A large Israeli property-management company was losing eight people to manual data entry. We automated the whole flow end to end — wrapped around a fifteen-year-old ERP nobody wanted to touch, and nobody had to.
The operations team was doing robotic work by hand: reading inbound emails, pulling data out of PDFs, and re-typing it into a fifteen-year-old ERP — eight people, around six hours each, every day.
Earlier attempts to bolt on "an employee chatbot" had failed, because they added yet another screen for staff to open. The bottleneck was never a missing chat window; it was the manual bridge between free-text requests arriving all day and a rigid legacy system that only speaks its own language.
Replacing that ERP would have cost millions and taken two years. The distance between the tools the team had and the system they actually ran on was the execution gap — and it was being paid for in hours.
No rip-and-replace. We left the legacy ERP exactly where it was and taught modern AI to work with it.
We stopped the "add another screen" approach and mapped the real service-call and pricing workflow, quantifying the load: eight people burning around six hours a day between them, purely on reading, extracting and re-keying.
Instead of a multi-year, multi-million replacement, we left the old core in place and built a custom integration bridge that lets modern AI read from and write to it — new records in, data out — without touching the system's source code.
We connected the inboxes to an AI tuned to the company's own terminology. It intercepts every inbound free-text request, understands it, and turns the unstructured text into structured data: customer number, fault, urgency, part codes and address.
When the AI recognises a simple request at over 90% confidence, it opens the service call in the old ERP itself, schedules a technician in the diary, and sends the customer a confirmation — with no human touch at all.
For anything complex or unusual, the system drafts the action and sends it to a manager in the company's messaging app. The manager taps "approve", and the system completes the work across the legacy systems in the background.
The whole thing runs behind the scenes, with zero change to how field staff work day to day. The interface they know stayed exactly the same; the hours of typing behind it simply disappeared.
Not a new system to learn. The same operation, with the manual bridge automated out of it.
Average handling time — from an email arriving to an action scheduled in the legacy ERP — dropped from about four hours to thirty seconds. Simple requests now complete themselves; only the genuine exceptions reach a person.
Roughly 80% of the manual data entry across the operation was eliminated, giving an eight-person team back the six hours a day they'd been spending on re-keying — time that moved to the work that actually needs a person.
The discipline is what makes it safe to run unattended: high-confidence work executes itself, everything ambiguous waits for a one-tap human approval, and the old system of record stays the single source of truth throughout. That balance — automation with a person on the edge cases — is exactly the shape of an AI system built to last in production. It's the same approach behind our facilities-management work.
Reef TRH