How we use AI

Founderboard is an AI product: the board members, the tools, the meeting write-ups and the reports are all generated. This page is the technical account of how that works, written for the people who have to sign off on it — IT, security and compliance reviewers at the companies and accelerator programs whose founders use the platform.

It is a companion to the Privacy Policy, which governs, and the sub-processor list, which is the authoritative list of who processes what.

1. The short answers

Question	Answer
Do you train models on our data?	No. We train, fine-tune and distil nothing.
Do your providers train on our data?	No. We use paid API tiers under terms that do not permit it.
Do you run your own models?	No. Every model is a third-party API. There are no model weights on our infrastructure.
Do you build embeddings or a vector index of our content?	No. There is no vector store in the product.
Do you keep copies of prompts and responses in an AI logging system?	No. We log token counts, not content.
Can a human at Founderboard read our chats?	Only through the database, under the access controls in the Privacy Policy. There is no review queue and no sampling of conversations.
Can our accelerator's coaches read our chats?	No. Chat content is excluded from every coach surface and every generated report, by design.
Does AI take decisions about people?	No. Every output is advisory. No automated decision with legal or similarly significant effects is made (GDPR Art. 22 does not apply).
Is anything sent on our behalf?	No. Outreach mail is sent by the founder, from the founder's own mailbox.
Where does data go?	Google (Gemini API) for text, vision and images; ElevenLabs for speech. Both may process outside the EEA under Standard Contractual Clauses.

2. Which models, and how they are chosen

All text generation runs on the Google Gemini API through Google's official SDK. Audio briefings additionally use ElevenLabs for speech synthesis, and the Venture Video tool uses fal.ai to run ByteDance's Seedance (the 30-second film) and Seedream (storyboard stills). There is no other model provider, no self-hosted model, and no open-weights model running on our infrastructure.

A film is made from text alone. What fal.ai receives is the script as it stands on the board — the founder's choices, the cast descriptions and the beats — composed into one prompt; no photo of a person is ever sent as a reference, and the prompt forbids real people and real brands. A venture may make two films, then waits a week, so a founder presses the button that sends it, every time.

Models are addressed in four roles rather than by name:

Role	Used for
Quality	Board member chat, tool work, board meetings, pitch decks, coach reads and reports, grounded web research
Fast	Titles, tagging, commentary, summarisation and compaction, transcription, short notes
Image	Venture visuals, brand logo marks, generated illustrations
Embedding	Configured, currently unused — the product builds no embeddings

Model identifiers are configuration, not code. They are read from environment variables at runtime, so a model can be changed, rolled back or pinned per environment without a release. Every call site routes through a single routing table that maps a named feature — boardMemberChat, coachReport, prospectPeopleSearch — to one of those four roles. That table is the complete list of places the product calls a model, which is what makes an inventory of AI uses possible at all.

Current model identifiers for a specific environment are available on request under NDA. We do not publish them, because they change.

3. Training — what is and is not done with your data

We do not train. Founderboard performs no training, fine-tuning, distillation, reinforcement learning, adapter training or any other procedure that turns customer content into model weights. We hold no model weights of our own. Nothing you write can end up inside a model we ship.

Our providers do not train on it either. Content sent to the Gemini API is sent on a paid tier, under terms that do not permit your prompts or the responses to be used to train or improve Google's models. The same holds for text sent to ElevenLabs for synthesis. Providers may retain content briefly for abuse monitoring and service operation under their own terms; that retention is theirs, and we set no instruction that extends it.

No retrieval index is built. The product creates no embeddings of your content and maintains no vector database. Where a board member "remembers" your venture, that memory is a written summary in your own project's row in our database — plain text, readable, exportable and deletable — not a representation in a vector space. It is regenerated from your current content, never accumulated across customers.

Nothing is shared between ventures. Context for a call is assembled from one project, scoped to the requesting member. Your content is never used as context, example or training signal for anyone else's board.

4. What is sent to a model, and when

A model call happens when you do something that needs one — send a chat message, ask a tool to work, generate a deck — and on a fixed schedule for a small number of background jobs (the weekly board meeting, the daily news crawl, the weekly aura recalculation, re-engagement checks). There is no continuous background processing of your content.

What travels with a call depends on the feature:

Feature	What is sent
Board member chat	The board member's persona, your venture profile, a summary of completed tool content, the recent history of that thread, recent news headlines, and your message
Asking the board on our homepage, without an account	What you type there and the conversation so far, which your browser keeps and sends with each message. We store none of it; only the number of tokens used is logged, linked to nobody. Your IP address is held in memory for an hour to limit how many messages one address can send. If you leave your name and email in the form the board may offer, they go to our team by email and never to the model
Tool chat and tool fill	The tool's prompt, the tool's current content, your venture profile, and a short memory block summarising recent board conversations
Board meetings, venture narrative, reports	Venture profile, tool content, progress and activity signals — never chat content for coach-facing reports
Dropboard capture	The note, link or file you captured, for titling and tagging
Transcription	The audio clip you recorded, returned as text into the composer — you send it, or you don't
Recording a meeting in the app	The audio, in pieces of under a minute, transcribed as above; the transcript is saved as that meeting's notes. Audio is kept on your phone only until its piece is written down, and never stored by us. When the meeting has a prep brief, every minute or so the brief and the recent transcript are sent to a model that returns which of your questions have been covered and at most two short suggestions for what to ask next; the suggestions are shown on your phone and not stored. You are responsible for telling the people in the meeting that you are recording, and the app reminds you
WhatsApp	The message you texted, and any attachment, which is filed in the Dropboard
Images	A text art direction and the brand palette — never a photo of a person unless you uploaded one to be processed
Adding a contact from a paste or a photo	The text you pasted — a signature, a profile — or the photo of a business card you chose, read back into the form's fields for you to check. The photo is not stored, and nothing is saved until you add the person
Looking up a contact	The name, company and role you entered for one person, executed as a live Google Search only when you ask for it. The result is offered for you to confirm. An email address is offered only when it is printed on one of the pages the search found — which our server then opens to check — and always with a link to that page; the model is never asked for one
Surveys	Nothing is sent to a model while somebody answers. Afterwards their answers are part of what the venture learned: open answers are read word for word, like an interview's notes — in the board member's chat about that survey and about what the venture learned, and when a founder asks for it to be read again. Answers to closed questions travel only as counts per question, never with who gave them. A founder may have the survey page put into another language: the page's own words and the founder's questions and texts are sent to a model to translate, with nothing about anybody answering. On a translated survey, open answers are sent word for word to be put into the founder's language when a founder reads them — the translation is kept beside the original, which stays what counts
Grounded research (Prospecting, news)	Your search brief and the company or topic being researched, executed as a live Google Search
Board member lookups	A single question the board member writes when your conversation needs a current fact — a company, a price, a regulation — executed as a live Google Search. The question is drawn from the chat, so it may contain names or details you mentioned; your messages themselves are not sent to the search
Speech synthesis	The script text of the briefing, sent to ElevenLabs

What is never sent to a model: passwords or password hashes, session tokens, refresh tokens, API keys, another member's private chat, another venture's content, or the contents of a venture you are not a member of. Access is checked before the call, not after it.

5. Where it goes, and under what terms

Provider	Role	Region	Transfer basis
Google (Gemini API)	Text, vision, image generation, grounded search	US / global	Standard Contractual Clauses; EU-US Data Privacy Framework where applicable
ElevenLabs	Speech synthesis	US	Standard Contractual Clauses
fal.ai	Video and storyboard images (Seedance, Seedream)	US	Standard Contractual Clauses

All three are on the sub-processor list with a data processing agreement in place. Our application, database and file storage are hosted in the EU; the model calls are the part of the stack that leaves it, which is why they are listed separately here.

The Gemini API key is held server-side only. The browser and the mobile app never hold a model credential and never call a model provider directly — every call goes out from our backend, behind an authenticated route scoped to your project membership.

6. What we log about AI calls

For every model call we write one accounting row: the feature name, the model identifier, prompt and completion token counts, a timestamp, and the account it belongs to (user, venture, company).

We do not log prompts or responses. There is no separate AI observability store holding copies of your conversations. The chat you can see in the product is the chat that exists; deleting a project deletes it.

Accounting rows are kept for cost attribution and capacity planning. Erasing an account detaches its rows from the person — the spend stays billable to the company, the row no longer names anyone.

7. Human oversight, and the limits of the output

Every output is advisory. A board member proposes; the founder accepts. Nothing a model produces is applied to your venture without a person putting it there, and the record of who wrote what inside a tool is kept per row, including when a proposal was taken up and by whom.

Nothing is sent on your behalf. Founderboard has no outbound channel to anyone you are contacting. Outreach messages are written in the tool and sent by the founder from their own mailbox; replies come back because the founder CC'd a campaign address.

Figures are computed, not generated. Where a slide or a report carries a number — market sizing, unit economics, demand counts, progress and completion — that number is calculated from your own stored content in our code, and the model is asked only for the words around it. A slide with no underlying work behind it is not sent to the model at all; it renders as a gap naming the tool that would fill it. This is the mechanism that stops an empty tool becoming an invented market size.

Grounded research must actually search. A research or prospecting turn that came back without having run a web search is retried once, then refused with an honest "the search did not run" rather than an answer recalled from training data. Sources are returned with the answer.

Reports about founders are a read, not a verdict. Accelerator-facing reports are synthesised from tool content, progress and activity signals, are always presented to a human coach, and carry no scoring of a person. They must not be used as the sole basis for admitting, ranking or removing anyone from a program. They are built without access to chat content.

AI output can be wrong. It is not legal, financial, tax or professional advice. The product tells you this at the point of use.

8. Transparency

• Chat surfaces carry an explicit notice that you are talking to an AI and that it can be wrong.
• Board members are clearly presented as AI personas. They are not, and are never presented as, real named advisers.
• Generated images and synthesised audio are produced within features that are described as generative; we do not present generated media as photography or as a recording of a person.
• Grounded answers carry their sources.
• Virtual customers are role-play personas, labelled as such. Conversations with them are kept out of the memory that feeds ordinary board chat, so a role-play is never mistaken for something you said.

9. Third-party personal data

The Prospecting and outreach features find real people at real companies, which means the product processes personal data about individuals who are not our users.

• Company and people discovery runs as a live web search. Hunter.io is the only contact-data provider, used in Prospecting for published work email addresses, and only when a founder asks for one. There is no scraping of LinkedIn, no crawling of sites, and no bulk contact database: the only web pages our servers open are the few that one founder's lookup of one named person returned (below), read once and not stored.
• A founder adding somebody can paste a signature or profile, or photograph a business card, to fill in the form. The text or photo is read once by a model and returned to the form; the photo is not stored. An address or phone number the model returns that is not in the pasted text — or, for a photo, in its own reading of the card — is dropped rather than shown.
• When adding somebody, a founder can ask for a live web search on that one named person — to confirm their current role and company. It returns candidates with their sources for the founder to choose from. A work email address is included only when it is published on one of those pages and reads as that person's; it is never generated, it is always shown with the page it came from, and that page is recorded on the person's thread. No phone number comes from a search.
• A venture can send a survey to people it chose, each through a personal link. Answers are linked to the person who gave them, are never changed by a founder, and can be withdrawn and deleted by the person from their own link at any time. The page carries no analytics or tracking scripts. A page a founder had translated says, with its notice, that the notice is a translation and that the English text applies, and shows the English one tap away.
• Spend is capped per venture per month, which also caps volume.
• The results live inside that venture's own tool content: covered by the data export, deleted with the project, and subject to the same access controls as everything else.
• Mail a founder CC's or forwards to a campaign address is stored against that campaign so replies can be tracked. It is the prospect's own words, held under the same retention rules.
• The founder decides who to look up and what to send them, and is responsible for the lawfulness of their own outreach under GDPR and ePrivacy.

10. Regulatory position

GDPR. Legal bases are set out in section 3 of the Privacy Policy. AI generation for features you use is performance of a contract. No solely automated decision producing legal or similarly significant effects is taken, so Art. 22 does not apply. Data subject rights — access, export, rectification, erasure — reach AI-generated content the same way they reach anything else, because it is stored in the same database and covered by the same export and erasure paths.

EU AI Act. Founderboard is a deployer of general-purpose AI models and the provider of the AI system built on them. Its use — generating business content and advice for the person who asked for it — does not fall within the Annex III high-risk categories. The product performs no emotion recognition, biometric identification or categorisation, social scoring, predictive policing, or automated decision-making on employment, creditworthiness, insurance, essential services or education admission. Where a board member is shown as concerned or excited, that is the persona's own written state, derived from your venture's activity — nothing infers anything about a person from their face, voice or any other biometric signal. Our obligations are the Art. 50 transparency ones — telling people they are interacting with an AI system and not passing off generated content as human or authentic — which section 8 describes. Where a program uses Founderboard's reports as an input to a selection decision, the program is the deployer of that decision and the human judgement in it is theirs; our reports are advisory and carry no ranking of individuals.

Confidentiality. Your venture content is yours. We claim no licence to use it beyond running the Service for you, as the Terms set out.

11. Your controls

• See it — everything the product holds about you, generated content included, is in the data export on your profile page.
• Remove it — deleting a project removes its content, including its AI output; deleting your account erases your identifiers and detaches the accounting rows.
• Limit it — the simplest control over what reaches a model is what you put into the product. Please do not put special categories of personal data into the tools or the chat.
• Turn it off — you cannot use Founderboard without AI; it is what the product is. If that is not acceptable for a given class of information, keep that information out of it.

12. Changes, and asking us things

We will update this page when the technical picture changes — a new provider, a new class of model call, a change in what is sent. Material changes are also reflected on the sub-processor list before they take effect.

Security questionnaires, DPIA input, architecture detail under NDA, or anything else a review needs: email privacy@founderboard.ai.