How our AI works

Every figure KAIA proposes comes from Kurum Group’s own archive of real, signed client proposals — the median of what comparable formations actually cost, carried with the 25th–75th percentile band and the number of proposals behind it, so you can see how firm the number is.

When there is no grounded figure for a line, the system returns zero and flags the line for confirmation. It does not estimate, interpolate, or let a language model fill the gap. A quotation you receive has either a price traceable to prior signed work, or an explicit gap a person has to close.

Work that involves no personal data may use a cloud model. Anything touching client personal data takes one of two paths: it is masked before it leaves, or it is processed on hardware we own in the UAE. There is no third path.

That gate is fail-closed. If masking cannot be applied, the request stops rather than proceeding unmasked — the failure mode is a refused request, never a quiet disclosure. Local inference is the default for routine work, with cloud models used as a fallback.

KAIA proposes; it does not send. A drafted message, a shared document, or a corrected price is written to an approval queue and stays there until a member of staff releases it. The proposing step and the applying step are separate pieces of code, and the second one cannot run without a recorded human decision.

The same applies to the system correcting itself: when KAIA notices a catalogue price drifting from what real proposals charge, it raises a suggestion for review. It never rewrites a price on its own.

When KAIA recommends a jurisdiction or authority, it returns the score behind that recommendation — which signals fired, what each contributed, and which did not apply. The ranking is deterministic: the same answers produce the same recommendation, and the reasoning is inspectable rather than a label attached after the fact.

Language models are used to explain and to converse. They are not used to decide what a jurisdiction costs or which authority is right for you.

Document upload in our public tools is optional — the flow completes without it, and skipping it costs you nothing. Uploads accept images and PDFs only, are size-limited, and are attached to your enquiry for the advisor handling it. They are never served from a public address, and the storage location is never returned to the browser that sent them.

Our e-signature system records each step in a tamper-evident chain — who acted, when, against which exact document — with the signed document bound to the record by cryptographic hash. Actions that change client records are written to an audit log independently of the code that performed them.

We hold no SOC 2 or ISO 27001 certification at the time of writing, and this page is not a certification. It describes controls we built and enforce ourselves; it is not an attestation by an outside auditor.

We also do not claim automatic filing to the Federal Tax Authority. Where a return has to be reviewed and submitted by a person, we say so rather than describing it as automated.

If you are evaluating us against a vendor whose privacy policy and engineering descriptions disagree with each other, the check that matters is whether the two match. Ours are meant to. If you find a place where this page and our behaviour differ, tell us at [email protected] and we will correct one or the other.

The clearest demonstration is the tool itself: start a company setup and watch it narrow to a jurisdiction, show the price band with the sample size behind it, and hand you to a person at the end. No account required.