Answering a question and changing an event registration carry different risks

One attendee asks your event chatbot whether the venue has wifi. Another asks it to cancel their registration.

If the answer about wifi is wrong, the first person may need to ask a member of staff or use mobile data. If a cancellation happens by mistake, you may leave someone without a ticket and release their place to another attendee.

Both tasks can be handled in a conversation. But you do not have to automate them in the same way.

Being able to do it does not mean it should be automated

AI can save you repetitive questions and help attendees handle requests. The judgement lies in choosing where that benefit outweighs the risk you are willing to accept.

Models are improving, but they can still give incorrect answers. OpenAI’s research on hallucinations explains why a convincing answer may contain false information. When preparing your event, you should not plan as though reliability were 100%.

That does not mean abandoning automation. It means considering what would happen if it failed.

Four questions to help you decide

How easily could it get this task wrong? A question with a clear answer in the event information involves less ambiguity than a request with exceptions. “What is the wifi password?” and “Can I cancel after the deadline because something has come up?” need different criteria. Test the tasks you want to delegate to find out where errors appear.

What problem would the error cause? Consider the cost to the attendee and your team: lost time, money, event access, privacy or a commitment to the client. Incorrect wifi information may inconvenience someone attending as a visitor. If that information determines whether an exhibitor can run their demonstration, the consequences change.

Can it be put right in time? Correcting an answer before someone acts on it is different from restoring a cancelled registration whose place has already been taken. Think about who would detect the failure, how much time they would have and what they could actually recover.

What do we gain by automating it? It may save team time or give the attendee a faster answer. Include the effort of reviewing and correcting results. If you need to check every result personally and support barely improves, delegating the task may not be worthwhile.

Controls help reduce errors

You can limit available information and actions, request confirmation before a change or leave certain decisions pending human review.

Another model can also act as a judge: it reviews an answer and checks, for example, whether event information supports it. Anthropic describes this kind of review between models.

These mechanisms can help reduce risk. The reviewing model can also make mistakes, and adding a check does not, by itself, demonstrate 100% reliability.

Ask your supplier which controls they apply and how they have tested them on tasks similar to yours. If you need help assessing the answers, an expert can help you decide whether the remaining errors are acceptable for that use.

You can delegate part of the work and keep the decision

You do not have to choose between automating the entire process and doing everything manually.

For a question about venue facilities, you may allow a direct answer using approved information. For a cancellation, you can let AI collect the request and prepare it for the team. A refund exception can stay with the person responsible.

The same judgement applies to other tasks: the harder a failure is to repair and the more serious its consequences, the more important it is to retain a review before acting.

Choose one specific task from your event. Write down what you would gain, what could go wrong and who would have to resolve it. That gives you a basis for deciding whether to automate it, let AI only prepare the work or keep it with the team.

To see how we apply this approach to attendee support, read how to prepare EVA, our Event Virtual Assistant, with information, permissions and organiser intervention.