Getting executive teams hands-on with AI

Our first AI session for a leadership audience ran on 2nd February 2023, nine weeks after ChatGPT launched: a hands-on workshop at a global consumer goods brand's international leadership conference, with executives generating campaign images for their own products using tools most of the room had never opened. This month the same company came back to start planning the next one.

We have run these sessions continuously since: for an international strategy consulting firm, for boards, for private equity firms, for broadcasters, for founders. The requests increasingly sound the same: "our leadership team gathers once a year; how do we use that time on AI well?"

This page is our answer. Use it to plan your own. And if we can help, we will run it with you.

Start with the goal, not the agenda

The most useful question in planning one of these days is about purpose: what do we actually want to achieve? Most sessions are designed backwards, agenda first. The agenda is the last thing to decide.

A leadership session can do six jobs. Pick two. A day that tries to do all six does none.

Then write the agenda as questions the group must answer rather than topics to cover. Steven Rogelberg (Chancellor's Professor, University of North Carolina at Charlotte), whose meeting research draws on surveys of more than 5,000 employees, finds a topic agenda invites drift while a question agenda forces a decision. It also gives you a scoreboard: at the end of the day, either the questions got answered or they didn't.

Know which room you are in

Every organisation we have worked with shows the same split. A small group does most of the innovating; the rest use AI like a slightly better search box.

That split gives you three different rooms, and they need different days. The pioneers need fuel: advanced material, competition, each other. The majority need nudges, foundations and proof from peers they recognise. The executive team needs personal fluency and decisions. Mixing them is the most common design mistake there is: content pitched at everyone lands with no one.

One more reason to take the audience question seriously: leaders reliably misjudge it. McKinsey's "Superagency" research (published January 2025, fielded late 2024) found C-suite leaders estimated 4% of employees used AI for at least 30% of their daily work; the employees themselves reported 13%, three times higher. The gap the day needs to close is often the leadership's own.

What works really well

• Hands-on beats watching, and the evidence agrees. Harvard Business Publishing and Degreed (2025) found people fluent with AI were twice as likely as everyone else to have learned by experimenting. Design for at least half the time spent doing. The strongest sessions we have run put a tool, a real task and 90 minutes in front of every person in the room.
• Their business, never a canned demo. Generic demonstrations show technology without context, ownership or stakes. At one firm's annual partner conference, tables of six built genuine client pitches from scratch in a morning: research, simulated client interviews, a finished proposal. The room was loud, competitive and surprised at what it could do. It worked because the work was real.
• Peers on stage beat outside experts. A short panel of colleagues showing real work they have already shipped grounds the day in proof rather than promise, and lands harder than any outside keynote. Moderna built the same principle into its rollout: a prompt contest surfaced its top 100 power users, who became a network of internal AI champions, with office hours in every business line. Its own chatbot, mChat, launched at the start of 2023 and passed 80% adoption, and Moderna set itself a six-month target of getting everyone proficient (OpenAI's case study on Moderna). The outside expert's best work happens backstage: finding the right presenters and coaching them, because smart people still flop unprepped.
• Nobody starts cold. Before the day, everyone gets set up and completes one small piece of homework that is submitted, machine-graded, and returned with personal feedback. The artefact matters less than the act: everyone walks in having done one real thing, from the same shared floor. And pre-work needs teeth. At one large conference we ran, where it was optional, barely one person in seven did it.
• A power user on every table. Identified in advance, ideally from usage data, and seated deliberately, so every table has one person who can already do the thing. In our experience the single biggest driver of a good room.
• Mix the room on purpose. The strongest research on offsites (Madeline Kneeland, Heidi Gardner and Adam Kleinbaum, Strategic Management Journal, 2025) found their durable payoff is rewired working relationships: new ties, new trust, a better map of who knows what. And only the people in the room get it. Seat across functions, protect the informal time, and choose the invite list like it matters, because it does.
• Structure who speaks. Hierarchy suppresses candour, even among the most senior. In Amy Edmondson's survey of 805 senior leaders (Harvard Business School), just 7% believed their own silence at work had had no impact. Small groups, round-robins and rotating facilitation surface more truth than an open floor ever will.
• Competition, with locked rules. Tables submit real work; it is judged; there are prizes across categories. AI can do the judging at scale, against criteria agreed with your leadership in advance, so the ranking reflects what the business actually values.
• Leadership opens, and stays. A few minutes from the chief executive on why this matters sets the whole room. McKinsey's State of AI (November 2025) found high performers were three times more likely to have senior leaders who visibly own and commit to AI, not just approve the budget. The room notices who leaves after the welcome.
• End with something written, and someone waiting. Every session we run ends with a report: what we saw in the room, what it says about where you are, recommendations and next steps. At conference scale, add commitment pairs: each leader swaps specific commitments with a cross-function partner and books a 15-minute check-in a month out (a mechanic recommended by the organisational psychologist David Burkus). The follow-through architecture is the deliverable; the day is how you build it.

What doesn't work

• Mixed rooms. One session for the pioneers, the majority and the board at once. See section II: pick a room.
• Death by deck, and a cast of thousands. Presentation marathons crowd out the conflict and the decisions the gathering exists for (Doug Sundheim, Forbes, 2025). Oversized invite lists do the same: entourages and observers turn a working session into a town meeting. Ban status updates, cap the room.
• The offsite high. Inspiration with no decisions, no owners and no dates evaporates within days. If nothing gets decided that could have been decided by email, the gathering was theatre.
• Leaving the plumbing to chance. The technology needs the same careful preparation as the content: hundreds of people downloading, logging in and submitting at once is a real load test. Pre-stage everything in small files, test on the venue's connection, run a live health check on your tools minutes before doors open, and agree a plan B for every dependency. Win the plumbing battle before the day, so the room spends its energy on judgement instead of downloads.
• An ending that fizzles. If the results reveal cannot physically happen before the room breaks up, the emotional payoff of the whole competition dies. The science backs the instinct: Daniel Kahneman's peak-end rule shows people judge an experience by its most intense moment and by how it closes. Size the judging to the break. Protect the ending.
• Changing the rules mid-game. Adjusting the scoring on the fly, however kindly meant, blurs the lesson for everyone. Lock the rules in advance and let them play out.
• Teaching speed without judgement. AI's first drafts are mostly impressive, and that is exactly what makes the weak minority dangerous: it arrives with the same confidence as the rest. Reviewing and improving the output is a skill in its own right, so build an explicit Check, Edit, Own beat into the run of show, and consider a red-team step where the model is asked to attack its own draft.
• The relapse. People revert to old habits within days unless managers hold the line and help is on call. Gallup (February 2026) puts frequent AI use at 78% among employees who strongly agree their manager actively supports it, and 44% among everyone else. Spaced reinforcement is what makes learning stick, so book the booster sessions before the day itself. A great day that ladders into nothing is a great day wasted.

Six ways the day can work

Formats we run, in rising order of ambition. Every one ends with the written report. Most rooms start from our signature keynote.

1. The conference keynote

   20 to 40 minutes · any size room · job: believe

   The keynote above, applied to your context, with live demonstrations. Belief change at scale, and a shared reference point for everything that follows.

2. The peer showcase

   60 to 90 minutes · any size room · job: believe + signal

   Your own people on stage, showing real work they have already shipped and what it changed. We find the presenters, coach them on content and delivery beforehand, and compere on the day.

   Proof from a colleague beats promise from a consultant, every time.

3. The structured debate

   45 to 60 minutes · any size room · job: believe + align

   A motion with teeth ("AI will replace us"), two teams of two or three arguing assigned positions, timed rounds, audience questions voted up live, and the same poll before and after. The winner is whoever moves the room. We did not invent this format; debating societies have run it for centuries. We have run it at conferences since 2018, and it is the best way we know to surface the issues and challenges a room actually holds, because assigned positions give people permission to make the uncomfortable argument properly.

   The first time we ran it, a room that began 77% against the motion ended roughly even. The side arguing for AI lost the question, and won the debate.

4. The executive working session

   Two to three hours · the executive team only · job: align + unblock

   Personal fluency first: every leader works with the tools on their own real work, starting with the simplest exercise we know (ask the model: "what else do you need to know about me to help me well?"). Then the team questions: what this means for us, what we will refuse, and what gets unblocked today. We arrive with a straw model to react to, because senior groups refine a draft well and stall at a blank flip chart (Bob Frisch and Logan Chandler, Harvard Business Review). Judgement over technical depth throughout, which is the same principle every major business school's executive AI programme has settled on.

5. The hackathon with an AI judge

   Half a day · 20 to 300+ people · job: accelerate + upskill

   Tracks by function. Each group gets a real business challenge and real data, and two hours to build. Submissions are judged by an AI panel against criteria agreed with your leadership in advance, so the ranking reflects what the business actually values. Locked rules, prizes, and a results reveal in the room.

6. The full immersion day

   One day · your most active users, with executive sponsorship in the room · job: accelerate + signal

   Keynote, peer showcase, hackathon, and an executive working session running in parallel. The pioneers get fuel, the leadership gets fluency and decisions, and the rest of the organisation hears about it, which is part of the point.

Where this thinking comes from

Most of the ideas on this page appeared first in David's Saturday AI Thoughts, the weekly email, where they were tested against a few hundred senior readers before they were ever run in a room. A sample, if you want the thinking behind the design choices:

Read the archive or join the Saturday email here.