Search "AI keynote speaker" and you will get thousands of people who describe themselves as one. A meaningful share of them have never deployed a model, never sat in a governance meeting, never had to explain to a board why the pilot did not scale. They have read the same reports you have and turned them into a talk.
You cannot easily tell which is which from a bio, because bios are written to be indistinguishable. Everyone is a leading voice. Everyone advises Fortune 500 companies. Everyone has been featured in a publication you have heard of.
Anyone can claim to be an AI expert. We check. This is what checking means, in enough detail that you can hold us to it, or run the same process yourself if you would rather.
Why the claim needs checking at all
The gap between knowing about a subject and having done it used to be visible. It is not any more, because the raw material is now free. A capable person with no operational experience can assemble a genuinely well-argued forty minutes on AI governance in a week, with current examples and confident delivery.
That talk will not survive contact with your audience. Not because it is wrong, but because the moment somebody in row three asks what happens when legal refuses to sign off on the training data, the answer will be a principle instead of a story. Your technical people will notice within ninety seconds, and once they have decided the speaker is narrating, the room is gone.
So verification is not really about catching liars. Most people are not lying. It is about establishing whether the knowledge was earned by doing or assembled by reading, because those two things behave completely differently under pressure.
What we check, and what fails it
| What we check | How we check it | What fails |
|---|---|---|
| Identity and the claimed role | Direct confirmation of the specific role and dates, not a scrape of a profile page | Titles that inflate on the way to the bio, dates that stretch |
| Whether the AI work is first-hand | Detailed questioning about one project: their decisions, their constraints, what went wrong | Fluent answers with no specifics, credit for a team's work, "we" that never becomes "I" |
| Stage skill | Unedited video of a full talk, plus footage of Q&A wherever it exists | Only a showreel; only a studio recording; no live footage at all |
| Audience effect | References from people who were in the room | References who can only speak to the person as a colleague or client |
| Currency of the material | What has changed in the talk in the last twelve months, and why | A deck whose examples predate the current generation of tools |
| Fee reality | What they have actually been paid recently, for what kind of event | An aspirational number with no bookings behind it |
| Format honesty | Whether they can hold a camera, run a room, or moderate, asked separately | Claiming all three without evidence for any |
The rest of this is what sits behind each row.
The claimed work
We start with the sentence in the bio that is doing the most work, and we go at that one.
If someone says they led AI adoption at a named company, we want to know what "led" meant. Whose budget, how many people, which function, what the mandate was, and what the state of it was when they left. Roles inflate in transit, and they inflate most in the gap between what someone did and what a bio needs them to have done to justify a fee.
This is not adversarial. Most people answer immediately and in detail, because most people are describing something they genuinely did and are pleased to be asked about it. The tell is not defensiveness. It is vagueness that stays vague when you ask a second time.
First-hand or narrated
This is the check that matters most, and it is the one legacy representation almost never performs, because an agent's confidence in a speaker is built from a personal relationship rather than from evidence. We wrote about why that structure produces the result it does in what is wrong with legacy speaker bureaus.
The method is simple. We pick one project the speaker mentions and go deep on it, for as long as it takes.
What was the constraint you did not expect. Who objected, and what was their actual argument. What did you get wrong in the first three months. What did you have to tell someone senior that they did not want to hear. What would you do differently.
People who did the work answer these in specifics, at length, and often with visible irritation at the memory. They tell you about the vendor who overpromised, the data that was worse than anyone admitted, the team that quietly kept using the old process for five months. They volunteer the part that was their fault, because that is the part they think about.
People who read about the work answer in frameworks. They restate the question at a higher level of abstraction. They are not being dishonest; they simply do not have the material, because you cannot read your way to the memory of a bad Tuesday.
Ask what broke. It is the most efficient question in this entire process, and you can use it in a fifteen-minute call without a bureau's help. It is one of the test questions in our guide to the ten kinds of AI speaker.
Video, and specifically the unglamorous kind
We want to watch a full talk, start to finish, from a real event. Not a showreel. A showreel proves that somebody can edit.
What we are watching for is the parts a reel removes. How they open before the room has decided about them. Whether the energy holds at minute twenty-eight when the argument is doing work rather than landing a line. What their hands do when a slide fails. Whether they are speaking to the room or reciting at it.
Q&A footage is worth more than the talk. A prepared forty minutes can be drilled. An unfriendly question cannot, and the response to one tells you almost everything: whether they engage with the actual objection, whether they can say "I don't know", whether they get smaller when challenged.
If no live footage exists at all, that is not automatically disqualifying, because genuinely good practitioners are sometimes early in a speaking career and we would rather find them before everyone else does. It does mean the next stage is a live session with us, and it means we tell you plainly that this speaker is new to the stage rather than letting a polished bio imply otherwise.
References who were actually in the room
Most reference checks are useless because they ask the wrong people. A former colleague can confirm that someone is smart and pleasant to work with. That was never in doubt.
We ask for people who watched them speak. An event organizer, an L&D lead, someone who sat in the fourth row. Then we ask questions that cannot be answered with a compliment:
- What did people say about it in the following week, if anything
- Was there a moment where the room lost interest, and where was it
- How did the Q&A go, especially the hardest question
- What did you have to fix or chase during the run-up
- Would you book them again for a bigger audience, and if not, why not
That last one produces the most useful pause in the whole process. Plenty of speakers are excellent for forty people and unconvincing for eight hundred, and the person who saw both knows.
Fee reality
We ask what a speaker has recently been paid, for what size and kind of event, and we compare it against what they want to be quoted at.
There is usually a gap and the gap is usually fine. Fees rise, and a speaker whose profile is genuinely growing should be asking for more than they got last year. What we are looking for is a number with no history behind it at all, because a quote invented from a competitor's website leads to a bad conversation with you and a booking that falls apart at contract stage.
Where we land, we put in front of you: the speaker's fee and our 17.5% commission on the same line, not folded together. The bands and what moves them are public in our 2026 speaker fee guide.
What disqualifies someone
Being early does not. Not having a book does not. Never having spoken to two thousand people does not, as long as we tell you that.
What does:
- A role or credential that turns out to be materially different from the claim
- AI experience that is entirely narrated, when the profile implies it is operational
- References who cannot confirm having seen them speak, or who hedge when asked the last question
- Core material that has not been meaningfully revised in over a year, in a field that changes quarterly
- Any dishonesty during the process itself, however small, because the whole product is our word
We do not publish a list of people we turned down and we are not going to. The point of the standard is that it is applied, not that it is performed.
Where we are honest about the limits
We launched in June 2026. We are early, and this process is done by hand, one speaker at a time, mostly by me. There is no automated verification engine behind it and I am not going to imply there is one.
That has real consequences. It is slower than a database lookup. It means the roster grows at the speed of actual conversations rather than sign-ups. It is why there is no public browse page listing names yet: we will not put a name in front of you before we have finished checking it, and checking takes as long as it takes.
The upside of being this size is that the checking is real. Nobody gets waved through because a partner vouched for them or because the quarter needs bookings. If that changes as we grow, the process will be written down and staffed rather than diluted, and this page will be updated to say what it actually is.
What you get on the other side
A shortlist where every name has been through the above, sent within 24 hours of your brief, with no account and no discovery call required first. What we verified about each person, and what we could not, comes with it. That whole flow is on how it works.
Behind it sit two promises that only make sense if the verification is genuine, because they are what we pay if it is not. If a speaker cancels, we replace them within 72 hours or you pay nothing. If your audience rates them below 8 out of 10, our commission comes back. Both are set out on the guarantee page.
And if you are a practitioner reading this and thinking that your own experience would survive these questions comfortably, that is exactly who we are looking for. You do not need an agent, a book or an introduction to start. You can apply directly, and you will be asked all of it. When you are the buyer instead, send us four things about your event and we will do the checking before the names reach you.