In short. The startups you support compete on a global market, often against American competitors, and yet they hand their most sensitive information (numbers, customers, technology, fundraising strategy) to American tools and their subprocessors: workspace, AI assistant, notetaker. Before choosing or tolerating a tool in a program, an incubator or accelerator should know what it captures, where the data goes and to whom, how long it is kept, whether it trains models, who can see it, how participants are told, and who handles an incident. Above all, it should ask its founders whether they agree. The Notion/Fireworks incident of November 2026 shows that the question goes beyond the main vendor: Notion AI conversations were exposed at an AI vendor almost nobody had heard of.
The question to ask your founders
In innovation, competition is global. Many of the startups you support face, or will face, American competitors that are often better funded. And during the program, they share with their coaches what gives them an edge: their numbers, their first customers, a technology not yet protected, their fundraising strategy.
A good share of this information goes through American tools: the workspace (Notion), the AI assistant (Claude, ChatGPT), the assistant built into the video call, and above all the notetaker (Otter, Fireflies) that a coach or mentor connected to their calendar. Each of them relies in turn on subprocessors the founders don't know about.
These tools are genuinely useful, and banning them would not work. One simple question remains: do your founders agree to their sensitive data leaving for the United States, and has anyone asked them? If nobody asks, the program makes the choice for them.
Questions to ask before allowing a tool
For every tool used in sessions or to store founders' work (Notion AI, a notetaker, the video call assistant, ChatGPT…), ask the vendor in writing:
- What does the tool capture? Audio, video, shared screen, call chat, documents, calendar, contacts.
- Where does the data go, and to whom? The list of subprocessors (AI model provider, host), their country, and any transfers outside Europe.
- How long is it kept, by the vendor and by its subprocessors? And how is it deleted?
- Is it used to train models? Is that forbidden by contract, including for subprocessors?
- Who can see it? Roles, access, access logs, strong authentication.
- How are participants told? And can they refuse the recording without standing out?
- What happens in an incident? How fast the vendor notifies you, who tells the founders, how access is revoked.
Simple rules for a program
- List the allowed tools when founders join the program, and say which ones use AI.
- Ask every participant for consent before recording a session, at every session, not just at sign-up.
- Turn the bot off for sensitive sessions: fundraising prep, co-founder conflict, intellectual property, customer data.
- Never paste passwords or API keys into an AI assistant (the Notion incident showed why).
- Name who handles an incident: who receives the vendor's email, who tells the founders, how fast.
- Have session notes reviewed by the team and the coach, instead of letting the bot's summary circulate as is.
An example: the Notion/Fireworks incident
Notion published an advisory on November 10, 2026, updated the next day. Here is the timeline as it gives it:
| Date | What happened |
|---|---|
| October 23 | An attacker gets into Fireworks' development environment through a third-party cloud provider (confirmed by Fireworks, according to BleepingComputer). |
| November 2, evening | Fireworks tells Notion that a small dataset was accessed. It reportedly contained only test data. |
| November 3, morning | Notion disconnects its systems from Fireworks, has the data on the vendor's side deleted and stops all communication with it. |
| November 6 | Fireworks warns that customer data may have been accessed. |
| November 8 | Fireworks confirms the affected dataset. Notion finds Notion AI chat transcripts used for the Insights feature (which shows admins how their team uses AI). |
| November 9 | Notion contacts affected customers. |
| November 11 | Notion revokes the Notion API keys found in affected users' messages. |
The incident affects fewer than 1% of Notion workspaces. For the vast majority of them, the transcripts were not tied to a user or a workspace. Their content is still what people wrote to the AI, though, and Notion recommends "treating this data as exposed" and rotating any password, API key or credential pasted into Notion AI during the period.
Data from Notion AI Meeting Notes (audio, transcripts, summaries) was not affected, unless it was used as input in a chat with Notion AI.
To check your workspace: if you are a workspace owner, open Notion on desktop, then sidebar, "?", "Security incident review". Notion says that if neither you nor your workspace owner has heard from them, your workspace has not been identified as affected.
What the incident teaches startup support programs
A contract is not enough
The most useful passage in the advisory is this one: Fireworks had committed by contract to retain no Notion data, and during the investigation Notion learned that data had been retained. Notion is a mature vendor, with annual audits (SOC 2 Type II, ISO 27001) and security reviews of its AI subprocessors. The commitment was there, and it was not kept.
For a program, this means a well-written DPA (data processing agreement) is necessary, but it does not tell the whole story. You also need to know how many links there are between the session and the server that processes it.
The chain is longer than you think
This incident has four links: the user, Notion, Fireworks (which provided inference to Notion), and Fireworks' cloud provider, through which the attacker got in. No user had chosen Fireworks, and very few knew it existed.
Yet in a support program, founders put things into these tools that they say nowhere else: the customer about to leave, the disagreement between co-founders, the price they don't dare announce yet, the patent not filed yet, the funding round that won't happen.
Notetakers follow the same path (plus consent)
A notetaker joins the call, records, transcribes and summarizes. The transcript goes to the vendor, which usually relies on an AI model provider and a cloud host of its own. It is the same chain as at Notion, applied to the coaching session itself.
Then there is consent. Very often, the bot is there because a single participant connected it to their calendar. In the United States, a class action was filed in August 2025 against Otter.ai in the federal court for the Northern District of California (Brewer v. Otter.ai). The complaint claims in particular that if the meeting host has connected their calendar, the notetaker "may join the meeting without obtaining the affirmative consent from any meeting participant", and that recordings are used to train the vendor's models. The case has not been decided yet, but it describes well what happens in a support session.
For a founder, the bot is often just one more name in the participant list. They don't know who invited it, or where what they say ends up.
What the French and European framework says
No public guidance specifically covers notetakers in startup support programs. Existing texts still give a fairly clear framework, and French regulators have spelled it out.
- GDPR and processors. For the CNIL (the French data protection authority), a processor is the organization that processes data "on behalf of another organization (the controller)", and that processing must be governed by a contract. When a program imposes or provides a tool to its founders, the tool's vendor is its processor, and the vendor's own subprocessors (like Fireworks for Notion) are part of the chain. (CNIL, working with a processor, in French)
- Transfers outside Europe. When data leaves the European Economic Area on the basis of standard contractual clauses, the exporter must carry out a transfer impact assessment (TIA). The CNIL published the final version of its guide on July 9, 2025. (CNIL, TIA guide, in French)
- The ANSSI and French Tech guide for startups. Published in February 2026 by the French national cybersecurity agency and the French Tech Mission, it recommends not sending sensitive information (API keys, business data) to AI models, especially SaaS ones, and setting a framework for AI use. It also names incubators and accelerators as well placed to offer these secure services. (Cybersecurity guide for digital startups, in French)
- The DINUM guide on AI use, written for French civil servants and useful for public programs. It points out that the location of servers is not enough, recommends for sensitive uses solutions "ideally SecNumCloud-qualified by ANSSI", and asks people to tell the people they work with when they use AI. (DINUM guide, in French)
How La Mashine handles it
We recorded and transcribed our own coaching sessions for months, and that is partly what allowed us to build La Mashine. So we know the comfort it brings, and what it asks in return.
In La Mashine, a session is recorded only if every participant agrees. The session notes are reviewed and signed by a team member and by the coach, then the audio is deleted. Every access to a project is logged, and emails and phone numbers are masked before being sent to the AI. And we offer every program a 100% European setup by default: data is hosted in France, the AI runs in Europe, and nothing goes through the United States, including for video calls held in La Mashine and their transcription. It is already live with one of our customers.
If you want to look at your program's tools together, book a call.
Frequently asked questions
Should founders be asked before using an American tool?
It is good practice. Tell them which tools the program uses and which ones send data outside Europe, so they can choose what they put in them. For recording a session, ask every participant for consent, at every session.
Was Notion hacked in November 2026?
No. Notion says its own systems were not breached. The breach happened at Fireworks, one of its AI vendors, and exposed Notion AI chat transcripts linked to the Insights feature, in fewer than 1% of Notion workspaces.
How do I know if my Notion workspace was affected by the Fireworks incident?
If you are the workspace owner, open Notion on desktop, then sidebar, "?", "Security incident review". If neither you nor your workspace owner has received a message from Notion, your workspace has not been identified as affected.
Can an incubator use an AI notetaker during coaching sessions?
Yes, provided participants are told, their consent is obtained before recording, the vendor is bound by a data processing agreement, and you know where the data goes, including to its subprocessors. For the most sensitive sessions (fundraising, co-founders, intellectual property), it is better to do without.
Is a tool hosted in Europe enough to protect founders' data?
Not always. The data can be hosted in Europe while the AI processing it runs elsewhere, at a subprocessor. The French government's DINUM guide puts it simply: the location of servers is not enough. Ask where the AI models run and who the subprocessors are.
Who is responsible if a notetaker leaks a session?
It depends on who chose the tool and for what. If the program imposes or provides the tool, it is generally the controller and the vendor is its processor. If a participant connected their own notetaker, they are the one who brought the tool into the session. Either way, the best protection is a list of allowed tools and a clear rule on recording.
