Security & data handling

What is sent out — and never used to train AI modelsNo model training

The findings themselves come from a deterministic rule engine, not a learning model. That part of the analysis makes no external call, so your documents cannot enter any model's training data through it.

To ground each finding in the most relevant passage, this deployment also uses semantic retrieval. When it is enabled — as it is here — passages from your documents are sent to OpenAI to compute embeddings (numeric vectors that let the system find the right passage), exactly as disclosed on the home page. Under OpenAI's API terms that content is used only to return the embedding and is not used to train their models. Semantic retrieval can be turned off (PI_EMBEDDINGS=off); the review then runs fully locally with no external calls at all.

So, plainly: your documents are never used to train any AI model, and they are processed only to produce your report. The only content that leaves the server is the document passages sent to OpenAI for embeddings while semantic retrieval is on — nothing else, and nothing for training. If that ever changes, it will be a disclosed change with its own data commitments — never a silent one.

Who sees your documents

Today, one named person — the founder running your engagement — under a confidentiality agreement. No anonymous team, no ad-hoc internal access, no third-party sub-processors handling your content. For the most sensitive work, the founder-run model means fewer hands on your data, not more.

Retention & deletion

Your documents are processed only to produce your report. You can have your documents and the report deleted on request, and the retention and deletion terms are written into your engagement agreement up front — so you know the lifecycle before you share anything.

Isolation — your data only ever produces your report

Each engagement is processed in isolation. Because the engine is deterministic and has no shared learning state between customers, one customer's documents cannot influence another customer's findings. The engine is the same for everyone; your inputs are yours alone.

Suitable for confidential projects

With this data-handling posture and a confidentiality agreement in place, the product is built for exactly this — a contained, independent read of board-sensitive work. For the most sensitive engagements, the contained founder-run model is an advantage over a multi-tenant cloud, not a compromise.

It's not a scraper, and it's independent

Project Intelligence reads only the documents you give it — nothing is scraped from you or about you. Any public reference material it draws on is licensed and obtained by permission. The review is independent of your delivery team, and it makes no benchmarked "X% of programmes" claims, because it uses no outcome corpus yet — only an honest, evidence-traceable read of what your documents do and don't contain.