Airbnb targets 2027 for its AI travel agent
The planned tool would pair maps and photos with help on reservation changes, payments and refunds.
By Saeed Ezzati8 min read
The audio edition
Listen to this newsletter
0:005:06
Read transcript
Airbnb is targeting the first half of 2027 to begin rolling out an AI travel agent—and it wants that agent to help people choose a trip, not just choose for them. In a November 5 shareholder update, the company described an experience built around travel rather than a standard chat window. A guest’s question could bring up a changing mix of maps, listing photos, comparisons, calendars and itineraries, assembled in real time to fit what they’re trying to work out. That design reflects Airbnb’s view that research is part of the pleasure. The company says 80 percent of travelers enjoy researching trips, and that its bookings average three guests. It wants people to compare options and make decisions together, rather than hand the whole choice to an automated assistant. Airbnb also points to how guests use its service now: more than 75 percent use maps when choosing a stay, over 90 percent browse multiple listing photos, and 70 percent read reviews. Those are figures about current behavior—not results from an agent that hasn’t launched. Airbnb’s other argument is what happens after someone picks a place. Because it handles reservations, payments and customer support, the company says its agent could help with changes, payment problems, refunds and rebooking, as well as 24-hour safety support. That connection to the booking, Airbnb argues, could let it do more than an outside assistant that can recommend a stay but can’t manage the reservation. The longer-term ambition is broader: planning a whole trip in one place, including homes, hotels, services and experiences. But that is a vision to build toward, not the promised scope of the first release. The first-half 2027 date is a target for starting a guest rollout—not a guarantee that the full trip-planning system arrives then. That push to move AI from answers into action is also showing up in workplace software. Cursor’s new Build Live Apps lets users prompt a cloud agent to create a web app, set up hosting and sign-in, and, if needed, provision a Postgres database. It’s available on paid plans with cloud agents enabled. Users can start from new or existing repositories on Cursor Origin or GitHub; GitLab support is described as coming soon. The key boundary is that generating an app doesn’t publish it. A user must choose a domain and explicitly release it, and Cursor says security checks look for common problems such as exposed credentials. Passing those checks is not a guarantee that an app has no vulnerabilities. For teams, administrators can add review rules, authorize a release that was blocked, see team apps and secrets, and take an app offline. So the promise is faster software creation, with a human release decision and some administrative controls still in the loop. The risks of putting AI into real workflows are more severe when it’s used to impersonate the people authorizing them. Federal prosecutors in New York have indicted three people over an alleged scheme to steal about 129 million dollars. The accusations describe emails posing as company executives, followed by calls that allegedly used AI to mimic a CEO’s voice—and sometimes appearance. A supposed executive would then refer an employee to a fake attorney, who allegedly directed large wire payments to accounts around the world. Prosecutors say the operation targeted thousands of companies over more than 18 months. That’s the alleged reach of the targeting, not a count of companies that paid. Court documents say the money was converted into cryptocurrency in an effort to conceal its origins. The indictment names Jacob Soall, Sharon Hazan and Guy Sheetrit; all three face charges including wire fraud, aggravated identity theft and money laundering. The accusations remain unproven, and an indictment is not a finding of guilt. The alleged method is a reminder that convincing voice or video can reinforce a familiar business scam: a confidential payment request, followed by pressure to act. The case is now moving through the courts, where the allegations will be tested. And while some tools turn prompts into a single app, Perplexity is offering agents meant to keep working on an assignment over time. Its Teammates feature is available on the web to Max subscribers. Perplexity says each agent keeps project context, skills and memory, runs in the background, and can bring in a person when it needs a decision. Agents can use desktops and browsers, coordinate with other agents, and follow up with coworkers in Slack. Users can inspect an agent’s activity, ask for a status update or change its plans. Perplexity’s examples include monitoring overnight engineering alerts and preparing a code change that still needs human approval before deployment. Those are company illustrations, not independently verified results. Each agent has a separate desktop and browser session, but uses the user’s existing skills, plugins and permissions. For now, access is limited to Max subscribers on the web; iOS, Android and desktop versions are promised without release dates. The launch includes promotional usage credits, but Perplexity hasn’t disclosed their size or expiry. After that, use draws on credits already included with the subscription, so the available launch allowance is unclear. Across these launches, the practical test is not just whether an AI can produce something useful. Watch what it can actually access, when a person must approve its next move, and whether the promised controls hold up once these agents are doing real work.
