Your AI Button Comes With a Supply Chain
Key takeaways
- A provider breach alone does not prove that customer documents were exposed.
- Promises about model training and data retention cover different things.
- SaaS companies should be able to trace customer data through their AI providers.
- Before enabling AI features, ask about data scope, retention, access, and incident response.
You click “Summarize” inside Notion. If an outside AI provider handles that request, the security questions extend beyond the app on your screen. The button is simple; accounting for everything behind it takes more work.
A breach leaves three questions to answer
Consider a hypothetical breach at Fireworks AI. That alone would not establish that Notion customer documents had leaked. You would first need to understand whether any relevant documents passed through the compromised environment.
An investigation needs to establish three things:
- Which systems did the attacker reach?
- Did those systems hold or process customer data?
- Did the attacker view or extract that data?
Skipping those steps turns a compromised internal account into a claim about stolen documents without the evidence connecting them.
The reverse shortcut is dangerous, too. Saying that a system stored no documents does not establish that data was safe while being processed.
The scope of the system breach and the scope of data exposure are separate findings. Two company names appearing together tells you much less than a clear account of what moved between them.
“No training” answers only one question
A promise that customer data will not be used to train AI models matters. It answers a specific question: will your documents help improve the model?
It does not tell you whether inputs are temporarily retained, whether error logs contain document excerpts, or who can access those logs. Each requires its own answer.
Imagine a meeting-summary feature designed to keep part of a request for troubleshooting. It could honor a no-training commitment while still retaining sensitive text. That is an illustration of how a service could work, not a claim about either company’s practices.
No training and no retention are different promises. Even when a provider retains nothing, access still needs to be controlled during processing.
A reassuring sentence in a data policy cannot answer every question about the data’s journey.
The app you chose should own the explanation
Customers choose the service whose name appears on the screen. They rarely select or manage every AI provider working behind it. A provider could change while the familiar summary button stays exactly where it was.
That is why the service operator should own the customer explanation and response.
A service such as Notion should know which features send data outside its systems, what those requests contain, and where they go. After an incident, it should be able to explain which customers and documents were affected—or clearly identify what remains uncertain.
“The incident happened at a supplier” may explain where the problem began. Customers still need to know whether their documents were involved, what action to take, and whether they can keep using the feature.
Those answers depend on preparation. The service and its provider should agree in advance on the records needed for an investigation, notification procedures, and response responsibilities. A carefully worded incident notice cannot compensate for missing evidence.
Three questions before you switch it on
AI procurement tends to start with output quality: does the summary capture the meeting, and does the answer make sense? Data handling deserves equally concrete questions.
- What gets sent? Just the selected paragraph, the whole document, or linked material as well?
- What stays, and who can see it? How long are inputs, outputs, and processing logs retained, and who has access?
- What happens after an incident? How are customers notified, and can transfers to the affected provider be stopped?
A public product announcement and an unpublished performance review might use the same summary feature. The consequences of exposure are very different. Understanding the data path helps a company decide which documents belong in that workflow.
Using an outside AI provider does not, by itself, settle whether a service is secure. A more useful measure is how well the operator understands its dependencies and how specifically it can account for customer data when something breaks.
One AI button can put several companies in the path of your document. The service you trust should be able to tell you where it went, when it is deleted, and who will contact you if something goes wrong.
Comments
Loading comments...