Signal

Vendors are training the AI inside their software on de-identified data from across their customer base, and opting out is starting to cost features. Atlassian, the Australian maker of Jira and Confluence, switched on de-identified and aggregated data contribution for its cloud customers in August 2026, to improve its apps and AI for everyone, and only Enterprise plans can switch it off entirely. Intercom trains its own support models on de-identified transcripts of its Fin agent's conversations across its customers, moved nearly all of those conversations onto its own Apex model in March 2026, and tells those who opt out that they may miss performance improvements. Freshworks calls its version Collaborative Models, trained on data from multiple customers, and two of its features stop working for anyone who leaves the pool. Japan has written the route into law, with a new exception that lets personal data go to third parties for statistics and AI training where the risk to individuals is low. And where rivals will not pool their data, they pool the learning: five pharma companies trained one model on their private molecular structures without a single structure leaving their buildings, and it beat anything one company could train alone. What your software learned from your data no longer stays with you. Whether your data joins the pool is becoming a line on the price list.

So what?

Your software now knows how the job is done from the data of everyone who used it before you. Keeping your own data out is becoming a price you pay, in features or in fees.

Evidence

The Register, 18 April 2026 (news): Atlassian told its 300,000 customers that from 17 August 2026 it would collect metadata and in-app content from Jira and Confluence to improve its apps and AI for all customers, de-identified and aggregated, and that customers on Free, Standard and Premium plans cannot switch the contribution off entirely.

Intercom, 8 September 2026 (policy): Intercom trains its Fin CX models on de-identified logs of its Fin agent's conversations with customers' users, offers an opt-out in workspace settings, and warns that customers who opt out may not benefit from certain performance improvements.

Apheris, 14 September 2026 (study): AbbVie, Astex, Bristol Myers Squibb, Johnson & Johnson and Takeda each fine-tuned a shared protein-ligand model on their own private structures through the Berlin company's federated platform, with only model parameters leaving each firm, and high-quality predictions rose from 35.6 to 52.1 percent on held-out structures.

Worth asking

If the software your organisation uses learned how you do your best work and shared that with every other organisation on the same software, what would you prefer to keep out of the pool?

Try what this looks like

Fictional concepts this signal echoes

Today's signals. Tomorrow's possibilities.

New science-fiction stories, emerging signals and essays exploring how life and work could change. Fresh perspectives and questions worth thinking about, delivered to your inbox.

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
Unsubscribe any time. Privacy Policy