What can go wrong when AI answers your attendees

Imagine an attendee arrives at a dinner convinced they have an invitation. Their ticket does not include access, but the event’s AI agent told them it did.

Your reception team has to resolve the situation in front of that person. To the attendee, the answer came from the organiser.

If you are considering an event chatbot to support your audience, you need to know how your supplier handles situations like this. A demonstration with easy questions shows how it converses. To decide whether you can trust it with a task, you also need to see how it handles what it does not know and what it has no permission to do.

Studies reflect the concern

In a PCMA survey of 92 events professionals, 59% identified security and privacy as a concern about generative AI. A further 48% mentioned a lack of technical expertise. These are responses from that sample, not percentages of the whole industry; the base for each question is not separately specified. You can read the original PCMA article and explore its source in Athena.

The scenarios in this article are examples for evaluating a solution. The study provides context about concerns; it does not measure how often these failures occur.

A convincing answer about something nobody confirmed

“Yes, you can change the name on your ticket until tomorrow.” It sounds helpful. But does it come from approved terms, or is it a plausible answer the agent constructed?

When someone makes a decision based on that answer, a refund policy, access condition or transport service may end up being promised without authorisation.

A correct source is the starting point. Appropriate technology can reduce errors by combining approved information, checks on answers and a route to resolve unanswered questions. To assess whether the result is acceptable for your event, ask for tests covering the tasks you intend to delegate.

Ask your supplier: how do you check that a claim is supported by event information? What happens when a fact is missing or two documents contradict each other?

Ask them to demonstrate a question whose answer is not in the documents. Observe whether the agent recognises the missing information, avoids making a commitment and leaves the enquiry accessible to the team.

Correct information reaching the wrong person

A timetable may be accurate but belong to a private activity. Or a conversation may contain another attendee’s personal data.

Protection should limit what information the agent receives for each user and what actions are available to it. Writing “do not share private data” in its instructions does not replace those limits.

Ask your supplier: what data about other attendees can the agent consult during a conversation? How are permissions enforced when someone requests information or an action reserved for another person?

Test the same service with two profiles that have different access rights. Check that each receives the appropriate information and that a chat request does not grant new permissions.

An answer that was correct yesterday

You have changed a room, cancelled a transfer or updated a time. The team knows the new version, but the agent continues answering with the old one.

During an event, you need to update information quickly and check what answer an attendee receives afterwards.

Ask your supplier: can you make the update yourself? When does the agent start using it? What happens if the old information still appears in another document?

Test a timetable change. Verify the new answer and agree how to notify people who already received the previous information. Updating knowledge does not automatically correct earlier communications.

A conversation nobody knows needs help

An attendee asks about a specific need. The agent cannot answer. The question remains pending while the team assumes attendee support is covered.

For that person, automation becomes a wait with no clear way forward. You need to find those enquiries, assign them and answer from the team when necessary.

Ask your supplier: where do unresolved questions appear? Can you intervene in the conversation? How do you know someone on the team has taken responsibility?

Ask to see the complete journey, from the question to the human answer. Then assign someone to review those cases during your event.

The agent starts discussing something you never approved

A conversation about timings may turn into a refund request, an access exception or a matter unrelated to the event.

Define which topics the service will cover and which actions it may take. Explaining a policy does not imply authorisation to grant an exception.

Ask your supplier: how do you limit conversation topics? What does the attendee receive when their request falls outside the scope? Which actions require a person’s approval?

Ask for a test of an out-of-scope request and another of an action without permission. Check how the answer or action is stopped and what alternative the attendee receives.

How to assess your supplier’s answers

Using ChatGPT or Claude to prepare a draft you then review is a different situation from having an agent answer on behalf of your event. The tool’s name alone does not tell you how its controls are configured.

Bring your own cases to the demo. Ask for test results comparable to your intended use and distinguish errors by their consequences: unclear wording and an incorrect access authorisation require different responses.

If you need help assessing the solution, involve experts who can review how data access is limited, how answers are checked and how the team intervenes. Ask them to translate their assessment into a practical decision: which tasks you can delegate, with what supervision and under what conditions.

At Sinapser, we apply this approach to attendee support. The next article explains how to prepare EVA (Event Virtual Assistant) with approved knowledge, user permissions and team intervention.