Someone has told you to book an AI speaker. Depending on who told you, they meant one of four completely different things, and the words they used were probably interchangeable.
Generative AI. Agentic AI. Responsible AI. AI governance. In a brief these arrive as synonyms. In practice they are four distinct subjects, aimed at four different parts of your organisation, requiring four different kinds of proof from the person on stage.
Book the wrong one and the room is polite and unmoved. This page exists so you can work out which brief you actually have before you send it to anyone.
The four briefs at a glance
| Brief | What it actually means | Who in the org cares most | What the speaker must prove |
|---|---|---|---|
| Generative AI | Systems that produce output a human then uses | Everyone doing knowledge work; L&D; marketing; operations | They have changed how real work gets done, at scale, and measured it |
| Agentic AI | Systems that take actions across tools and workflows | CIO, ops, product, engineering, internal audit | They have shipped something with permissions, failure modes and a rollback |
| Responsible AI | What the system does to the people it affects | Legal, HR, DEI, comms, regulated business lines | They have stopped or changed a system and can describe the cost |
| AI governance | The machinery that makes standards hold | Risk, compliance, procurement, board, general counsel | They have run or been audited by an actual control framework |
Everything below is the long version of that table.
What a generative AI keynote speaker actually covers
Generative AI is the part your audience has already touched. Drafting, summarising, coding, translating, generating images and analysis. The system produces something and a person decides what happens next.
That last clause is why the useful talk here is not about the technology. It is about the human system around it. Which tasks genuinely get faster. Which get worse in ways that are hard to see for six months. What happens to quality control when everyone's first draft is competent. How junior people learn a craft when the entry-level version of the work has been automated.
A good generative AI keynote speaker in 2026 should be able to name specific tasks where adoption failed and explain why, because everyone has those and almost nobody talks about them. If the whole talk is upside, you are getting a 2023 deck with new screenshots.
Book this when: your audience is broad, the questions are about daily work, and the goal is a shared and realistic picture. This is the most common brief for all-staff sessions and it is often better served by a workshop than a keynote, because the value is in people trying it rather than watching someone describe it.
The test question: "What did your organisation try that did not work, and how long did it take you to admit it?"
What an agentic AI keynote speaker actually covers
Agentic AI is the shift from output to action. Instead of drafting an email for you, the system sends it. Instead of summarising a ticket, it updates the record, contacts the supplier and books the engineer.
That changes the problem completely, and this is the bit that gets lost when the two words are treated as fashion labels. Once software acts, you inherit a set of questions that generative AI never forced you to answer. What is this thing permitted to do. What is it permitted to spend. What happens when step four of a seven-step chain is confidently wrong and steps five through seven proceed anyway. Who is accountable. How do you roll it back. How do you even know it happened.
An agentic AI keynote speaker who cannot talk fluently about permissions, evaluation, failure modes, human checkpoints and rollback is describing a demo. A great many talks currently sold as agentic are chatbot talks with a new vocabulary, and the difference is audible inside five minutes if you ask about production.
Book this when: your organisation is already funding projects where systems take actions, or your executives are about to approve some. The audience is usually narrower and more technical, which makes a panel or a moderated session a strong format — practitioners argue productively about this in a way they will not from a lectern.
The test question: "What is the worst thing your agent did in production, and what stopped it the second time?"
What a responsible AI keynote speaker actually covers
Responsible AI is about effect on people. Does this system treat people fairly. Can a decision be explained to the person it affected. Who is harmed if it is wrong, and are they the same people who are always harmed when things go wrong. Was consent meaningful. Is the training data something you would be comfortable explaining publicly.
This subject attracts more commentary than practice, which is the thing to watch for. There is an enormous supply of people who can deliver a principled, well-argued forty minutes on AI ethics without ever having had to make one of these calls with a deadline and a stakeholder who disagreed.
The distinguishing evidence is specific and it is always the same: a responsible AI keynote speaker who has done the work can tell you about a time they said no. What the system was, who pushed back, what the argument was, what it cost them internally, and what happened afterwards. Sometimes the answer is that they lost. That is a better answer than a clean one.
Book this when: you are deploying something that affects customers, employees or the public, and you want the room to take the consequences seriously before the launch rather than after. Also when a sector body or regulator is putting your industry under scrutiny.
The test question: "Tell me about the time you said no. What did it cost you?"
Why an AI governance keynote speaker is a different person
Responsible AI is the standard. Governance is the machinery. They are related the way a safety policy and a safety inspection are related, and confusing them is the single most common briefing error we see.
An AI governance keynote speaker talks about inventories of models nobody knew existed, approval gates that people route around, procurement clauses that do more than a policy document ever will, audit trails, model documentation, named accountable owners, and what happens when a regulator asks for evidence and you have a slide deck instead.
It is less inspiring than an ethics talk and considerably more useful to a risk committee. It is also the fastest-growing brief we get asked about, and the thinnest part of the market, because most people with real governance experience are busy doing it rather than speaking about it. Our governance and policy briefing covers where enforcement has actually moved.
Book this when: the audience is risk, compliance, procurement, internal audit or the board, and the outcome you want is a control that exists next quarter rather than a conviction.
The test question: "Show me what an approval gate looks like on a Tuesday when a business unit is in a hurry."
How to tell which one you actually need
Work backwards from the room's unanswered question.
| If the room is asking | You need |
|---|---|
| "How do I use this in my own job without producing rubbish?" | Generative |
| "What do we let it do on its own, and who is accountable?" | Agentic |
| "Is this fair, and can we defend it to the person it affected?" | Responsible |
| "Who signed this off, and where is the evidence?" | Governance |
Three more shortcuts. If the person who asked you to book a speaker sits in Legal or Risk, the brief is almost always governance even when they said ethics. If it came from the CIO, it is usually agentic even when they said generative. If it came from HR or internal comms, it is generative and the real subject is anxiety, which is a workplace psychology brief as much as a technology one.
When you need two, and how to run it
Some events genuinely need more than one of these, and forcing them into one speaker produces forty minutes of nothing in particular.
The pairing that works most often is a practitioner and a governance voice on the same agenda: someone who has shipped, followed by someone who has had to control what was shipped. They will disagree in front of your audience, which is the point. Structure it as a keynote plus a moderated conversation rather than two keynotes — the reasoning is in our panel and moderator guide.
If budget only stretches to one, choose by outcome, not by breadth. A speaker who covers all four covers none of them at the depth your room will remember. A narrow one gives you a decision. The fee bands for specialists are usually kinder too, as set out in the 2026 fee guide.
What we do with this when you brief us
We ask which of the four you are in, and if you are not sure, we ask about the room instead. Then every name that comes back has been checked against that specific brief, not against a general AI label, using the process in how we verify speakers. Anyone can claim to be an AI expert. We check which kind.
No account, no discovery call first, and the speaker's fee and our 17.5% commission printed on the same line. Four questions is all it takes to get a shortlist, and it comes back within 24 hours.