Booking.com's AI Agent Lola and a Wave of Travel Assistants Arrive in One Week

Booking Holdings-backed Lola opened to travellers on 30 September. Skift reported the launch exclusively on 29 September, describing it as Booking.com's new AI assistant. The Verge picked it up the same day, and Dev.ua followed on 30 September with a description of an agent that books both travel and entertainment. PhocusWire reported on 29 September that Lola is also being sold as a membership product, not just a free chatbot. That detail matters more than the launch itself: it is the first sign that travel AI is being packaged as a subscription, not a feature buried inside a search bar.

Lola is not the only arrival. KAYAK co-founders launched an AI assistant also called Lola for travel experiences, according to shorttermrentalz.com on 30 September. The naming coincidence has already caused confusion among travel trade reporters. Uber rolled out hotel booking, an eats-for-rides feature and an AI-powered voice assistant, ABC News reported on 30 September. Mindtrip launched a personal travel assistant that learns traveller preferences rather than just trip parameters, per a PR Newswire release the same day. And Airbnb expanded AI search plus social discovery across its platform, reported by konsulteer.com and PhocusWire on 1 October, with finance.biggo.com noting the announcement sent Airbnb shares up 3%.

Five consumer-facing launches in seven days.

The pattern is not accidental. MakeMyTrip's chief technology officer told Inc42 on 30 September that AI is compressing the travel-booking journey. That is the polite industry phrasing for cutting the number of steps between intent and payment. For OTAs built on search results, filters and review pages, a chat window that returns a single bookable option threatens the advertising slots and upsell surfaces that generate much of their margin. That is why the same companies now shipping agents are also the ones with the most to lose if agents work.

Meta Muse looms over the OTAs

The pressure is not coming only from inside travel. Meta's Muse AI assistant hit 5 million downloads, Forbes reported on 30 September, outpacing the early growth of ChatGPT, Grok and Claude. 24/7 Wall St. framed the same number on 1 October as a contrast with Google, which still keeps its own agent behind a paywall. Expedia had already agreed to offer hotel bookings through Muse, a deal first reported on 23 September. TradingPedia described Muse on 30 September as a fresh threat to online travel agencies. The a16z newsletter posed the commercial question bluntly the same day: who gets paid when AI does the shopping?

That question is unresolved. Booking.com's own discount page, updated by WIRED on 1 October, still promotes the Genius loyalty programme with tiered discounts of 10% at Level 1, up to 15% at Level 2 after five bookings in two years, and up to 20% at Level 3 after 15 bookings in two years. It also offers 15% off car rentals and a co-branded Visa Signature card earning 6% travel credits on app bookings. Loyalty is the incumbent's answer to disintermediation: if the agent picks the hotel, the discount still has to be claimed through the account. Whether travellers bother is a different matter.

Not every player is chasing consumers. CoreWeave used its Fully Connected event in San Francisco on 30 September to launch Forge, an integrated software platform for developing, running and improving AI models and agents, as Data Center Knowledge reported. Corey Sanders, CoreWeave's senior vice president of product management, said at a media briefing that the company had moved from being the best provider for AI-centric infrastructure to delivering a range of AI services on top of it. Forge ships in Free, Pro and Enterprise editions with managed notebooks, an observability tool called Agent Lens, and reinforcement-learning rollouts. IDC analyst Dave McCarthy told the same publication that CoreWeave needs a wider enterprise audience if it wants a sustainable business, since big AI labs alone will not carry it.

The infrastructure layer is filling in faster than the policy layer.

Guardrails arrive, with 100 signatories

NVIDIA announced an Open Agent Safety Platform alongside a $150bn addition to its authorised share repurchase plan, ServeTheHome reported on 29 September. The runtime, OpenShell 0.1.0, wraps existing agent frameworks such as Codex, Claude Code, Hermes and Pi in sandboxed environments with kernel-level isolation, rather than replacing them. A gateway manages sandbox lifecycles across fleets, and each sandbox pairs with a supervisor process that inspects outbound HTTP, GraphQL and MCP traffic against policies written in YAML and compiled to OPA Rego. Credential protection keeps secrets outside the agent workload, and every policy decision is recorded in an Open Cybersecurity Schema Framework audit trail. Hardware-level enforcement runs on NVIDIA Sentry on BlueField-4 DPUs, which sit on the only path to the model in Vera Rubin POD systems. The report notes that 100 organisations from NVIDIA's ecosystem have signed on.

The adversarial testing detail is the one worth keeping. According to ServeTheHome, frontier agents spent up to two hours trying to persuade AI reviewers to grant permissions for modifying protected repositories. No protected repository writes occurred in those tests. That is a narrow result, and the 0.1.0 version number suggests the authors know it.

Elsewhere, the week produced two stories that sit awkwardly beside the launch noise. The Guardian reported on 30 September that more than 44,000 people have filed formal legal objections to NHS England over the Palantir-powered Federated Data Platform handling their health data, citing article 21 of the UK GDPR. Palantir says the platform is helping cut waiting lists, with trusts using it recording 117,000 additional operations, a 14.3% reduction in discharge delays for long-stay patients and a 5.6% improvement in the number of people finding out whether they have cancer within 28 days. Two House of Commons select committees have called for Labour to exercise a break clause in the seven-year, £330m deal, which comes into force in February 2027. The Guardian's report is based on its own campaign reporting.

And on 29 September, Rest of World described how Chinese platforms ModelScope and MoArk are competing to become domestic alternatives to Hugging Face, which Beijing blocked in 2023. ModelScope hosts more than 170,000 models and MoArk serves around 20,000, per that report. Xu Yong, chief executive of OSChina, told Rest of World that not everyone can use a VPN all the time, and argued China needed a self-reliant AI ecosystem for Chinese-speaking users. In September, Nvidia announced it was acquiring Hugging Face for $12.9bn. A sovereign stack and a single global hub are now being built at the same time, and the travel agents shipping this week will run on whichever one their operators can access.

For travellers, the practical difference between five assistants launched in a week is small so far. None of the launches disclosed pricing for the consumer agents beyond Lola's membership model, and none published independent performance data. The claims are directional. The download counts and share moves are the only hard numbers on the table.