The short version
A clean room is a place where two firms can join their data without seeing each other's rows. The sales pitch is about what stays in. But you pay for what comes out, and for what you can still do when the deal ends.
We read the public docs for fourteen ways that data leaves nine tools, as they stood on 9 October 2026. The docs show well how a result gets out. Some public rules limit how a result may be used. But the docs do not say what you may keep or use once the deal ends. We did not read any private contract. Read the full paper.
Start with the job
Before you sign, name the job you must keep doing after you leave. It might be a report you reopen, a model you run or a test you repeat. Then list what that job needs. A way to export is not enough on its own.
Four checks for each result
Take one result and ask four things.
- Get it. Does it reach you, or only a partner or an ad tool?
- Keep it. Can you still have it after the deal ends?
- Use it again. May you use it for a new job?
- Build it again. Could the same job run somewhere else?
The docs show the first. They show some rules for the second and third, but do not settle them. We did not check the fourth, because that needs a named job and a named place to run it. See the chart.
More leaves than you think
People say only totals leave a clean room. That is not quite right. Some tools let lists of matched people leave, under rules. One lets a trained model leave. Others send lists only into their own ad tools. Ask which thing goes to whom, and under which rule. See the table.
Small rules that bite
- Some results are kept for 30 days, or less if the room is deleted, unless you copy them.
- A model can leave while the data that taught it stays behind.
- Some results need the other side to take them in before they arrive.
- If you leave a shared room, you can lose sight of its logs.
Copy what you need in time. Save the logs before you leave. Get both rights in writing. Read more.
Proof has to cross the line
A log can show that a job ran. It may not show that the file got where it was going. Some tools can log that a file was deleted, but only that one file. Old copies can stay behind. Ask for proof at each step. See the chain.
Bots: check what they do
Firms are adding AI agents that can run these tools. A few general ones are fully out. The ones made for ads are still in test, in a closed trial, or say no stage at all. No stage does not mean not ready. It means ask. See the table.
Dates to watch
One old Databricks link for agents shuts on 31 October 2026. Snowflake plans to shut its old clean rooms in two steps in 2027. Publicis has agreed to buy LiveRamp, but on 9 October 2026 the deal had not closed. A firm can move a date it set. See the dates.
The case for keeping data in
Rules that keep data in can be good. They guard people's privacy and the other firm's rights. You may not need a copy of everything. A sound answer, a note of how it was made and the right to use what you learned can be enough. Check that before you pay for more. Read the case.
Tools and tests
The full paper gives a form for each result, a short list of what to take when you leave, a sheet to price the exit, and thirteen tests to run with fake data before you sign. We did not run them ourselves. See the tools.
We fixed our own pages
This site has guides on the same tools. When we checked them, we found errors, and we fixed them. A review then found that our first draft said more than the facts showed. We fixed that too, and list each change in the full paper. See the fixes.
Who wrote this
Evgeny Popov wrote this. He works at Samba TV, a TV data firm. He helps lead a group that writes one of the ad agent standards named here. He also runs the firm that writes this site's guides. Read the full note.