Airbnb reported its Q3 numbers recently, and one line in particular caught my eye: its customer support cost per booking fell more than 17% year on year, which Brian Chesky put down “in large part” to improvements in its AI assistant. According to CX Dive’s write-up, the assistant resolved around 40% of all support issues in the quarter without a human agent, it’s now being extended to voice support in the US, and customer service NPS is, in Chesky’s words, “as high as it’s ever been”.
Now, I’ve read a lot of AI success stories this year, and most of them follow the same script: “we deployed AI and cut X people”. This one is more interesting to me, and not because of the size of the number. It’s because of what they chose to measure.
Cost per Booking Is the Right Metric
Let me explain. “We cut headcount by 20%” tells you something happened to your payroll. It doesn’t tell you whether customers are better or worse off, and it tells you nothing about what happens when volume doubles next summer. “Support cost per booking went down while NPS went up” tells you the unit economics of the business improved and the people paying you didn’t notice anything except getting their answers faster. That’s two numbers, pulling in opposite directions if you cheat, which is exactly what you want from a metric.
I keep seeing companies measure AI adoption by inputs: how many licences, how many prompts, how many people “use AI weekly”. That’s like measuring a sales team by how many emails they send. If I were reviewing an AI programme as a CTO, the first question I’d ask is: what’s the unit cost of the thing we do, and what’s the quality signal that stops us gaming it? For support it’s cost per booking and NPS. For engineering it might be lead time per change and change failure rate. Pick the pair, then let the AI argue for itself.
The 60% Nobody Talks About
The other half of that statistic deserves attention too: if around 40% of issues are resolved without a human, roughly 60% still need one. Those are, almost by definition, the messy ones: the cancelled flight, the host who isn’t answering, the guest standing outside a locked door at midnight. So the human support people who remain are handling a harder mix of problems than before, and they need better tools and more authority, not a smaller team being told to “do more with less”. Don’t get me wrong, the savings are real; but the companies that pocket the savings and forget about the humans in the loop are going to see that NPS line move in the wrong direction.
And the UI Bit
One more detail I loved: Airbnb says it’s building its own AI travel agent for the first half of 2027, which “draws from a library of hundreds of travel-specific interface components” to build its responses in real time. Read that again. The screen isn’t designed up front any more; the agent assembles the interface it needs for your question. I’ve been saying the UI as we know it is on its way out, replaced by agents that do the clicking for you, and this is what the transition looks like in practice: the agent first, the pixels generated on demand.
Bottom line: if your AI programme can’t name its unit cost and its quality signal, it isn’t a programme yet, it’s a hobby. A very expensive hobby, in some cases (ask anyone who has seen this year’s token bills) 🙂





