Updated
For AI apps and agents
Sell AI Product Data to AI Companies
Short answer
Yes, some AI data programs ask for real usage data from AI apps, agents and developer tools. They ask for a product in production for 6+ months, complete session logs, and user terms that allow licensing de-identified data, or a way to get consent. Your lawyer should check your user terms and model provider terms first.
To sell AI product data to AI companies, you license real usage data from your app, agent or developer tool. The program may use a copy under that license. Your company keeps and owns the logs.
Two questions decide most of this. Are your session logs complete? Do your user terms let you do this at all? This page walks through program requests, poor fits and what your model provider's terms may say.
We are a data agency. Our job is to match companies with the AI data programs that fit them. A referral fee from the program pays us, so your company gets no bill from DillisData.
- Real users only
- Full sessions matter
- No bill from DillisData
What programs ask for
What do AI companies look for in AI product companies?
No two programs use the same rules, and rules change, so use this as a guide. A session is one user's full run with your product, from the first prompt to the end. A tool call is when the model uses an outside function, such as a search or a file edit.
| What some programs ask for | Why they want it |
|---|---|
| A product in production for 6+ months with real users | Real users ask messy, surprising things. A demo or a test script doesn't. |
| Complete session logs: prompts, responses and every tool call | The full path shows how an agent got to a result, including wrong turns. Value depends on session volume and completeness. |
| User terms that allow licensing de-identified usage data, or a way to get consent | Programs want this settled up front. De-identified means details that point to a person are removed. |
| Usage-data programs: roughly 5 to 200 people on the team | Small teams can fit here, since the ask centers on the logs. |
| Whole-company programs: around 50+ people, counting desk jobs alone | These programs study a team at work. A further group begins near 20 or 30 full-time staff. |
| Whole-company programs also want 5+ tools in daily use and 3+ years of history | Young AI companies may fall short here. The usage data path is the likelier fit. |
Who is not a fit
Which AI products are not a fit?
If your logs fall into one of these groups, some programs will say no. It helps to find out early.
Synthetic or scripted traffic
Sessions from test bots, evals or scripts are not what these programs ask for. Filter them out before you count your volume.
Logs that can't be stripped of personal information
Users paste names, health details and passwords into prompts. If you have no way to remove that, some programs will turn the logs down.
Truncated or partial traces
Some teams cut long prompts to save storage, or keep only a slice of traffic. Logs like those don't show the full session. Check your settings before you take the quiz.
Products with no real usage yet
A product still in beta, or live for under 6 months, misses the ask above. Come back once your production history is longer.
Defense and export-controlled work
Some programs won't take these records at all. If your product serves that field, keep those sessions out or expect fewer matches.
Your records
| Records | What they show | What to check |
|---|---|---|
| Coding sessions saved with the original repo or environment | What the agent did and what the code looked like when it did it | Whose code it is. It may belong to your customer. |
| Sessions with planning, to-do and clarifying-question steps | How an agent breaks up a task and asks for what it needs | That those steps are in the log, not hidden from it |
| Sessions that search the web beyond documentation | How an agent looks for facts and judges what it finds | That results and pages opened were kept |
| User ratings or preference votes | What real people thought of an answer | That each vote links back to the right session |
| Sessions run on open-weight models | Real usage on models whose weights are public | The license for each model, since they differ |
Rules to clear
Do your user terms and model provider terms allow this?
This is the real work for an AI product company. Take this as general information and not legal advice. Your lawyer should read your terms and your provider's terms side by side.
What you promised your users
Read your terms and privacy policy as they stood when each session happened. An FTC staff post from January 2024 says model-as-a-service firms must keep their promises to users and customers. That holds wherever the promise appears.
Changing your terms now
Quietly editing your terms is not a safe shortcut. A February 2024 FTC staff post warns about quiet, retroactive changes to allow AI training. It says they may be deceptive or unfair. Ask your lawyer how to handle sessions from before any change.
Getting consent
If your terms fall short, consent may be the path. Some teams could ask users directly, then license only sessions from people who said yes. Your lawyer should write the wording, and it should be plain enough to read.
Model provider terms
Your model provider may limit how outputs are used. One large provider's services agreement bars using output to build models that compete with it, with narrow exceptions. Another's commercial terms bar training competing AI models. A third provider's API terms have a similar limit. Whether a license reaches those limits is a question for your lawyer.
Business customers
Enterprise contracts and DPAs often limit how you may use customer content. If you promised a customer you would not train on its data, keep those sessions out. Sort logs by customer before anything else.
Personal details in prompts
Tools can help find them. Presidio is an open-source tool for finding and masking personal data. Its own docs warn there is no guarantee it will find everything. Scrubbing helps, yet no method is perfect.
Other people's content
Prompts can hold pasted contracts, private code and customer lists. Your users may not have had the right to share those with anyone else. Filter where you can, and ask your lawyer about the rest.
Run the numbers before you read the fine print
A few one-tap questions cover your team, your records and your region. The estimate shows up right there. No logs and no API keys are involved.
The process
How does an AI product company reach a license?
We do the matching and stop there. No logs ever pass through us. You sign directly with the program you choose, and the transfer stays between you and it.
Do the quiz
Give 8 to 14 one-tap answers on team size, records and region. You share no prompts and connect no tools.
2 minutesLook over your top 3
An estimate appears, plus your top 3 program fits and the reason for each. We flag any terms that call for a careful read.
Audit your logs and terms
Check that sessions are complete. Your lawyer reads your terms by date range. You pick what goes in, and what stays out goes into the contract.
Sign and transfer
Join your pick in one click, or wait for its representative to contact you. The program copies the agreed sessions, strips personal details and pays your company.
Public facts
What public reports show about usage and work data
We played no part in these deals. None of them involves an AI product company, and none tells you what your logs would bring.
Reddit's data licensing contracts, signed in January 2024, as reported in its IPO filing. What people typed online became something AI firms paid to use.
SEC Form S-1, Feb 2024How many deals at one startup shutdown firm now include a code or data sale, according to that firm.
MIT Technology Review, Oct 2026Planned 2026 capital spending at four tech giants: Alphabet, Amazon, Microsoft and Meta. Most of it is for AI, largely chips and data centers rather than data.
Sherwood News, Apr 2026Timing
Why should an AI startup look at this soon?
Current prices may not hold, and no one knows for how long. These points apply with extra force to young AI companies.
Logging choices can cost you
If your tool deletes old traces or cuts long prompts, sessions disappear each day. Fix that now, even if you never license a thing.
Real agent sessions are scarce
Records a program can't get elsewhere tend to pay best. Full traces with real users and every tool call are not on the open web.
Rivals log the same kinds of tasks
Once a program holds sessions close to yours, it tends to need yours less. Coding agents and support bots are crowded fields.
Prices may move
Expect pay to shift as more companies join. Labs keep trying other ways to train, though nobody can say when that will bite.
What might your sessions be worth?
Programs typically pay out $100K to $5M+, largely depending on a company's size. That is a typical range and no promise. A small team should expect its estimate to reflect that. Check yours and decide from there.
Questions
Questions owners ask
Our product has been live for 4 months. Is that too soon?
It is short of what some programs ask for, which is 6+ months in production with real users. Keep your logs complete in the meantime. Then come back, since your history only grows from here. Terms and rules may also change by then.
We never stored tool calls. Can we still qualify?
Some programs ask for prompts, responses and every tool call, and they list partial traces as not a fit. Older logs without tool calls may fall short. You can turn on full logging today, if your terms allow it, and build a complete set going forward.
Can we change our terms and then license old sessions?
Be careful here. FTC staff warned that quiet, retroactive changes for AI training can be deceptive or unfair. Sessions from before a change may need consent, or may need to stay out. Your lawyer should map which sessions fall under which terms.
Does our model provider's contract block this?
It might. Several major providers limit using outputs to build competing models. Those limits differ, and the answer can turn on who trains what. Ask your lawyer to read your contract, not just the public terms. An enterprise deal may say something different.
Do sessions on open-weight models avoid that problem?
They may reduce it, but read the license for each model. Open-weight licenses differ a lot. Some programs say sessions on these models can add value. Your duties to your own users stay the same, whichever model ran the session.
We only sell to businesses. Is that easier or harder?
It can be harder. Business customers often sign contracts that limit how you use their content, and some bar training outright. You may need each customer's consent. Free or self-serve users on your standard terms may be a simpler place to begin the review.
Do we need a big team?
Not always. Some programs built around usage data look at companies of about 5 to 200 people. Programs that want whole-company records ask for more staff and 3+ years of records, which many young companies won't have. The quiz checks this for you.
Does DillisData see our prompts or logs?
No. We look only at your quiz answers and the rules each program sets. Your logs can't reach us, because they go straight to the program if you sign. For a wider view, read how businesses sell data to AI companies.
Next step
See what your production logs could earn
Answer the quiz in about 2 minutes. No prompts, no logs and no keys change hands. Then read your terms with your lawyer and decide.
Nothing owed to DillisData. No logs needed. An estimate, never an offer.