An AI assistant for a trade show portfolio

The problem

Most questions that reach a show team are repeats. Dates, hours, parking, how to exhibit, whether a registration went through. They show up at night and on weekends, and every one pulls someone off real work.

A generic chatbot can recite an FAQ, but it can’t see your own data, it doesn’t know when to stop and get a human, and nothing it learns makes it back to the CRM.

What I built

An AI assistant on the show websites, plus a version for staff inside Microsoft Teams.

• One assistant per show. Each gets its own facts, its own allowed links and its own prompt, so answers never bleed between events.
• Answers from a curated knowledge base, not the open internet. During registration season it could also check whether someone’s registration went through.
• Knows when to hand off. Every reply is either an answer or a handoff to a person, nothing in between, with two tiers of escalation.
• Consent first. The chat only loads after a visitor accepts cookies.
• Remembers. A summary of each conversation is written to the contact’s record in the CRM, so the sales team sees what a prospect asked.
• Staff version in Teams. Staff ask for registration counts, session sign-ups or exhibitor staffing gaps and get the answer in the chat, without opening three systems and at no hosting cost.

How it works

1. 1 Visitor asks Web chat on the show site
2. 2 Automation picks it up Pulls the conversation so far
3. 3 AI answers or hands off Show's own knowledge base, strict answer/handoff format
4. 4 Reply goes back Into the same chat thread
5. 5 CRM is updated Summary on the contact record

A few design choices that mattered more than the model:

• A fallback model. If the main model is unavailable, a smaller one answers instead of the assistant going quiet.
• Answers only from what it’s given. If the knowledge base doesn’t cover it, the assistant says so and routes the question.
• Separate per show, shared where it’s safe. Only a common block about the show family is copied between assistants.

The result

It’s scoped to handle about 85% of inbound questions. Visitors get an answer at 11pm on a Sunday, and the questions that do need a person arrive already summarized in the CRM.

It also paid for itself: retiring a $5,000 a year tool covered it, for a net saving of about $3,100 to $3,700 a year.