Nine things AI still can't do for an event team

Sixty-five percent of event professionals use generative AI tools. Sixteen percent say it has significantly improved planning and execution (Northstar/Cvent PULSE, 2026). Another 49% report modest gains. The survey was run with Cvent, a company that sells AI-powered event software, which makes the modesty of the result more notable, not less.

A smaller PCMA study, with only 92 respondents, found a similar shape: 91% of business-events professionals use AI in some way, but only 15% integrate it strategically (PCMA, 2025). The same study found the top application was content creation and summarization, at 46%.

That is roughly where AI earns its place on an event team: drafting, summarizing, reformatting, first passes at research. This list covers the other side. Nine jobs that still need a person, why, and what to do about each.

1. Read the room while it is happening

AI can summarize a transcript afterward. It cannot notice the VP who went quiet after the pricing question, or the table where conversation died by the second course.

Organizers are already worse at this than they think. Freeman found that attendees who experience a "peak moment" are 85% more likely to return, yet only 40% of attendees report having one, while 78% of organizers believe they deliver them (Freeman, 2025). An AI summary of the agenda will not close that gap.

What to do: assign one person who is not the host to watch the room and take notes on energy, not content. Ask two or three guests directly, the next day, what the best moment of the evening was.

2. Know who is not speaking to whom

Buying groups are rarely harmonious. Gartner found 74% of B2B buyer teams show "unhealthy conflict" during the decision process (Gartner, 2025). Add the things no public source records: who just left whose company, which two guests lost a deal to each other last year, which customer is quietly evaluating a competitor.

That information lives with account executives and customer success managers. It does not live in a model.

What to do: send the draft guest list to every account owner with one question. Is there anyone here who should not be grouped together, and why?

3. Confirm where someone works today

Language models will produce a confident job title for almost anyone. It is often out of date. The CRM is not a reliable backstop either. In Validity's 2024 survey, 49% of CRM admins reported expired data and 48% said customer data had decayed faster over the prior 12 months (Validity, 2024).

What to do: ask for current title and company at registration. For senior guests, have a person confirm against the guest's own recent public profile in the week before the event. Never let a model's guess reach a name card.

4. Judge chemistry

Who sits next to whom decides whether a dinner produces conversation or small talk. Freeman found 51% of attendees prefer industry or topic-specific discussions as a networking format, and 54% are willing to share on-the-job challenges to improve networking (Freeman, 2025).

AI can group guests by industry and role. It cannot tell that two operators with the same problem and opposite solutions will produce the best conversation of the night, or that one talkative founder will flatten a table of junior staff.

What to do: let AI draft groupings by role and topic, then have the host rearrange them by hand.

5. Replace the host

The value of an in-person event is that a person is in it. In a 2017 field experiment with 45 participants, Vanessa Bohns found in-person requests were 34 times more effective than email requests, even though participants were equally confident in both (Bohns, Harvard Business Review, 2017). It is a small, older study, but the direction is hard to argue with.

Attendees notice who shows up for them. Freeman found 93% of attendees consider capable booth staff extremely important, but only 78% feel exhibitors deliver (Freeman, 2025).

What to do: spend the hours AI saves on host preparation. A host who has read a one-page brief on every guest will outperform any tool.

6. Clean the CRM on its own

Validity found 76% of organizations say less than half their CRM data is accurate and complete, and 57% moved to manual data cleaning while cutting dedicated data staff (Validity, 2025). Bain reports that in sales AI projects, sometimes as much as 80% of existing data has to be removed because it is old, inaccurate or confusing (Bain, 2025).

Validity's Cynthia Price described the pattern as organizations "layering AI on top without addressing foundation" (Validity, 2025). Validity sells data quality software, but the survey numbers make her case.

AI can propose merges and flag anomalies. It cannot know which of two conflicting records is true.

What to do: let AI suggest matches and corrections. A named person approves every merge, and the approval is logged.

7. Prove ROI from data that was never joined

Only 1 in 5 enterprises has integrated its primary event platform with the rest of its martech stack (Forrester, 2024). A Forrester Consulting study commissioned by Cvent found only 36% of organizations capture and integrate event data across all their events, and 71% say disconnected event tools limit visibility (Forrester Consulting for Cvent, 2026).

Ask a chatbot what your Q2 dinners returned, with no join between attendance and opportunities, and it will produce an answer anyway. That is the danger. A fluent paragraph about pipeline influence is not evidence of pipeline influence.

What to do: join attendance to CRM records first. Then ask AI to describe the joined table, never to estimate what is missing from it.

8. Be the source a buyer believes

Buyers use AI heavily and trust it selectively. Gartner found 45% of B2B buyers used generative AI during a recent purchase, and 69% prefer to validate AI-generated insights with a sales rep (Gartner, 2026). In a Censuswide survey for Cvent, 85% said AI-generated content has made trust harder to build, and 81% said audience trust rose after in-person events (Cvent, 2026). Both of those last figures come from vendor-commissioned research with undisclosed samples, so treat them as directional.

The event is where that validation happens, face to face, with someone who can answer a follow-up question.

What to do: put subject matter experts in the room, not only sellers. Freeman found 84% of attendees say connecting with subject matter experts is extremely or very important (Freeman, 2025).

9. Tell you when it is wrong

Models state errors in the same confident tone as facts. Buyers have noticed: 51% of B2B buyers say they are likely to meet misleading information from generative AI, slightly more than the 49% who say the same of sales reps (Gartner, 2026). Among B2B marketers using AI for content, 12% say quality got worse (Content Marketing Institute, 2025). And among go-to-market professionals who avoid AI, 80% cite accuracy concerns (ZoomInfo, 2025), a vendor survey, but a telling one.

What to do: every AI output that leaves the team has a named human checker. For anything containing a number, the checker traces each number back to its source row.

The pattern

Read the list again and a pattern shows up. AI is weakest where the information lives in someone's head, or in data that was never connected. Those are also the parts of the job that justify having an event team at all.

The PULSE survey does not say what separates the 16% who report significant improvement from everyone else. This list suggests where to look: the teams that gave AI the drafting and kept the room for themselves.