- A hands-on AI agent workshop for teams should end with each team owning one working agent, built in the AI tool the company already approved, such as ChatGPT or Claude, and tested on a task the team repeats every week.
- An AI agent is a model that uses tools in a loop to finish a job, not a product label. For most office teams the first agent is a saved assistant with written instructions, approved reference files and a defined output.
- The data rules come before the build: which tools are approved, which information stays out, and where a person must sign off before an agent's output is used or sent.
- Practice runs on synthetic or redacted material. A workshop agent should draft and check work; it should not send email, change records or touch live systems.
- The session ends with a decision for each agent: keep, change or stop, and who owns it.
What teams mean when they ask for an agent workshop
The request is usually concrete. A manager has a list of jobs the team wants an AI agent to take on, such as preparing meeting briefs, sorting incoming questions or tracking where assets are, and wants a session where people build the first one instead of hearing about agents in the abstract. That is the useful version of a hands-on AI agent workshop for teams: the team leaves with an agent that runs on its own work.
We have seen what makes this work in community rooms first. At the Dallas AI hackathon on April 18, 2026, 37 people signed up, and attendees worked in teams for one three-hour afternoon and presented what they had built to the room. At the May 17 agents meetup, the session walked through what an agent is and how several agents with clear roles can split up a complex task. The company workshop keeps the same rule: build something real in the session, then show it.
Agent, workflow, chat: the difference that matters
Most confusion in these sessions comes from the word "agent". In the definition Jack uses in his Agent 101 notes, an agent is a model that uses tools in a loop to finish a job. It is not a product label. Three levels are worth separating:
- Chat: one person types a request and reads one answer. Nothing is saved, so next week starts from zero.
- Workflow: a written set of steps with a defined input and output, which anyone on the team can rerun and check.
- Agent: a workflow saved inside the tool, with its own instructions, approved reference files and, where the tool allows, tools such as search or file reading. In ChatGPT this is usually a custom GPT or a project; in Claude, a Project.
For a non-technical team the first agent should be the third level, built on top of the second. If the team cannot write the workflow down, the agent will not behave the same way twice.
How the session runs
Our 90-minute AI workshop for teams follows the same order whether the goal is a workflow or a first agent:
- Data rules first. Agree which AI tools are approved, which information stays out, and where a person must sign off before an output is used. An agent makes this more important, not less, because it runs the same instructions every time.
- Choose the job. Each person lists the tasks they repeat every week and scores them: frequent enough to matter, clear input and output, quick to check, and safe to practise without sensitive data. The strongest one becomes the agent's job.
- Write the instructions as a spec. A prompt is a spec, not a spell. The agent gets its role, the input it will receive, the format it must return, what it must never do, and an acceptance test a person can apply in under a minute.
- Build it in the approved tool. Save the instructions and the approved reference files in the tool the company already uses, then run the agent on synthetic or redacted sample material.
- Break it on purpose. At least one test case is deliberately flawed: a missing field, a contradiction, a request outside the agent's job. Watching how the agent handles it shows the team what to review before anything goes out.
- Decide what happens next. Keep, change or stop the agent, name who owns it, and agree how it will be checked two weeks later.
If the team arrives with a longer list of agents it wants, the list is handled the same way: score the list, build the one or two strongest in the room, and leave the rest as written specs to build after the session.
Data safety for team agents
An agent built in a workshop drafts, sorts and checks. It should not send email, update customer records or connect to live business systems during training. Those actions need a person's approval and, often, an IT decision about integrations. Our workshops are AI training, not a CRM, ERP or automation build, so the agent works on the parts around those systems: preparing notes, summarising updates, checking exports.
Three answers should be written down before the agent is shared with the team: which tool it lives in, which information it may never receive, and who signs off on its output. The free AI data boundaries one-pager is a starting point.
When this does not work
It does not work without an approved tool. If people can only build agents in personal accounts, the session teaches them to keep company work there. Approve one tool first; the AI tool approval checklist is built for that decision.
It does not work when the job is vague. "An agent that handles our marketing" cannot be tested. "An agent that turns our approved event notes into a five-point recap in our format" can.
And it is the wrong session if what you need is software. If the agent has to read and write records in your systems on its own, that is an implementation project with IT, not a training workshop.
What to do in the next 14 days
- List the jobs your team would hand to an agent, and score each one on frequency, clear input and output, how fast it can be checked, and whether it can be practised without sensitive data.
- Write down the three answers: approved tool, off-limits information, and who signs off.
- Pick the one job with the highest score and gather two or three redacted examples of its input and a good output.
If you want your team to build its first agent together, see how the 90-minute team workshop runs, or book a 20-minute workflow call to talk through your list.
The approach in this post is what we teach in 90-minute AI workshop for teams.
Questions people ask
What is a hands-on AI agent workshop for teams?
A live session where a team builds and tests its first AI agent on its own work. It starts with data rules, picks one repeated task, writes the agent's instructions as a spec, builds it in the company's approved AI tool, tests it on synthetic or redacted material, and ends with a keep, change or stop decision.
Do we need developers to build an AI agent?
No. For most office teams the first agent is a saved assistant inside a tool the company already uses, such as a custom GPT in ChatGPT or a Project in Claude, with written instructions and approved reference files. No coding is required.
What is the difference between an AI agent and an AI workflow?
A workflow is a written set of steps with a defined input and output that anyone can rerun. An agent is a model that uses tools in a loop to finish a job; in practice, a workflow saved inside the AI tool with its own instructions and files. A team that cannot write the workflow down will get an agent that behaves differently each time.
Is it safe to use company data in an agent workshop?
Practice should use synthetic or redacted material, and only the AI tools the company has approved. No customer personal data, passwords or confidential documents should go into a tool during the workshop, and the agent should not send messages or change records without a person signing off.
How long is the workshop and how is it priced?
Choose a 90-minute team workshop, full-day onsite training, or 4 live 90-minute sessions. Every format is priced by a scoped quote. Group size, location and preparation are agreed in the written quote.