A chat assistant on a destination website promises a lot. It answers at 10pm, in the visitor’s own words, from everything the site knows. Our own platform sells one, Concierge AI, so we have an interest in saying yes. This article explains when the honest answer is “not yet”.

Two visitors, one question

Two families ask the same thing on a Thursday night: “Is the falls walk open this weekend, and can we take the pram?” The first family reads a page that says, plainly, that the walk reopened in September, that the first 400 metres are sealed and the rest is gravel. The second asks an assistant. If the page behind it says those things, the assistant gives a good answer in seconds. If the page doesn’t, the assistant guesses, and it guesses fluently.

An assistant is a new way to read your content, and it can’t create knowledge your site doesn’t hold.

You own what it says

Two points from Australian regulators are worth putting in front of your board before anyone signs a contract.

The ACCC’s December 2025 industry snapshot on artificial intelligence cites the 2024 Canadian case Moffatt v Air Canada, in which the tribunal “held that Air Canada was liable for a misleading representation made by an automated chatbot to a customer on its website” (ACCC, December 2025⁠ (opens in a new tab)). It is a Canadian case, but the principle is easy to follow: the assistant speaks for the organisation that runs it.

Treasury’s October 2025 review of AI and the Australian Consumer Law makes the same point at home. A business “may engage in misleading or deceptive conduct if it allows consumers to form a reasonable expectation that the chatbot will assist them” with something and then it fails to. The review also notes that a small business buying an off-the-shelf chatbot service may have recourse to its supplier (Treasury, October 2025⁠ (opens in a new tab)). Recourse is not the same as not being responsible.

The privacy regulator adds two expectations. The OAIC’s guidance on commercially available AI products says organisations should make sure public-facing tools such as chatbots are “clearly identified as such to users”, and reminds them that “AI systems are known to produce inaccurate or false results” (OAIC⁠ (opens in a new tab)).

This is not legal advice, but it is a reason to treat an assistant as an editorial product.

Travellers check AI answers

The travellers most likely to use an assistant are also the ones most likely to check it. Smartraveller reported in July 2026 that 30% of Gen Z travellers use AI tools to plan overseas trips and 82% are likely to cross-check what AI tells them; its advice lists local tourism bodies among the sources to check against (Smartraveller, July 2026⁠ (opens in a new tab)). The research is about overseas travel, but the habit carries over. An assistant that contradicts your own pages loses the trust it was meant to build.

Three tests to pass first

1. Is the content current enough to answer from? Opening hours, seasonal closures, event dates, accessibility details and getting-here information should be right on the pages before an assistant repeats them. If your team couldn’t confidently answer the twenty most common questions from the site alone, an assistant can’t either. Start with the counter questions in What a visitor information centre knows.

2. Can someone take the handover? Good assistants know when to pass a conversation to a person: a complaint, an accessibility question that the content doesn’t settle, a group booking. If nobody on a team of three can answer those within a working day, the handover becomes a dead end, and a dead end is worse than a contact form.

3. Will you own every answer? Someone needs to read a sample of conversations every week, fix the content behind wrong answers, and decide what the assistant should decline to answer. If no one has that job, the assistant will drift. Budget it as staff time every week, not only as a subscription, and ask any vendor, including us, for the add-on’s cost for your site up front.

If you pass all three, an assistant can be useful, especially out of hours and for visitors who prefer asking to browsing. If you fail any of them, the same budget spent on the pages will help both people and the external assistants that read your site (How an AI assistant reads a destination page).

What to ask any vendor, including us

  • Does it answer only from our content, or from the open web as well?
  • How does it say “I don’t know”, and what does it do next?
  • Can we see every conversation, and fix the content behind a wrong answer?
  • How is it labelled as an assistant on the page?
  • Which model does it use, who owns the knowledge we load into it, and where is conversation data kept?
  • What happens to the handover outside office hours?

The national context

Governments are starting to fund this at scale. In July 2026 New Zealand’s Minister for Tourism announced $800,000 to expand Tourism New Zealand’s AI travel assistant, adding “more detailed information from regional tourism organisations” (New Zealand Government, July 2026⁠ (opens in a new tab)). The assistant is only as good as what the regions give it.

On Roam

Concierge AI is a live add-on for Roam sites. It answers from your own products, events and stories, in your destination’s voice, 24/7; you choose OpenAI, Anthropic Claude or Google Gemini models, keep ownership of your destination knowledge, and get analytics on visitor conversations; and it hands over to your team with full context when a person is needed. It won’t place your destination in anyone else’s AI answers, and it isn’t the right first step for every team. See how Concierge AI works.