How Heimby runs 247 apartments on Proptonomy
Thirty days of production data: 15,257 guest messages in, 94% of the replies sent by the AI, and 670 conversations closed without anyone on the team typing.

Heimby manages short-term rentals in seven Norwegian cities and is the largest operator in Bergen. It is also the operation Proptonomy grew out of, which makes it the portfolio we can be most specific about: we can open the database and count.
So we did. Everything below is measured over the thirty days to 25 August 2026.
The numbers
- 365 apartments synced, 247 active, and 166 of them had a live guest conversation in the window.
- 15,257 messages came in across five inboxes: the Guesty inbox, WhatsApp, Hybel.no, the Microsoft 365 shared mailbox and the internal ops channel.
- 10,741 replies were sent by the AI, 698 by a person. Of everything Heimby sent, 94% was written and sent by Proptonomy.
- 670 guest conversations closed with no staff message at all. Of 1,407 conversations that had a guest message in them, the AI answered 988, and in 670 nobody on the team needed to type anything.
- 2,448 times the AI raised something to a human in the ops channels, and 1,577 follow-ups were scheduled so nothing waited on someone remembering.
- 21,918 AI runs completed, and the knowledge base grew by 461 entries with another 875 updated, learned from the work itself.
The number we find most telling is not the 94%. It is the 2,448 escalations. The AI is not trying to handle everything; it is trying to handle what it should and put a person in front of the rest.
Three jobs, start to finish
Aggregates are easy to wave at. Here are three real cases from the same window, with the guest details stripped out.
A broken toilet seat, closed in 13 hours. A guest reported it in the evening. The AI asked for a photo, got one back within the hour, raised the replacement job against the right apartment, and came back to the guest to confirm once it was done. Nobody on the team touched the thread.
A kitchen that had not been cleaned properly, escalated in 2.6 hours. The guest sent photos. The AI attached them to the existing cleaning task, recorded the defect, then took the photos to the cleaning partner and asked for a confirmed time for the re-clean, rather than just logging a complaint.
A wrong door code, caught twice. A cleaner flagged that the door code on a job was wrong. The AI corrected it, and then went looking: it checked every other open task on that building for the same bad code. That is the kind of work that gets skipped at 17:00 on a Friday.
The spare key that took a week
The clearest picture of what "follows every job through to done" means is a case that ran for seven days across three companies, so here it is with the guest details and the live keybox codes taken out.
Monday evening. Njål asks where the spare key for one of the Bergen apartments is. The AI answers in ninety seconds from the property knowledge: a black keybox in a building down the road. Six minutes later he replies that the box is empty.
The same evening. The AI establishes there is no emergency, because the guest in the flat has access by another route. Then it goes into the history and reconstructs where the key went: last documented use 17 July, when the maintenance partner took it to fix a toilet seat, with nothing recorded about it going back. And on 2 August, when two guests were locked out, the spare-key route failed and a cleaner handed over a key instead. Two leads, from logs a month old.
Twelve minutes after that it opened three lines of enquiry in parallel, each in that party's own chat: the maintenance partner, the cleaning company, and the current guest. It raised a task and asked the ops channel whether it should just order a new key.
Then it kept asking. Tuesday, Friday, Saturday, Sunday, Monday, bundling the key question in with each partner's other open items rather than sending a nag on its own. On the Monday the cleaning company confirmed they had never had it, so the AI reported back and changed tack: it asked the maintenance partner to cut a copy of the master at the next turnover.
Tuesday, 14:35. It turned out the maintenance partner had the original all along, on a Winnie-the-Pooh keyring. No copy needed, just put it back. Three minutes later: "reservenøkkel på plass". The AI closed both tasks and reported it done. Njål's reply was "Helt konge. Godt jobbet."
Seven days, three companies, one key worth a few hundred kroner. Nobody held that thread in their head. That is the part worth paying for: not the dramatic saves, but the small thing that would otherwise have been quietly forgotten until the next guest stood outside a locked door.
What Heimby says
"Proptonomy is steering the wheel. Quality went up and the daily stress went down."
Endre Nesvoll-Jenssen, CEO of Heimby.
Heimby got there the way we would ask anyone to: observe mode first, drafts reviewed for weeks, auto-reply switched on inbox by inbox as it earned it. It now runs with the guest agent on auto-reply and a person handling the 2,448 things a year of that volume genuinely needs a person for.
Heimby has published its own account of the same thirty days, in Norwegian: Slik driver Heimby 247 leiligheter med AI. The story of how the product got here is in our launch post, and the systems it connects to are listed under integrations.
If that sounds like your operation, tell us your listing count and we will show you the same numbers for your portfolio.