Here is a 45-minute lunch-and-learn plan on AI agents: five minutes to open with a real question, ten on what depth means, ten walking through one held action from the audience's own work, ten on a paired exercise deciding verdicts, seven for questions and three for next steps. Each block below has its timing, script notes and materials.
What is a lunch-and-learn for, and what can it not do?
A lunch-and-learn is a short, informal session held over a meal break, usually voluntary. The format is described in what a lunch-and-learn is. It is good at one thing: giving a broad audience a shared picture and a reason to want more.
It cannot certify anyone. Forty-five minutes will not make a person ready to approve an agent's held payment. Say so at the start, so nobody leaves thinking they are trained. Use the session to build vocabulary and curiosity, then point people to the training that proves competence, as set out in the guide to training teams for AI agents.
Run one when an agent is about to arrive in a team, when it gains a visible new capability, or when people outside the team keep asking what it does.
How do you prepare?
Pick one agent the audience will meet, and one real held action from its record, anonymised. Prepare three short cases for the exercise, each with a clear right answer: one the agent should be allowed to do, one that should be held for a person, and one the platform withholds for every agent at every depth. Bring a one-page handout with the depth scale and the four verdicts.
Use the audience's own work for every example. A finance team learns more from a held supplier payment than from a generic chatbot demo. The habits that make a session stick are covered in how to run a lunch-and-learn on AI agents that people remember.
The session plan
| Time | Block | What the presenter does | What the audience does | Materials |
|---|---|---|---|---|
| 0 to 5 | Open | Ask: "Who decides when the agent gets something wrong?" Take three answers. State what the session is and is not. | Answer aloud | None |
| 5 to 15 | Depth in plain words | Explain depth 1 read, 2 draft, 3 routine action within bounds, 4 money or hard-to-undo actions, 5 coordinating other agents. Explain that depth is earned through evidence, not switched on. | Listen; mark on handout where their own work sits | Handout |
| 15 to 25 | One held action, start to finish | Walk through the anonymised held action: what the agent drafted, which bound held it, who decided, what was recorded. | Ask questions as it goes | Screen or printout |
| 25 to 35 | Paired exercise | Hand out the three cases. Pairs decide: allowed, held, refused or withheld, and why. | Decide in pairs, then share | Case sheets |
| 35 to 42 | Questions | Take questions. Park any about specific people or incidents for a private follow-up. | Ask | None |
| 42 to 45 | Next steps | Name the training that follows, who it is for and when it starts. | Note the date | Sign-up link or list |
A worked example
An illustrative procurement team of 20 is about to work with an agent that drafts purchase orders at depth 2 and raises them within budget at depth 3.
- Open: "If the agent raises an order to the wrong supplier, whose name is on it?"
- Depth block: the presenter shows that drafting orders is depth 2, raising them within budget is depth 3, and paying suppliers sits at depth 4 with finance, not procurement.
- Held action: an order the agent held because the supplier was new and its details came from an emailed quote, an external source.
- Exercise cases: (a) a repeat order to an approved supplier within budget: allowed; (b) an order 10 above the budget line: held; (c) a request to reverse a payment already settled on the organisation's own rail: withheld for every agent.
- Next steps: reviewers on the held rota begin their certified training in two weeks; everyone else gets the shared foundation.
Some teams record the session for people on shift. A recording helps with reach but loses the paired exercise; see what a recorded session is for how to handle that.
What should follow the session?
For reviewers, structured training and assessed decisions. For the team as a whole, regular calibration once the agent is live, using the calibration session agenda template. Much of the rest happens at the desk, which is why the question of whether staff can learn to supervise AI agents on the job matters for planning what comes next.
Questions people ask
Should attendance be compulsory?
No. Compulsory sessions over lunch breed resentment. Make it voluntary, make it useful, and make the formal training that follows compulsory for those who need it.
Who should present?
Someone from the team who has used the agent, with the agent owner on hand for hard questions. An outside presenter can explain concepts, but cannot answer "what happens to my queue?".
How many people is too many?
Beyond about 30, the paired exercise becomes noisy and questions dry up. Run two sessions instead of one large one.