Asked often, answered plainly.

The models that do your AI work run on hardware you own, in your building. Your records are read and written locally. If a step needs the internet, we show you exactly what leaves the machine and you approve it in the workflow spec before we build it.

Often no. We start by sizing the workload against what you already own. If new hardware makes sense, it is one purchase, one invoice, and it is yours. We will never put you on a hardware rental plan.

It stays where it is. Code and tests run on synthetic data we generate, and real records never enter a repository, a log, or a training run. The data-flow map we write for every project lists what moves and what never moves.

Not on the first month. On the second year, usually yes. Cloud billing is per call, forever. Local is one hardware cost plus power and upkeep. We will give you the real math for your workload before you decide, including the cases where renting is still the right answer.

Yes, by design. Every workflow has named approval gates: a specific person says yes before money moves, content goes out, or a decision is finalized. The software is built to stop at those gates and wait.

Every workflow ships with a fallback path: the work routes to a human, the model's confidence is shown, and the miss is logged so we can tune against it. A wrong answer is an incident with a runbook, not a mystery.

A first deployment for a single workflow typically runs four to eight weeks, in small reviewed slices. You see each slice, approve each slice, and you can stop at any slice with what works so far.