Skip to content

AI data handling

Know exactly what the AI sees, what it keeps and why it cannot invent your data.

One provider, zero retention

Kabaido's AI runs on Anthropic models only, configured for zero retention: your content is not kept by the provider after a request completes and is not used to train models. We do not train on customer data either. No other AI vendor receives your data.

What is sent

An extraction call carries the content of the request being read (its text, tables and any scanned pages), your organisation's units and currency context, your schema field names and the relevant domain knowledge packs. It does not carry your whole catalogue, your customer base or your pricing history.

Cite or abstain

Every extracted value records the exact place in your document it came from. A value the source does not state stays empty and raises a clarification. The AI never invents a specification, tolerance, price or part number, which means an extraction error is visible and traceable rather than silent.

Where the AI is not involved

Fit scoring, match bands, pricing resolution and every engine calculation are deterministic code, not model output. Prices come from your own catalogue, formulas and rules. The AI reads and ranks; it does not price.

Connected accounts

Connected inbox and calendar tokens are encrypted at rest and used only to run the features you connected.

Catalogue search

Catalogue search and product matching run in Postgres inside our own infrastructure: text search over your rows, part number similarity and filters on your typed attributes. Making your catalogue searchable sends no product data to an AI provider or any other third party.

Why the split matters

Reading is a language problem and pricing is not. A language model is very good at working out that "12 dia x 4 flute alicrn" is a 12mm four flute end mill with an AlTiCrN coating, and it has no business deciding what that costs you. Keeping the two apart is what makes an answer auditable: the reading step shows its citations, the pricing step shows its rule chain, and neither is asked to vouch for the other.

It also means the parts you would want to reproduce in front of a customer are reproducible. Run the same enquiry against the same catalogue tomorrow and the fit scores and prices are identical, because no model was involved in producing them.

Kabaido Meet

Meetings hosted on Kabaido Meet involve no AI at all during the call, and the transcription that runs afterwards is an open source speech model on our own hardware, offline. No meeting audio and no transcript reaches an AI provider.

Spend protection

AI usage is metered per call with model, token and cost accounting in a usage ledger. The monthly usage allowance and, on the free plan, a hard provider cost cap stop a runaway document from running up usage unnoticed. Rate limits sit underneath that as a burst guard: the AI routes run on a per organisation token bucket, so a loop in somebody's integration is throttled rather than billed.