Data clean room: how it works

Two companies often know complementary things about the same customers, or the same market, but neither wants to hand over its records. A clean room settles that standoff. The data goes into a space neither party fully controls, the rules for what can be asked are set in advance, and only answers that pass those rules come out.

Where the idea came from

The idea grew up in advertising, where a brand and a publisher want to measure overlap between audiences. The IAB Tech Lab, the advertising industry's standards body, has published guidance and recommended practices setting out common principles, functions and privacy technologies for clean rooms. Large cloud providers now offer them as a service too.

1. 1

   Invite

   One party sets up a collaboration and invites the others.

2. 2

   Connect

   Each party links its tables. In AWS Clean Rooms, for example, the service reads data where it already sits rather than copying it to a partner.

3. 3

   Set rules

   The parties agree analysis rules: which columns can be joined, the minimum group size an output must have, and sometimes exact pre-approved queries.

4. 4

   Query

   The permitted party runs analyses. Outputs that break the rules are blocked.

5. 5

   Audit

   Logs record who ran what, so each side can check the agreement was kept.

Some products add further protection, such as differential privacy, which adds calibrated noise to results, or cryptographic computing that keeps data encrypted while it is processed.

A hypothetical scenario, for illustration only. A pharmacy chain holds sales records by store and product. A consumer health brand wants to license them, but first wants evidence that the data covers its products well. The chain will not release even a sample of store-level sales before a contract.

The broker proposes a short clean room trial. The chain loads its sales; the brand loads its product list. The agreed rule allows only counts and totals by region and month, with no output covering fewer than a set number of stores.

What each side learns

The brand learns how many of its products appear and how sales trend by region. The chain reveals nothing at store level. If the numbers look useful, the parties sign a licence and move to a normal delivery; if not, both walk away having exposed very little.

The broker's work in a trial like this is mostly in writing the rules: deciding which questions answer the buyer's doubt without giving the dataset away for nothing.

• Not automatically compliant. A clean room reduces exposure but does not by itself give a lawful basis to combine personal data.
• Rules can be gamed. Repeated narrow queries can sometimes reveal individuals unless outputs are restricted carefully.
• Both sides usually need technical staff, or at least someone who can prepare tables in the required format.
• Platform lock-in: some clean rooms work best when both parties use the same cloud or advertising platform.