You have sat through this panel. Four people in matching chairs, a moderator reading questions off a card, forty-five minutes of polite agreement. Someone says it is still early days. Someone else says it is really about the humans. The audience checks their phones, and the only memorable moment is when a mic cuts out.
Panels fail for boring, fixable reasons. Every panelist was picked for being safe, so there is no tension. The moderator treats the job as timekeeping, so there is no direction. And four people who broadly agree have nothing to push against, so there is no content. AI panels are especially prone to it: the topic is hot, everybody has an opinion, and organisers assume that booking smart people is enough. It is not. A panel is a designed thing, and most of them are not designed at all.
Use a panel only when the disagreement is the point
Default to a keynote when one person has a coherent argument the room needs to hear end to end. A panel is a bad way to deliver a thesis. It is a good way to show a collision that no single speaker can represent honestly.
Book a panel when the question is genuinely contested and you want the audience to see the disagreement rather than a tidy resolution. Or when the right answer depends on context: a healthcare security lead and an early-stage founder will read the same AI regulation completely differently, and that gap is the content. Or when you want practitioners comparing notes on something each of them has actually done.
Do not book a panel because you have one strong speaker and want to give more people stage time. Stage time is not a prize you hand out. The moment you add a fourth person for political reasons, you are already building the bad version.
Compose for productive disagreement
The composition is the panel. Get it wrong and no moderator can rescue you. What you are engineering for is substantive, informed conflict — not personality clashes, not manufactured hot takes.
Spread across three axes at once.
Viewpoint. Put a genuine sceptic next to a genuine believer. If everyone on your list thinks agents will reshape their industry within a year, find the person who thinks that is nonsense and can defend the position. The sceptic is not a foil. They are there because a large part of your audience quietly agrees with them and has never heard the case made well.
Sector or vantage point. Mix where the work happens. A researcher, an enterprise buyer and an operator who has shipped to real users will disagree about what "ready" means, and those three definitions are the session.
Seniority. An executive and a hands-on engineer see different truths. One talks strategy, the other knows what breaks at 2 a.m. Both are right, and the contrast is what people remember.
Three rules save more panels than anything else. Three or four people, never more. No two panelists who would answer every question identically — if you have a duplicate, cut one. And screen specifically for people who can disagree well, meaning they can say "I think that is wrong, and here is why" without making it personal. Some excellent speakers cannot do this. Find out before you book, by asking them directly how they handle being contradicted on stage.
When we build shortlists for panels and moderators, we spread candidates across those axes deliberately rather than handing over four versions of the same voice, because a uniform shortlist produces a uniform panel. Every name is verified individually first — the standard is the same one in our booking guide: what have you shipped, what went wrong, who will vouch for you. Describe the tension you want on stage in a written brief and we come back within 24 hours; you can do that here without an account or a call.
The moderator is the product
If composition is the raw material, the moderator turns it into a session. This is the role organisers most often underspend on and the one that most determines the outcome.
A strong moderator does five specific things. They come in with a thesis, so the conversation has a spine instead of wandering question to question. They notice disagreement and pull on it: "you two just contradicted each other, say more." They cut off the rambler without rudeness and bring in the panelist being talked over. They ask the follow-up the audience is thinking, which is almost always "concretely, what did you actually do?" And they know the material well enough to recognise an answer that is too convenient.
Vetting one is different from vetting a speaker. A reel of solo keynotes tells you very little, because moderating and presenting are separate skills. Ask for video of them running an actual panel. Get on a call and watch how they handle you: do they listen and redirect, or wait for their turn to talk? Ask how they would deal with a panelist who dominates, and with two who agree on everything. Someone who has done the job will answer immediately, because they have lived both. Someone who has not will improvise something vague.
A note on roles, since they get conflated. A moderator runs one session. An emcee runs the day: transitions, energy, housekeeping, recovering the room after a session runs long. They are different bookings and often different people, which is why we list emcees and hosts separately. Asking your emcee to also moderate the hardest panel of the day is how you get a mediocre version of both.
Prep that actually changes the outcome
Most panel disasters are briefing failures. People arrive cold and default to platitudes because nobody told them what the session was for.
Send a written brief a week out. It needs four things: the thesis and the disagreement you are after, with the fault lines named; the rough question arc, so people know the terrain without memorising lines; who else is on the panel and what angle each brings; and hard limits on time and self-promotion, stated plainly. Telling panelists you want them to push back changes their behaviour more than any other single intervention.
Brief the moderator separately and in more depth. They need the full plan, the landmines, and explicit permission to interrupt. A twenty-minute pre-call between moderator and panelists — not a rehearsal, just a conversation — surfaces where the real disagreement lives so the moderator can go straight to it on the day.
Design the running order on purpose too. Open with the sharp framing, not with introductions; bios belong in the programme, and the moderator should land the central tension inside ninety seconds. Write tight questions and plan two layers of follow-up for each, because the first question is just the door. Decide in advance how many topics you can genuinely cover, since three done well beats seven rushed. Reserve the last ten to fifteen minutes for the audience, with the moderator filtering so you get questions rather than "more of a comment than a question."
Failure modes and the fix
| What goes wrong | Root cause | Where you fix it |
|---|---|---|
| Everyone agrees | Composition without contrast | At booking, not on stage |
| Serial monologues | Passive moderator | Brief the moderator to interrupt and cross-examine |
| Someone keeps pitching | No stated limit | In the written brief, then live redirection |
| Six voices, no depth | Slots handed out politically | Cut the cast to three or four |
| Best moment at minute 43 | Structure front-loads throat-clearing | Open with the tension |
| Dead audience Q&A | Unfiltered open mic | Curate through a moderated queue |
Room and timing mechanics
The physical setup does more work than people expect. Chairs angled slightly toward each other rather than in a straight line facing out makes panelists talk to each other instead of taking turns addressing the audience. Every panelist needs their own mic; a shared handheld kills interruption, and interruption is where panels come alive. Put the moderator at one end where they can see every face without turning their back on the room.
On timing: forty minutes is usually right for four voices, fifty if you have a strong moderator and a contested topic. Beyond that, energy drops regardless of how good the cast is. Panels work better mid-morning or mid-afternoon than immediately after lunch, and they are a poor closing session because they rarely resolve — which is the point, but a room wants to leave on something conclusive.
Virtual magnifies every weakness. Latency destroys natural interruption, so the moderator has to direct traffic explicitly by name. Keep remote panels to three voices and shorter runtime, and test every connection in advance, because one frozen video takes the segment down with it. The same production discipline applies as for any remote session; the mechanics are in the virtual keynote guide, and the wider agenda question is covered in the 2026 conference planning playbook. For hybrid, name one person whose only job is bringing remote participants in, or they will simply disappear from the conversation.
Fix the cast first. A brilliant moderator with four people who agree still has nothing to work with, but a competent moderator with genuine disagreement in the room will get you a session people argue about in the hallway afterwards.