“Is the AI going to have
access to my stuff?”
It's the first question every owner asks, and it deserves a structural answer — not a reassuring paragraph. This page is the whole stack: how Odell touches your knowledge, what the AI providers do and don't get, what we log, and where the honest limits are.
“Odell is like a sharp employee with keys to your filing cabinets. Somebody asks a question, Odell opens only the cabinets that person is allowed to open, reads what's needed, answers — and that's the end of it.”
Permissions come before reading
Most systems retrieve first and apologize later. Odell filters first: an employee's question can only draw from cabinets that employee is cleared to open. Pricing a tech isn't cleared for doesn't get summarized, hinted at, or leaked through a clever question — it simply isn't in the room. The owner sees everything; there are no locked drawers against the owner.
The engine is rented. It doesn't keep or learn.
The AI models that phrase Odell's answers are called under paid, business terms that contractually exclude training on your data, with zero-retention settings wherever the provider offers them. Model families whose terms force retention near customer data are excluded from the roster entirely. Your interviews, documents, and questions are never training material.
Conversations are ephemeral
An Ask session doesn't become content. When a session ends, Odell forgets the conversation and re-retrieves from the cabinets next time. What persists is what should: the cabinets themselves — that's the product — and the audit record that a question was asked, by whom, and from which cabinets it was answered.
Every look is logged — and you can read the log
Odell's audit log is append-only: who asked, what was served, from which cabinet; who opened or downloaded a document; who changed a permission; who approved an entry. Compare that to right now, where your prices live in a binder anybody can open and nobody logs a thing. Your secrets are safer in Odell than in the break room — the break room doesn't keep a log.
Cabinets are encrypted, and there's a deepest drawer
Cabinets and documents are stored encrypted and access-logged. Entries can also be marked owner-only at the deepest level: Odell will serve them only in the owner's own session, and will not recite them to anyone else. Some formulas are for the owner's eyes only; the system treats them that way by design.
You hold the keys
You decide what goes in, who can open what, and who approves changes. On the Dedicated lane, the whole brain lives on a server you own, under accounts you control — revoke our management key and the box is simply yours. And on any lane, if you ever leave, your brain goes with you. It's yours.
What we will never promise you.
Nobody honest can promise “never, ever, any risk” — not us, not your bank, not the safe in your office. If a vendor guarantees zero risk, they're selling, not engineering.
Ephemeral conversations do not remove storage risk: the cabinets have to live somewhere. That's why the answer is a stack — no-training contracts, zero retention where offered, encryption, permission-before-retrieval, total audit visibility, and a dedicated-box option — and why every layer of it is something you can check.
One more, specific to Dedicated: putting the brain on your server changes where knowledge lives. The few entries needed to answer a question still travel to the AI provider — under your own account and terms, which you chose. We say it up front because it protects you from anyone who'd promise otherwise.
Tested like it matters, because it does.
Before Odell's pilot ever touched a real company's knowledge, the permission engine was run through an automated evaluation built from that company's own cabinets — dozens of questions across owner, estimator, crew, and office personas, graded on pricing exactness, citation correctness, honest “I don't know”s, and above all, leaks.
- Zero leaks. Every permission-block, document-gate, and cross-company check in the evaluation passed — a question asked by the wrong person never surfaced protected content.
- Conflicts get flagged, not buried. When two approved sources disagreed during the pilot load — two documents, two different warranty terms — the entry was flagged unsettled rather than silently resolved. That's the system working as designed.
- Approvals are human. Model output is a draft until a person approves it. Always.