removes or transforms identifiers. Mode's release puts this at 15 to 30 days.

Mode describes the company's side as "a few hours from your team." Use those hours well: keep a log of which accounts and date ranges were exported, confirm that excluded folders and channels were actually excluded, and ask whether you can see a sample of the de-identified output before it leaves the buyer's workspace. micro1 says companies "have the opportunity to review representative samples before use" (micro1); ask any buyer for the same.

Step 5: Payment

Payment follows whatever trigger the contract names. Note the differences in the table: micro1 pays after it accepts the data, Mercor ties payment to delivery of the anonymized dataset to the buyer, and Mode's FAQ says the amount and schedule are fixed in writing "before export."

What happens during de-identification?

Buyers want the workflow, not the people. Novaes described the goal as keeping "the operational decision-making of companies" while names change, "Dan's going to become… Derek" (The LFG Show). Mode's site says "personal identifiers such as names and contact details are removed or transformed" and shows customer and contact fields replaced with "Company 01" and "Person 01" (Mode).

That covers obvious identifiers. It does not settle three harder problems:

1. Re-identification. In the Spirit case, the consumer privacy ombudsman wrote that "there is no external way to verify if a private entity has privately deidentified 'anonymized' data," and 121 members of Congress warned that documents without personal identifiers "can potentially be reassembled" (Stretto; Travel Weekly).
2. Confidentiality is not identity. The flight attendants' union objected that the Spirit agreement's screen addressed whether records link to a consumer, not whether their contents are confidential. A client strategy memo with the names swapped out is still confidential.
3. Referential integrity. The Spirit agreement requires de-identification "while preserving referential integrity across the data set," which keeps the data useful but makes linking records easier. The pilots' union argued individuals would be identifiable as a result.

A useful benchmark: the Spirit sale agreement adopts California's deidentification standard, Cal. Civ. Code §1798.140(m), which requires reasonable technical measures, a public commitment not to re-identify, and contractual obligations on every recipient. It also makes the agent's certification a closing condition. Asking for equivalent terms is reasonable. See de-identified vs anonymized data for the definitions.

What terms should you negotiate?

Public content deals show how structure varies. Reddit disclosed that its 2024 licensing arrangements, worth $203.0 million in aggregate over two to three years, deliver content "through continuous access to our data API as well as quarterly transfers" (Reddit S-1). Informa's agreement with Microsoft paired an initial data access fee of "$10m+" with recurring payments over three years and "limits on verbatim text extracts" (Informa). Smaller operational deals will be simpler, but the same levers exist: one-time versus recurring, access versus transfer, and use limits.

What happens after you're paid?

The license outlives the payment. Three things continue:

1. Retention and deletion. Your contract should say how long the buyer and its recipients keep the copy. micro1 says data "is retained only for agreed periods" and that "deletion requests are honored according to the terms of the engagement." Get the equivalent in writing from whoever you sign with.
2. Refresh offers. A buyer may come back for newer records; Reddit's and Informa's deals both build in recurring deliveries or payments. Treat any refresh as a new deal with its own scope review rather than an automatic extension.
3. Internal communication. Employees may learn about the deal. In October 2026 coverage of companies selling workplace messages, Business Insider quoted an Emory law professor calling employee consent to resale "an ethically gray area." Decide in advance what you will tell staff, and when.

Keep the signed agreement, scope document, export log and payment records together. Ask your tax adviser how to treat the income; we don't give tax advice.

What should you check before signing?

Some records cannot be licensed on ordinary terms whatever the contract says:

• Law firms: client-confidential files and privileged material fall under duties such as ABA Model Rule 1.6. See our law firm page.
• Accounting and tax firms: tax-return information is restricted by IRC §7216. See our accounting page.
• Medical practices: protected health information needs a HIPAA de-identification method or must stay out.
• Employers with California staff: employee personal information has been covered by the CPRA since the exemption expired at the end of 2022 (CPPA). Plan employee notice and, where required, consent.
• Your own contracts: customer agreements, NDAs and SaaS terms can restrict reuse of records or claim ownership of service-generated data. In the Spirit case, a software vendor objected that its contract gave it ownership of data its platform generated (Stretto).

We are not a law firm or a CPA firm, and this is not legal or tax advice; talk to your counsel. For the full risk list, see is selling company data to AI safe.

We are a paid referral partner of Mode, which runs one of the programs described here, and earn a fee if a company we introduce licenses data through Mode; see our disclosure. The process and clauses above apply whichever buyer you choose, and comparing written offers from more than one is sensible if your agreements allow it.