AI in first-line tenant contact: what works and where you lose your tenant

Monday morning, you open your system's inbox and scroll through the automated replies from the weekend. You stop at ticket 47. A tenant asked on Friday evening whether the replacement of his boiler had been scheduled. The bot replied neatly with an explanation of the reporting procedure and a link to the tenant portal. The same tenant called on Saturday morning, irritated, with the same question. You look at the transcript and think: I could have solved this manually in ten seconds, and now I have an unhappy tenant on top of it.

That's the pattern almost every property manager runs into once AI sits somewhere in the first line of contact. The idea that a chatbot catches standard questions round the clock holds up, but it breaks down the moment a question touches something specific. And that exact distinction between "works brilliantly" and "tenant gone cold" is rarely made in advance.

Where AI genuinely works in first-line contact

The questions suited to automation are the ones you get a hundred times a month, phrased in exactly the same way. What are the opening hours, how do I report a repair, when is my rent collected, where do I find my annual statement. A large share of that traffic comes in outside office hours. A conversational chat implementation described in CorporatieGids shows that roughly ten percent of messages arrive outside opening hours. That's exactly the flow you want to catch, so a tenant doesn't have to wait until ten o'clock on Monday morning for a status update that's already sitting in their portal.

What works here is that the question has one clear answer, the answer doesn't change over time, and the tenant isn't expecting escalation. The bot then confirms what the tenant already suspected, and everyone's happy.

Where it breaks down

It starts to grate the moment a question needs context that sits outside the general FAQ. A tenant asking "why hasn't my repair been scheduled yet" doesn't want an explanation of how reports are categorised. He wants to know why his report from three weeks ago still hasn't gone through. An AI that misses that and gives him a generic answer feels, to the tenant, like a landlord who isn't listening.

Research by the Woonbond (the Dutch tenants' association) showed that twenty seven percent of all reports to its helpdesk concern landlords who don't respond to complaints. An automated answer that fails to actually answer the question falls into that same category for the tenant. If anything, a bot that fires something back quickly without engaging with the substance of the report feels even more cynical than no response at all.

The Aedes benchmark 2025 backs this up. In the repairs process, a good explanation of why a repair can't be completed straight away is worth one and a half to two points of improvement on the tenant satisfaction score. That's exactly the kind of explanation a first-line AI can almost never give well, because the real reason sits in the file, not in the FAQ.

What helps you draw the line

There's a rule of thumb that works well in practice. If a question can be answered without looking into that specific tenant's file, the AI can handle it. The moment the answer needs person-specific context, automation breaks down. Directions, opening hours, generic procedures, payment references, explanations of energy labels, notice periods: those are all questions AI handles fine. A specific report, an ongoing repair, a payment arrangement, a complaint about a neighbour: that belongs with a person.

What AI can do within that second category is prepare the ground. Triaging a report based on free text, flagging that the same tenant already called last week, drafting a reply your staff member can personalise in fifteen seconds. That's a different path from answering automatically, and it often saves more time than a bot handling the question itself.

The hidden cost of getting automation wrong

The Aedes benchmark also shows that for repairs, the average score is 8.5 when the repair is resolved in one visit, and drops to 6.6 once an engineer has to come round four times or more. The same mechanism applies to communication. A tenant who gets an unsatisfactory AI answer three times for the same question files a formal complaint the fourth time. That complaint costs you an hour of casework, while the original question could have been dealt with in two minutes.

So the gain from AI in first-line contact doesn't come from handling as many messages as possible, but from handling the right messages. A forty percent deflection rate on the right category is worth more than sixty percent on a mix where a quarter of tenants come back dissatisfied anyway. The question for your own portfolio isn't whether you use AI, but which question flows you explicitly exclude from automation, and how you make that visible to the tenant.