Documents nobody wants to open. Cameras nobody watches. A process that runs on one person's memory. Describe it in a paragraph - we will come back within two working days with a straight answer about whether this is a good fit, and what a two-week proof of concept would look like.
Usually less than people expect. Document work often starts with zero labels - we define the schema and the models generalise. Vision work typically needs a few hundred examples per class to get a useful baseline, and the pilot itself generates most of them, because every reviewer correction is a label. Where you genuinely have nothing, we start with anomaly detection trained only on good samples.
None, before we have seen your data - and be careful with anyone who does. What we commit to is a measured baseline within two weeks, per field and per class, on your own material, along with the list of cases that fail. You decide whether those numbers are good enough to build on. In practice, clean document families land above 99% field accuracy and hard handwriting sits closer to 90%, which is why the threshold and review queue matter more than the headline number.
No. Your data trains models that serve you, inside your boundary, and nothing crosses to another customer. If a workload benefits from a third-party frontier model, that is an explicit, per-workload opt-in you make in writing - and there is always a path that keeps everything on your own hardware instead.
That is a normal deployment for us, not an exception. Air-gapped installs run with offline model updates delivered on approved media, and edge nodes store-and-forward through outages, reconciling when the link comes back.
It should be wrong loudly rather than quietly. Every output carries a confidence; anything below your threshold routes to a person with the page or clip already open at the relevant region. Automated writes are logged with a defined undo path, and the correction feeds back into the next training round. A system that silently guesses is worse than no system, and we design against that first.
Yes - SAP, Oracle, Dynamics, Tally, OPC-UA and MES historians, DICOM and HL7/FHIR are all standard for us. Where a supported API does not exist, we build the adapter as part of the engagement rather than handing you a CSV export and calling it integration.
The proof of concept is a fixed fee scoped in the first call. Production is priced on volume - per document, per stream, or a flat platform fee for larger deployments - plus infrastructure you own and control. We will give you a cost-per-unit figure in the pilot report, so the business case is arithmetic rather than optimism.
Then we tell you, usually in week one or two, and you have spent very little finding out. Roughly one in five scoping conversations ends this way - most often because the data does not exist yet, or the downstream process nobody has agreed on is the real bottleneck. We would rather say it early than bill through a pilot we do not believe in.