Skip the all-hands demo. A team learns to use AI when each role gets one workflow to practice and one weekly artifact to produce, not when everyone watches the same forty-minute tool tour and goes back to their inbox. The demo teaches nobody a habit. The habit is the point. Give the salesperson a workflow (draft the follow-up email from call notes) and a weekly artifact (five sent drafts, saved to a shared folder). Give the bookkeeper a workflow (flag anomalies in this week's transactions) and a weekly artifact (the flag list, with a human decision next to each). The artifact is the proof: it either exists at the end of the week or it does not, and that binary tells you more than any survey about whether the training stuck.
The all-hands demo fails because it has no owner and no output
The all-hands demo fails on a mechanism, not on effort. A room of twelve people watches one person type a clever prompt into ChatGPT, everyone nods, and then Monday arrives with no changed step in anyone's actual job. Nobody owns a specific task. Nobody produces anything by Friday that they did not produce before. The knowledge decays because it was never attached to a workflow that repeats. You can measure this yourself: run a demo, then check the tool's usage logs two weeks later. Most seats never log in again. The demo felt productive and taught a habit to no one.
Contrast that with how anyone actually learned a spreadsheet. Nobody learned Excel from a lecture on Excel. They learned it because they had one recurring job (the monthly invoice reconciliation, the quarterly headcount roll-up) and they did that one job in the tool, badly at first, until it got fast. AI is the same category of skill. It sticks to a task, not to a training slot. So the design question is not "how do we teach the team AI," it is "which one task does each role repeat weekly, and how do we move that task into an AI-assisted workflow they own."
Assign one workflow per role, chosen from work they already repeat
Pick the workflow from the job the person already does every week, not from a list of impressive AI use cases. The test for a good starter workflow is boring on purpose: it recurs weekly or more often, it takes real time today, and its output is text or structured data (which is what current models handle well). A support lead's ticket-tagging and weekly-theme summary qualifies. A marketer's turning one long blog post into five social posts qualifies. A recruiter's screening resumes against a rubric qualifies. "Have AI rethink our five-year strategy" does not qualify, because it does not repeat and you cannot practice it.
Here is a starter map you can adapt in an afternoon:
- Sales: draft each follow-up email from the call notes. Artifact: the week's sent drafts in a shared folder. - Support: cluster the week's tickets into themes and draft the top three canned replies. Artifact: the theme summary plus the three replies. - Marketing: turn one core piece into a week of channel posts. Artifact: the posts, scheduled. - Bookkeeping and ops: flag anomalies in this week's transactions or inventory. Artifact: the flag list with a human decision beside each flag. - Recruiting and HR: screen inbound resumes against a written rubric. Artifact: the ranked shortlist plus the one-line reason per candidate. - Owner or GM: summarize the week's numbers into a five-line update for the team. Artifact: the update itself, sent.
Notice what every row shares. One task, one tool session, one thing that exists on Friday that did not exist on Monday. The role owns the workflow, which means the person, not a central "AI team," decides how the prompt evolves and where the output lands. Ownership is why it survives past the training week.
The weekly artifact is the whole enforcement mechanism
The weekly artifact does the work that willpower and good intentions cannot. Learning a new tool competes against every established habit in the person's day, and the established habit wins by default because it is faster this week. A required artifact changes the default. When the Friday check-in expects five AI-drafted follow-ups in the shared folder, the person opens the tool on Tuesday instead of promising themselves they will "get to it." The artifact is a forcing function, and it is a visible one, which matters because it removes the ambiguity that kills most adoption efforts.
The artifact also exposes failure honestly, which a survey never will. Ask someone "are you using AI more?" and they will say yes, because they want to be a good sport and because they genuinely intend to. Look at the folder and count the drafts and you get the truth: three drafts this week, zero the week before, all of them edited heavily before sending. Now you know something actionable. Heavy editing means the prompt or the source material is weak, so you fix the prompt. Zero drafts means the workflow does not fit the real job, so you change the workflow. The artifact converts a vague sense of "adoption is slow" into a specific defect you can repair.
Keep the artifact small enough that producing it is never the hard part. If the weekly artifact takes two hours to assemble, you have built a reporting chore, and the chore will be the first thing dropped when the week gets busy. Five drafts saved to a folder takes seconds beyond the work itself, because the work is the drafts. The artifact should be a byproduct of doing the workflow, never a separate deliverable. When the artifact and the work are the same object, honesty is free and skipping is obvious.
Run it as a four-week cycle, then let each role own its workflow
Structure the first month as a short cycle with a clear end, not an open-ended program that quietly loses momentum. Week one: each person sets up their one workflow with a real example from their actual queue, and produces the first artifact by Friday. Week two: they run it solo and bring their artifact plus one thing that broke to a fifteen-minute team check-in. Week three: they tighten the prompt based on what broke, and the artifacts should visibly improve (fewer edits, faster turnaround). Week four: they write down their workflow as a short saved prompt or a checklist so a teammate could run it cold. That written-down version is the real graduation, because it means the habit now lives outside one person's head.
The check-ins stay short and they stay about artifacts, never about opinions. Fifteen minutes, everyone shows the thing they made, one person names one improvement they will try next week. No slide decks, no "how do we feel about AI" discussions, no roundtable of vibes. The meeting has one job: make the artifact public, because public artifacts create the light social pressure that turns a four-week push into a standing habit. When a peer shows a clean shortlist and yours is empty, you run the workflow next week. That pressure is a feature, and it is far more durable than a manager's reminder.
Do not centralize the prompts into a locked company template that nobody can change. The person who runs the workflow every week is the person who knows why the model keeps misreading the call notes, so that person needs the authority to edit the prompt on the spot. A governed workflow that the role owns beats a perfect prompt handed down from above that fits last month's version of the job. Central teams can share a starting prompt and a good example, then get out of the way. The workflow belongs to the role that runs it.
Pick tools the role already lives in, and one model to start
Choose the AI surface that sits closest to where the work already happens, so the workflow adds a step instead of a destination. A support team living in a help desk should use the AI features inside that help desk or a browser tab kept open beside it, not a separate app they have to remember to visit. A marketer drafting in a doc should draft with the assistant in that doc. Every extra context switch is a place the habit leaks, and new habits leak at every seam. The best tool for week one is usually the capable one your team can reach without changing where they work.
Start with one strong general model and one workflow before you buy anything specialized. A current top-tier assistant (Claude, ChatGPT, or Gemini) handles nearly every starter workflow on the map above: drafting, summarizing, clustering, screening against a rubric. You do not need a dedicated tool for each function to begin, and buying six point solutions before anyone has a single sticky habit is how budgets get spent with nothing to show. Prove the habit on one general tool first. Specialize only when a role's artifact shows a specific limit the general model keeps hitting, and let that real limit, not a sales demo, justify the next purchase. That buy-versus-run-in-house decision is easier once you have artifacts telling you where the general tool actually falls short.
What the artifacts tell you after a month
By the end of the cycle the folder of artifacts is your only honest scoreboard. Count them per role. A role with four weeks of steady, improving artifacts has a habit that will outlast the program, and that person is your candidate to help the next role start. A role with a scatter of half-finished artifacts has a workflow that does not fit, and the fix is to change the task, not to run another demo. A role with an empty folder is telling you the assignment never connected to real work, and pretending otherwise with a completion certificate helps no one. The scoreboard is unforgiving and specific, which is exactly why it works where the all-hands demo did not.
Team AI adoption is not a knowledge problem, it is a habit problem, and habits form around one repeated task with a visible output. Give each role one workflow drawn from work they already do, require one small artifact that is a byproduct of doing it, run a tight four-week cycle, and read the artifacts to see who has the habit and who needs a different workflow. Skip the demo. Build the folder.
Related
- ai-champions-for-smb-teams - smb-ai-enablement-buy-vs-run-in-house - smb-ai-strategy-one-workflow