ai-literacy · deliverable

Facilitator note — running “Train the AI, Check the AI” in a 15-minute huddle

For: Support team leads · Group: 4–10 agents, one screen or everyone on their own laptop · Prep: 2 minutes (open the lesson, have one real anonymized draft ready if you have one)

This lesson is designed to run inside an existing huddle, not instead of one. The pedagogy is deliberate: consequence before instruction (Scene A shows a confident wrong draft), detection before procedure (Scene B hotspot hunt), procedure under feedback (Scene C ladder), choice with stakes (Scene D branch). Your job is to keep the pace and land the debrief — the lesson carries the content.

Minute-by-minute

Time	What happens	Your move
0–2	Scene A: read the draft aloud together	Ask: “Would you send this?” Take a quick show of hands. Do not reveal anything.
2–6	Scene B: hotspot hunt	Let people click and argue. If the room finds all 3 fast, ask why the decoys are safe — that distinction (invented fact vs. style) is the actual skill.
6–11	Scene C: Verify ladder	Run it as call-and-response: read the KB excerpt, ask “what do we do with this clause?” before clicking the answer. Wrong answers are good here — let them happen.
11–13	Scene D: the send decision	If anyone picks “send the original”, play the consequence email aloud without commentary. The email does the teaching; you do not scold.
13–15	Debrief — the only part that is yours	Three questions, 40 seconds each (below). Close on the `#ai-catch` tag.

The three debrief questions

“Which risk would you have caught before this session?” — honest self-baseline; usually 1 of 3. That gap is why the ladder exists.
“Which rung feels slowest in real tickets?” — this is your process feedback. If rung 1 (KB search) is slow, that is a KB gap to fix, not an agent problem.
“What do we do when we catch one?” — land the norm: tag #ai-catch, no blame, the catch is the win. Volume of catches is our quality metric, not our shame metric.

Facilitation guardrails

Never frame catches as individual failure. The draft-assistant produces confident text; catching it is a designed team behavior. Praise the catch, not the catcher’s cleverness.
If someone says “I just don’t use the assistant” — that is a valid choice, but note the guideline applies to any draft that leaves the team, hand-written or AI. The ladder is about claims, not about tools.
If the room wants to debate policy (e.g., “should Billing confirm every amount?”) — park it. The huddle teaches verification; policy questions go to the quarterly guideline review.
Time-box discipline: the hotspot hunt is where huddles die. At minute 6, move on even if one risk is unfound — Scene C reveals it anyway.

What good looks like after the huddle

Agents can name at least 2 of the 3 risk types without prompting.
The team knows the #ai-catch tag and uses it in the first week (check the count in your weekly ticket review).
At least one agent reports a rung that is slow in practice → you have a concrete improvement for the KB or the guideline review.

Reuse & adaptation

The lesson is a single offline HTML file — no LMS, no accounts, no network. To adapt it to your own product: swap the draft email, the KB excerpts, and the consequence email; keep the four scenes and the ladder structure. If you change the draft, keep exactly 3 real risks and 2 safe decoys — the decoys are what teach detection instead of paranoia.

Fictional disclosure: Nordwind Software and all referenced tickets/persons are a fictionalized composite for this portfolio. Note drafted with AI assistance; final judgment and copy are human-owned.

Prefer the source? Download the raw markdown.