Every business seems to want an AI assistant on its website. Most of the ones we see have the same flaw: ask them something they don’t know, and they’ll make up a confident, friendly, completely wrong answer. On a website, a wrong opening time or a price that doesn’t exist isn’t a quirk. It’s a customer turning up to a locked door.
We recently built an AI assistant into a city destination platform: a concierge that helps visitors decide what to do, plan a day and find their way around. Here’s the principle it’s built on, and what it means if you’re thinking about AI for your own business.
The rule: the AI is never the source of truth
The assistant doesn’t know anything by itself. It doesn’t decide what time a bus leaves, whether a café is open or how much a ticket costs. Those facts come from the platform’s own data:
- Bus times from two operators’ own timetable feeds: 61 routes and more than 150,000 timetabled stops.
- Events from two ticketing and listings services, checked before they’re published.
- Places and businesses from a directory of hundreds of listings.
The AI’s job is the part people are good at and databases aren’t: understanding a messy question (“something to do with two kids on a wet afternoon that doesn’t cost a fortune”), finding the right facts, and turning them into advice that sounds like a helpful local. Facts from the data, voice from the AI.
Why that matters
An AI model on its own is trained to produce plausible answers. Plausible and true overlap most of the time, which is exactly what makes the gaps dangerous: the wrong answers sound just as confident as the right ones.
Keeping the facts outside the AI turns that around:
- When the data changes, the answers change. A cancelled event or a new timetable is right the moment the data is, with no retraining.
- When the data doesn’t know, the assistant can say so, instead of filling the gap with something that sounds right.
- You can check it. Every fact the assistant uses can be traced back to where it came from.
What we learned building it
- The data work is the real work. Pulling two bus operators’ timetables into one place, removing duplicate businesses that arrive from several sources, handling events that move or get cancelled: that took far longer than connecting the AI.
- Write the rules down. The platform has dozens of architecture documents, because with an AI in the mix everyone has to agree what the AI is allowed to decide and what it isn’t.
- Useful beats clever. The assistant is judged on whether a visitor ends up somewhere good, not on how impressive the conversation feels.
If you’re considering AI for your business
- Start with the facts you already have. Prices, stock, availability, opening hours, policies. If they’re not in one reliable place, fix that first. It’s valuable with or without AI.
- Decide what the AI may never decide. Prices, legal answers, medical advice, anything with a promise attached should come from your data or a person, never from the model.
- Make “I don’t know” an acceptable answer. An assistant that hands over to a human is far better than one that guesses.
- Test it with awkward questions, not the easy ones. Ask it things it shouldn’t know and see what it does.
- Ask whether plain software would do. Plenty of “we need AI” problems are solved better by a good search box or a clear form.
If you’ve got a problem where AI might genuinely help, and you want it done without the nonsense, that’s our platforms and AI work. The data side of it is data and integrations. Tell us the problem.
Got a problem like this? Good. We like those.
Tell us what you’re dealing with and we’ll tell you honestly what we’d do.
