Skip to content

AI Workflow Audits

Airbnb's CTO says AI resolves about half of its support tickets. The useful part is what it chose not to automate.

Airbnb's CTO gave Latent Space an interview published on October 2, 2026, with large numbers attached. The direct answer for a business owner: do not copy the numbers, copy the method. Airbnb picks which work AI handles, tests each use case against real examples first, keeps people on the high-stakes cases, and requires humans to understand what the AI produced. Those four habits transfer to a business of any size.

By Fabio Rabelo · Founder, ATLACIS ·

What happened

On October 2, 2026, Latent Space published an interview with Ahmad Al-Dahle, who became Airbnb's CTO in January after leading generative AI at Meta. He described Airbnb's push to become an "AI-native company" and offered several figures. About 60% of Airbnb's code is now AI-authored. The company has shipped nearly 80% more features and improvements year over year, and pull-request throughput for the average engineer is up about 1.6 times. On customer support, he said roughly half of tickets are now resolved purely by AI. Latent Space notes that Airbnb's own second-quarter results put the figure at nearly 45%. In those results, Airbnb also reported that support cost per booking fell about 16% year over year, driven in part by its AI assistant. These are the company's own numbers, given in an interview and a earnings release. They are not an independent audit.

Why it matters for business owners

Most AI headlines give you a capability. This one gives you an operating pattern, and the pattern is more useful than the percentages. Al-Dahle called support "the hardest problem to deploy" because the cost of a mistake is high. So Airbnb did not turn it on everywhere. He said the team builds the agent, then generates a battery of synthetic test cases before it goes live, and stays "deliberate about the tickets we don't choose to solve yet," safety issues being one example. In other words, half automated does not mean half careless. It means the other half was chosen on purpose. He also said Airbnb runs evaluations for each use case, using sampled production queries and including edge cases. And it does not use one model for everything. Coding gets the strongest available model because a defect is expensive. Search is latency-sensitive at scale, so it uses smaller, faster models. The point is matching the tool to the job and the cost of being wrong.

What owners should not misunderstand

Do not treat 50% as a target. Airbnb is a very large company with a large engineering team, years of support history, and the ability to customize its own models. A business with a small support inbox is in a different position, and its safe percentage may be lower or higher. Do not read "resolved without a human agent" as "the customer was satisfied." It is a resolution measure, not a quality guarantee. Airbnb pairs it with testing, but the interview does not give satisfaction data, and we should not invent any. Do not assume that 60% AI-authored code means 60% less engineering. Airbnb's own CTO is worried about something else: whether junior engineers still build judgment when AI does so much of the work. His answer is a rule that every engineer must be able to explain what they built, even when AI generated the pull request. And do not forget the source. This is an executive describing his own company in a friendly interview. It is useful as a pattern. It is not proof that the same results are waiting for you.

The operational lesson

The expensive mistake in AI adoption is not choosing a weak tool. It is automating a workflow without deciding where it must stop. Airbnb's approach reads like a short checklist. Choose the work AI should handle. Test it on real examples, including the awkward ones, before customers see it. Keep people on the cases where a mistake is costly. Make sure someone on your team can still explain the output. Notice what is missing: a promise that the tool will be accurate on its own. The reliability comes from the testing and the boundaries around the tool, not from the model.

What a serious business should do next

Pick one workflow, such as support replies, quoting, or internal reporting. Split its work into three piles: routine, judgment-heavy, and high-stakes. Automate only from the routine pile first. Before you switch anything on, collect twenty to thirty real examples, including a few messy ones, and check the AI's answers against what a good employee would have done. Write down which cases will always go to a person, and how a customer can reach one. Then name an owner. Someone must be able to explain what the AI sent or produced, and to turn it off. If nobody can, the workflow is not ready. If the workflow is unclear, or you cannot say what a mistake would cost, do not automate yet. Map it first.

The Atlacis view

Atlacis helps owners slow down before an AI decision, understand the workflow, and decide what to automate, what to keep human, and what to leave alone. Airbnb's interview is a good illustration of the same discipline at a much larger scale: the results came from choosing and testing, not from switching AI on. If you are weighing automation for support or another customer-facing workflow and want a second opinion on where to draw the line, that is a good conversation to have before you buy a tool.

The short version

  • On October 2, 2026, Airbnb's CTO told Latent Space that AI resolves roughly half of support tickets and authors about 60% of its code. These are the company's own figures.
  • Airbnb's own Q2 2026 results put the share of issues resolved without a human agent at nearly 45%, and support cost per booking down about 16% year over year.
  • The transferable part is the method: test each use case on real examples, keep people on high-stakes cases, and require humans to be able to explain AI output.
  • Do not copy the percentages. Airbnb's scale and custom models are not yours, and "resolved" is not the same as "satisfied."
  • Decide what AI must not handle before you decide what it will handle.
Tags:AI workflow auditAI customer supporthuman reviewAI evaluationbusiness AIAirbnb
FAQ

Common questions

Should my business aim for half of support tickets handled by AI?
No. That is Airbnb's reported figure at its own scale, and it is not a benchmark for other companies. Start with the routine, low-risk requests, test on real examples, and expand only if the results hold.
What should always stay with a person?
Cases where a mistake is costly or hard to undo, such as safety, money disputes, legal or contract questions, and upset customers. Airbnb says it deliberately leaves some ticket types, including safety issues, out of automation for now.
How do I test an AI workflow before launch?
Collect real past cases, including awkward ones, and compare the AI's answers with what a good employee would have done. Keep that set and rerun it whenever you change the tool or model.
Keep reading

More from the blog

The AI workflow audit checklist: what to check before you automate anything

Automation is now a switch inside software you already pay for, which makes it easy to automate a mess. Here is the short screen to run before you flip anything on.

A new benchmark tested AI coding agents on real company code instead of public GitHub repos. Here is what business owners should check before picking one.

On September 12, 2026, startup Specific Labs published Real-SWE, a benchmark that tests AI coding agents on licensed, private production code rather than public repositories. The best model and tool combination solved fewer than four in ten tasks, and on six of the ten disclosed tasks, every model tested solved less than 15 percent of the time. The more useful number for a business is not who topped the leaderboard. It is that the top-ranked model cost more than double, per working change, than the cheapest one tested.

Stanford's newest data shows the AI hiring gap for young workers has widened to 19 percent, and economists still disagree on why. Here is what business owners should know before cutting entry-level roles.

Stanford researchers revised their ADP payroll study this month and found employment for 22-to-25-year-olds in AI-exposed jobs is now 19 percent below where it would be if it had kept pace with peers, up from 15 percent a year earlier. Economists studying the same data disagree on whether AI is actually the cause. Here is what that means for a business owner deciding who to hire next.

Make better AI decisions, starting with one call.

Book a free AI Fit Call. We will tell you what to use, what to avoid, and where to start. No jargon, no pressure.