The problem with most AI chatbots
A language model will answer any question confidently, including questions about your business it knows nothing about. Put one on your website without guard rails and it will invent opening hours, prices and policies. That is worse than having no chatbot at all.
So the first rule of every AI assistant we build is simple: it answers from your data, or it does not answer.
How we build an AI assistant
- Retrieval first. The question is matched against your own content (listings, products, policies, help pages) before the model sees it.
- The model proposes, the platform decides. The model can suggest an answer or an action, but only from what retrieval returned. Anything else is refused.
- A grounded draft before the clever bit. The page shows the answer built from your data straight away, then the conversational version arrives. If the model is slow or down, the visitor still gets an answer.
- Honest about gaps. When the data does not have the answer, the assistant says so and points to a person.
- Provider-agnostic. We have built on models from more than one provider, and design so you can switch.
DeeDee, Visit Dundee's city assistant
On Visit Dundee we built DeeDee, an assistant that answers questions about the city from the platform's own listings: places to eat, things to do, events and businesses. The model never writes a recommendation of its own. It proposes one action and names a place, and anything not already returned by the search is refused. It has a visitor mode and a local mode, which decide which part of the platform it searches first.
Because the grounded answer appears first, the page went from 3.7 seconds to 0.31 seconds, and it still works with JavaScript switched off. You can try DeeDee on the live site.
Automation that does not need a prompt
A lot of the most useful automation has no AI in it at all. It is jobs that run on a timer and do the dull part of the day:
- Doorstep Glass Recycling's platform builds every round's stop list overnight, re-orders the route when a stop is added, and sends each customer a reminder the evening before by push, text or email.
- PLZ pulls fixtures, results, squads and video in on a schedule, and alerts us when any feed goes quiet.
- Visit Dundee merges event listings from several feeds into one, without duplicates.
We will often suggest automating the process before adding AI to it. Once the data is clean and in one place, AI becomes much more useful and much less risky.
AI-drafted content, with a human in charge
Where AI helps write content, we keep a person in the loop: drafts are generated with their sources attached, reviewed, and only then published. Nothing goes out on its own. For a frank look at where AI helps and where it does not, read what an AI assistant can actually do for a small team.