Jack's Fish Club Operations & software, Miami, FL, USA
AI privacy

What a model sees, and what it never gets to.

A model may read a document you send it. It never gets your price list, your customer file, or your books: QuickBooks connects to our code, never to the model. Here's exactly where the line is.

AI privacy & security

Your data stays yours. Here's exactly how.

The model reads. It never decides.

It transcribes a document. Code prices, totals and reconciles. A person approves.

Your prices and books never leave.

No reference data in any prompt. QuickBooks connects to the code through Intuit's own login; no model is ever connected to it.

Your account. Your key.

Model access runs on your own account, under terms with no training on your data and zero retention where offered.

Keys in a vault. Never in code.

Never in chat either. Logs redact document contents. Staff logins need two-factor.

Stored on your systems.

Documents and extractions live on your storage, with a retention rule you set. Purged on request.

You can say no.

Redact before sending, run a private model, or route a sender to human-only. Trade-offs in writing, before the first document goes anywhere.

The detail, including the workarounds, is right below.

AI information exposure

Every document you let a model read is information leaving your building. Say so plainly, then control it.

When a generative model transcribes a purchase order or an airway bill, the whole document goes to the model provider: the distributor's name, the ship-to address, the quantities, whatever prices they printed. That is the exposure, and pretending otherwise is how people get surprised. The job is to keep that surface as small as it can be and to make sure nothing else ever rides along with it.

So the model gets the document and a fixed schema. It does not get your price list, your account tables, your margins, your customer database or your credentials, because none of that is in the prompt. It answers once, the answer is checked by code against your own tables, and the document is stored on your own account under your own retention rules.

Non-negotiables

  • Models transcribe and classify. They never price, total, allocate or decide.
  • One call per document, against a fixed schema, with no reference data in the prompt. The price list never leaves your systems.
  • Model access runs on your own account and key, under terms that exclude training on your data and set retention to zero or the minimum the provider allows.
  • QuickBooks and every other system of record connect to code through their official APIs and a login you authorize and can revoke. No model is ever given access to them, and nothing a model produces reaches the books without passing validation and a person's approval.
  • Keys live in a vault, never in code, never in chat. Logs redact document contents.
  • Every model output is validated by deterministic code before it can touch money, and a person approves before anything leaves.
  • Documents and extractions are stored on your storage, with a retention rule you set, and can be purged on request.
  • You are told, in writing, which providers see which documents. No silent vendor changes.

Workarounds when the exposure isn't acceptable

  • Redact before sending. Deterministic pre-processing masks the header block (names, addresses, account numbers) and sends only the line-item region. The crosswalk still resolves items; the model never sees who ordered.
  • Private or local models. Run the extraction step on a model you host, on your own hardware or a private cloud. Costs more to maintain and reads scans less well, but nothing leaves.
  • No generative model at all. For distributors with a consistent format, classic OCR plus a per-format template does the job. Slower to onboard a new format, zero exposure.
  • Per-document opt-out. Flag a sender or account as "human only" and the pipeline routes those to a person for transcription while still validating and drafting the invoice.
  • Provider choice per client. A provider with a signed data-processing agreement, a zero-retention API tier, or a region requirement, chosen with you, not for you.

The honest trade-off: the more you restrict what the model sees, the more formats need hand-built templates and the more a person is in the loop. We'll tell you where that line is for your documents before the first one is sent anywhere.

Not sure where your line is?

Tell us what your documents contain. We'll tell you what a model would see, and how to keep it out.

Talk to us →How we build