Nothing to import
Onboarding a new property was slow because every asset, from an air conditioner to a boiler, had to be entered by hand.
Most of that information isn't in a file anywhere. It's on the identification plate on each machine. There was nothing to import. It was a data-shape problem, not a missing tool, and it's the only part of the problem I pointed AI at.
Watching the work
Instead of designing from the request, I watched people set up assets. They were already working from photos. They'd photograph the equipment in the field, walk back to a computer, and type the details off the photo.
The workflow was already photo-driven. It was just happening in the slowest possible order, and every asset was handled twice. So the product collapsed it into one step in the field: take the photo where the asset is, and let the system read it.
What we built
A user photographs the equipment and its identification plate. AI reads the equipment type, manufacturer, model and serial number, and creates the asset record. I owned it end to end, from discovery to general release.
Piloting the two things we couldn't predict
Two risks couldn't be resolved in testing: what each asset would cost us to process, and how accurate extraction would be on whatever people actually photograph, as opposed to our test images. Per-asset cost decides whether an AI feature is a product or an expensive demo.
So we piloted it with real customers and watched both numbers before opening it to everyone. The pilot ran longer than the build.Deliberately: the build answered what we already knew how to test; the pilot answered what we didn't.
That gate wasn't new for this feature. We'd learned earlier that a module can be called finished before anyone has shown it gets used. Since then, large modules go through a pilot with real customers, and the pilot is judged on usage data.
Result
"We reduced asset onboarding from 30 days to under a week."an Actabl customer, on actabl.com/hotel-ai