We didn't read about it in a report.
It happened to us.
When we launched PulseCosta, our leisure app for the Costa del Sol, questions started pouring in at all hours: opening times, prices, whether there was a terrace, whether children were welcome. In Spanish, English, Finnish and Swedish. And almost always out of hours.
Talking to venues, we saw the same thing: restaurants, hotels and apartments were losing bookings because nobody replied at night or at weekends. The chatbots they had tried answered with generic phrases or, worse, made things up.
“A customer who asks at midnight and gets no reply books somewhere else.”
We built our own assistant and learnt three things the hard way:
1. Without your data, AI makes up answers. In early tests, the model even recommended venues that didn't exist. Since then, our assistants only answer with the business's own information and cite where each fact comes from.
2. Sending data to an external API is expensive and it isn't yours. We started with a cloud AI service. We moved to our own model on European servers: the cost per query dropped to zero and the data stopped leaving.
3. A public chatbot can be attacked. Before going live, a security specialist tried to trick it into revealing internal information. Today, every one of our assistants passes those tests before it is published.