What should be in an AI startup fundraising data room?
An AI startup fundraising data room is the indexed packet after the deck: corporate and cap table, financials, IP assignments that name model weights and embeddings, a training-data provenance list, foundation-model licenses, and the evals that support the claim.
Do not dump weights or customer training data into the open room.
AI startup fundraising data room (definition). The indexed evidence file behind the deck.
Provenance, licenses, evals, and IP that covers weights sit in folders an investor can reach in one or two clicks.
A teaser is not this room.
Weights and raw customer data stay out of the first-look pack.
AI startup fundraising data room
AI diligence is provenance, licenses, evals, and IP that names weights. It is not a SaaS deck plus a cap table with an extra folder. For the biotech cut, see What should be in a biotech fundraising data room?. For architecture, a BOM, or hardware proof, see What should be in a deep tech fundraising data room?.
Montague Law, writing 24 March 2026, is the outbound analog: an IP audit, not a vendor checklist. Their AI and data questions are training-data sources and rights, whether employees paste company material into outside AI tools, and whether customer and privacy disclosures match actual use. OSS is an SBOM, license compatibility, and copyleft. See Software Startup IP Audit Checklist for Seed and Series A. The structure is the packet. Drive can hold files. It is still not a room.
What goes in the folder
First-look is the open room. Confirmatory is still gated. Keep-out is what does not belong in a first look. Put inference vs R&D as separate lines only if you have them. Do not invent a GAAP lecture. If assignment contracts only say “software,” say that.
| Folder | First-look | Confirmatory | Keep out |
|---|---|---|---|
| Corporate + cap table | Yes | Yes | - |
| Financials | Burn, runway, model | Inference vs R&D as separate lines if you have them | - |
| Model card + evals | Task, metrics, eval set description | Harness, historical runs | Cherry-picked playground dumps as the only proof |
| Training-data provenance + licenses | Inventory: source, license, consent. No raw dumps | Deeper license memos | Raw datasets, customer transcripts |
| IP assignment covering weights / embeddings / fine-tunes | Yes, if the contracts actually say so | Full contractor chain | - |
| Foundation-model / OSS mapping | Provider + license | SBOM | - |
| Weights / checkpoints | No | Supervised access if asked | Open-room dump |
Three permission levels
Split early access from confirmatory diligence. Universal access to the whole tree is unnecessary.
| Level | Who | What they see |
|---|---|---|
| 1 Initial | First look, often pre-CDA | Deck, high-level model card, eval metrics, team. |
| 2 Serious investor | After CDA, real interest | Provenance inventory, IP summary covering weights, foundation-model licenses, corporate, financials. |
| 3 Confirmatory | Term-sheet path | Full IP, weights under supervised access if asked, harness and historical runs, deeper license memos. |
Weights stay out of the first-look pack
Do not dump weights or customer training data into the open room. Share a model card, an eval set description, a provenance inventory, and IP assignments that actually name weights. If an investor’s engineer needs the checkpoint, use supervised access, not a zip of the run.
Viewers get a view-only link, logged page by page. That is not DRM. Weights belong behind confirmatory access the way unpublished sequences stay out of a first-look biotech pack.
A data room can run a model on the files without training on them. See Does a virtual data room train AI on my documents?. Many Series A AI buyers will ask for a security questionnaire or a cert. Absence of one is not a Cashpian certification.
Teaser vs room vs banker VDR
DocSend and email win the deck. This room wins the packet. Datasite and Intralinks win once a banker is running the process. For the biotech cut of that split, see DocSend vs a data room for biotech.
| Job | Who wins | Why |
|---|---|---|
| Send the deck, see who read which slide | DocSend / email | One file. Not this room. |
| Assemble the packet, three access levels, keep it between raises | This room | Indexed folders. 1-2 clicks from the question to the file. |
| Banker-led process, later M&A | Datasite / Intralinks | A later tool. Not a Series A fundraising packet. |
See What is Cashpian? and pricing.
FAQ
Do investors expect the weights in the room?
No.
First-look is the model card, the evals, and IP that names weights.
Checkpoints stay out of the open room.
Supervised access is confirmatory, if they ask.
What is a training-data provenance list at seed?
An inventory: source, license, and consent for what trained or tested the model.
No raw dumps.
Montague asks whether contracts and privacy disclosures match those uses.
Deeper license memos wait for confirmatory.
Is this the same as a SaaS data room with an "AI" folder?
No.
A SaaS room can stop at deck, cap table, and financials.
An AI room adds provenance, foundation-model licenses, evals, and IP that covers weights.
It is not a SaaS packet with one extra folder.
Sources
- Montague Law, Software Startup IP Audit Checklist for Seed and Series A, 24 March 2026, retrieved 12 September 2026. AI and data questions: training-data sources and rights, employee use of external AI tools, customer and privacy disclosures aligned with actual use.
If the files are already in a folder, grade which row they are on. Cashpian is a gated room for that packet. It is not a listing portal. It does not sell the company or run the deal.