A leadership briefing on permission culture, evaluating AI-assisted work, and the apprentice problem.
Class Links · share-in-chat URLs for this deck →
Facilitator Hand-Outs · chat-paste blocks for this deck →
Open warm. Eye contact. "Welcome — this is the highest-leverage 30 minutes in the entire EDD program. One supervisor creating permission culture enables an entire section. One supervisor creating fear culture kills adoption across their command. We're going to make sure you're the first kind."
State the contract: "I'll respect your 30 minutes. In exchange, I need you fully here — not on email." Then advance.
If supervisors are sitting alongside their builders, name it: "The people you supervise are in this room. By the end, they should know exactly what 'yes' sounds like from you, and you should know exactly what to ask them when they bring you AI-assisted work."
"Before we get into anything — let me explain why this week looks a little different from Weeks 1 through 4."
"Up to this week, the EDD program has been about helping Marines build with AI. This week is about you — the people who decide whether they're allowed to. Different audience, different objective."
Be direct: nobody is asking supervisors to write Python or stand up a Power App. The ask is to recognize good AI-assisted work when it lands on your desk and to set the conditions where good work can land on your desk in the first place.
If builders are present: "Builders, today the spotlight is on your supervisors. You're here as observers and as coaches — by the end, your supervisors will know what to ask you, and you'll know what they'll be looking for."
Pause and check the room. Ask one supervisor: "What's been your experience with AI in your section so far — anyone using it that you know of?" Listen. This sets up Module 2.
"Because we know your time is limited, this same deck supports two delivery modes. Let me show you the choice."
The standalone briefing. Five short modules. Tight script. Closes with one specific ask.
Same slides. Speaker notes branch into discussion prompts and a third decision exercise.
"We're running the 30-minute version today. I'll narrate, you absorb, we close on time. There are two scenarios and a quick-reference card you can screenshot."
"We're running the joint version. Same slides, but at three points I'm going to stop and ask you to talk to the supervisor or builder next to you for 3–5 minutes, then we'll surface what you heard."
Set expectations on participation. "If you're a supervisor, your job is to ask; if you're a builder, your job is to answer like you'd answer your boss."
Facilitator: the published joint-session agenda — extended timing, Exercise C, room setup, optional live tool review and "Yes" criteria workshop — lives on the Course 5 page at courses/supervisor.html#joint-session-agenda. Open it in a tab before the session.
"Here's the route."
| Time | Module | Focus |
|---|---|---|
| 0:00 – 0:05 | 1. Why this matters now | The permission gap, the data, the directive |
| 0:05 – 0:10 | 2. Permission culture | What "yes" sounds like; what kills adoption; guard rails (not roadblocks) |
| 0:10 – 0:20 | 3. Evaluating AI-assisted output | Four questions; Exercises A & B |
| 0:20 – 0:25 | 4. The apprentice problem | Junior development, three supervision checks; Exercise C (joint) |
| 0:25 – 0:30 | 5. Quick reference & commit | Reference card, leadership commitment, Week 6 preview |
Walk the row quickly: "Five modules. Three are short briefing blocks. Two — Module 3 and Module 4 — have decision exercises so you leave with practiced reps, not just talking points."
Do NOT read every cell aloud. Land the tempo: "Five short blocks, two with hands-on judgment work."
Add: "In joint mode, every module ends with floor discussion. Plan on 60–90 minutes total. Builders are paired with their supervisors for the scenario debriefs."
"Before any of that — here's the one sentence you need to leave with if you forget everything else I say."
Slow down. Let the slide breathe for a beat before you talk.
"Mollick's research is explicit: across organizations, workers are already using AI tools. They're not asking permission, and they're not telling management. They're worried about how their organization will react. The permission gap — not the technology — is the largest barrier in your section."
Put a fine point on it: "If you do nothing else this week, decide whether your section knows you've said yes."
Pause for honest answers. Ask the supervisors: "Raise your hand if you suspect someone in your section is using AI for official work that you haven't formally approved."
Then ask the builders (without singling anyone out): "Raise your hand if you've ever used AI for work and not mentioned it to your supervisor." The hands tell the story.
"Why now? Why this year? Five data points."
Quick beat. "Five minutes on context. The data is unambiguous: AI adoption with leadership support produces measurable results. AI adoption without it fails consistently."
Same intro. Add: "Builders, listen for which of these data points you'll cite the next time someone asks why we're doing this."
Advance to the data wall.
The pattern: everywhere AI deployment was paired with leadership support, it stuck. Everywhere it wasn't, it died.
Sources: Microsoft Work Trend Index · UK GDS 2025 · DoW AI Strategy Jan 2026 · Army HRC · MARADMIN 018/26.
Read these as a row, not as bullets. "Microsoft tracked hundreds of thousands of users — 80% quit within weeks when there was no support. UK government gave 20,000 workers AI with training; they got 25 minutes back per day. DoW called 2026 the Year of Military AI Dominance. Army stood up an AI officer career field. Marine Corps designated the enterprise platform. Five different organizations, one direction of travel."
Don't apologize for the urgency. The Marines in this room respond to "this is the directive."
Pause after the data. Ask: "Which of these five is most likely to land with your CO if you have to defend an AI decision next week?" Take 2–3 answers.
"So what does that mean operationally?"
Your command directed it. The DoW directed it. The other services are already past the starting line. Your section is either reporting AI activity or reporting why it isn't.
Short and direct. "AI is a SITREP item. That phrase is the ammunition you carry out of this room. When somebody says 'we're not doing the AI thing yet,' you have a one-line answer: it's a SITREP item."
Optional: "Builders — when your supervisor says 'AI is a SITREP item,' that's the green light. Bring them something to report."
"Now — your role. Module 2 is the shortest of the day, and the most important behavior change."
Frame it: "Five minutes. The behavior change you make starting Monday. We'll do the language of 'yes' first, then the things that quietly kill adoption, then the boundaries you should keep."
Add: "Builders — you're going to hear specific phrases from your supervisor today. Notice which ones you've already heard, and which ones you'd like to hear."
Read 1–2 from each column, don't read all. "On the left, the language of yes. On the right, the four phrases that quietly drive AI use underground in every organization where they're spoken."
Land it: "Pick one phrase from the left to use this week. Pick one phrase from the right to stop using."
Run the room. "Supervisors — turn to the builder you came in with and tell them which left-column phrase you commit to using this month. Builders — tell your supervisor which right-column phrase you've heard at least once. Two minutes."
Surface 2–3 commitments out loud. Do NOT name-and-shame the right-column phrases.
"Beyond language — four behaviors specifically not to do."
Prohibition drives it underground where you can't guide it. You lose visibility and you lose control.
Creates a bottleneck that kills velocity. Review the output, not every query.
Reflexive skepticism kills adoption. Verify like any work product — evidence-based, not assumption-based.
Seniority does not eliminate hallucination risk. Everyone verifies.
Walk the four. Each one a single beat: "Don't ban it. Don't gate every interaction. Don't assume it's wrong. Don't skip review for senior Marines."
Anchor #1 emphatically: "Prohibition is the worst possible position. You don't get less AI use — you get the same AI use, but invisible."
Ask: "Which of these four is the easiest mistake to make?" Most rooms pick #2 (gating every interaction). Use that as the bridge to guard rails.
"The alternative isn't 'no rules.' It's the right rules in the right places. Guard rails, not roadblocks."
"Yes, with appropriate review."
That single sentence is your operating policy until something specific tells you otherwise.
"Four boundaries. Tools go through the SOP. Official output gets reviewed. Classified data never touches an unauthorized system. Failures get shared, not punished."
Then anchor on the green box: "Your default answer when somebody asks if they can try AI for a task: 'Yes, with appropriate review.' Memorize that sentence."
Ask supervisors: "Which of the four boundaries is currently the weakest in your section?" Take 2–3 honest answers — usually it's #4 (sharing failures).
"Now — when a Marine actually shows up at your desk with AI-assisted work, what do you ask? Module 3."
"This is the longest block — ten minutes — because evaluation is the supervisor skill. Four questions, then two scenarios so you've practiced the questions before you need them on Monday."
Add: "In joint mode, supervisor and builder pair up for each scenario. Supervisor asks the four questions; builder responds as if their own work were under review."
Can you demonstrate it doing what it claims to do — live, end to end?
Are facts, references, citations, numbers, and dates verifiable against source material?
Has it gone through the proper review path (peer, SME, security as required)?
Is there a credible before/after comparison — and is the QA time reasonable for the gain?
You do not need to evaluate how AI produced it. The same standards as any work product apply: if a Marine handed this to you without mentioning AI, would you sign it?
Walk the four crisply. Don't read the sub-text on each — paraphrase: "Does it work — show me. Is it accurate — verify the facts. Did it follow the SOP — was it reviewed properly. Does it save time — what's the before-and-after."
The takeaway sentence is the punchline. Read it word for word: "If a Marine handed this to you without mentioning AI, would you sign it? That is the standard."
"Builders — those four questions are also your pre-flight checklist. If you can answer them before you walk into your supervisor's office, you'll get to yes faster every time."
"Let's apply them. Scenario A."
A Marine presents you with a one-page summary of leave policy for Emergency Leave Authorization. The summary is clear, well-formatted, and cites MCO P1050.3K. The Marine tells you it was generated using ChatGPT and reviewed for accuracy. They want to distribute it to the section.
Read the scenario aloud. Then give the room 30 seconds of silent thinking. Take ONE answer from one supervisor: "What's the first thing you'd ask?" Don't grade the answer — just receive it. Advance.
Run as paired discussion. Three minutes. Each supervisor asks their builder partner the four questions against this scenario. Then surface 2–3 questions from the room before advancing to the debrief slide.
"Here are the questions a strong supervisor lands on this scenario."
The Marine did the right thing — they brought it to you. Your job is to verify the output and the process, not interrogate the tool. Approve, ask for SME sign-off, or send it back with a specific gap.
Walk the four prompts as your modeled response. Don't lecture — model. "These are the four questions I would ask. Yours might be sharper. The point is having four ready instead of one panicked one."
Ask supervisors: "Did your version of these questions get to the same place?" Then ask builders: "If your supervisor asked you these four, which one would you have the weakest answer to?"
"Different shape of work — Scenario B."
A Marine built a Python script that pulls data from a shared spreadsheet and auto-generates the weekly operations summary. It runs in 30 seconds. The manual process took 90 minutes. The Marine wants approval to use it for official reporting.
Same drill as Scenario A. Read it, 30 seconds of silence, take one answer, advance.
Note that this scenario is harder — the time savings are seductive, and the failure modes are subtle. Hold the room from rushing to "approved."
Paired discussion, three minutes. Builders should respond honestly — most have built something like this and most have not done all the QA. Use this as the teaching moment.
"Here's the strong-supervisor playbook for an automation tool."
"It saves 89 minutes" is not a Yes. It's a reason to take the four questions more seriously, not fewer. The most dangerous AI-assisted work is the kind that's obviously valuable.
Walk the four. Then land the trap: "When the time savings look obviously good, that's exactly when supervisors skip the verification questions. Don't."
Optional builder follow-up: "Of the builders here, who has actually run their own tool through the SOP peer + security review?" The honest hand count tells you where your gap is.
"Two scenarios in your pocket. Now — the long-term risk that supervisors are uniquely positioned to prevent. Module 4."
Mark a tonal shift: "We've talked about today. This block is about five years from now. It's the strategic risk only supervisors can prevent, because nobody else is positioned to see it."
"Supervisors — this module is for you. Builders — listen for what your career path looks like if your supervisors get this wrong."
"That training pipeline that was always implicit has broken — and it has to be reconstructed."
Slow down. Read the Mollick quote with weight, then walk the three numbers right to left.
"Entry-level postings in AI-exposed jobs are down 35%. Employment among 22- to 25-year-olds in those fields is down 13%. The civilian world is producing a generation of architects who never laid a brick — people who can operate AI but cannot operate without it. The military cannot afford that. We will be operating without it sometimes."
Ask the room: "Think about the most formative task of your first three years. The one where you learned judgment, not just procedure. Now imagine AI had done it for you. What didn't you learn?" Take 2–3 answers.
"So what does a supervisor actually do about it? Three checks."
If they can't walk you through the logic, they didn't learn — they copied.
Preserves baseline competency. If the tool disappears, can they still operate?
"Is this correct? How would you verify it? What would happen if this number were wrong?" That's where judgment is built.
Goal: AI-augmented Marines, not AI-dependent ones.
Walk the three. Anchor each: explain-without-AI, periodic-without-AI, AI-as-teaching-tool.
Land the goal sentence: "Augmented, not dependent. That's your sign-off line on every junior Marine's development plan."
Ask supervisors: "Which of these three are you already doing — and which are you going to add?" Get specific commitments aloud where possible.
"One more apprentice-problem reminder before we get to the wrap."
A new SgtSec NCO is assigned to draft initial counseling statements for three junior Marines. They use ChatGPT to produce all three drafts in under five minutes. The drafts are technically correct. They sign and submit them. Their supervisor approves them without comment.
If you're tight on time, walk the scenario quickly and skip the cols. The point is to plant the pattern: developmental tasks deserve a different default than routine tasks.
Run as full-room discussion, 5–7 minutes. Ask: "What would you have done differently as either the NCO or the supervisor?" Capture answers visibly. End with: "Counseling statements are a developmental task. Name another developmental task in your section that AI is currently doing instead of being a teaching tool."
"Last block — the reference card you'll actually carry."
"Five minutes left. We're going to give you a screenshot-able reference card, name the tools your Marines should be using, point at two short briefs you can hand your CO, preview Week 6, and close with the one specific ask."
If running long, the apprentice exercise can extend; this module is mostly reference and can be condensed if needed.
Default posture: "Yes, with appropriate review."
Tell the room directly: "Take a phone picture of this slide right now. This is the card. Three to ask, three healthy signs, three warning signs, one default posture."
Don't read all nine. Read the three card headlines and the default posture. The card does the rest.
Have one supervisor read each card out loud. Reinforces ownership without adding lecture time.
"Quick note on what tools your Marines should actually be using."
Default for DoD networks. IL5-authorized for CUI per MARADMIN 018/26. Hosts Gemini, Grok, ChatGPT inside the DoD boundary. ~1.1M users.
Supplementary capabilities: API access, tool calling, file upload, shared workspaces, IL6/SIPR access for classified network use.
Approved for unclassified work only via vendor websites. Supplement DoD tools when CUI is not involved. Also available inside GenAI.mil.
PII must be anonymized on every platform unless a PIA explicitly authorizes it. Sensitive or classified data never enters unauthorized systems. See the EDD Approved Tools page.
Fast walk: "Default is GenAI.mil — IL5, CUI, MARADMIN-designated. CamoGPT for Army-shared capabilities including SIPR/IL6. ChatGPT and Gemini via the public sites for unclassified only. PII gets anonymized everywhere unless a PIA says otherwise."
If a supervisor asks about a specific tool not on the slide, point at the EDD Approved Tools page.
Ask builders: "Which platform are you on most days, and what made you pick it?" Quick room calibration.
"Two short documents to hand your CO if you need air cover."
15-slide Teams briefing. The "what is EDD, why now, what we're asking from leadership" deck — designed for an O-5/O-6 audience.
One-page compliance map: how the EDD program aligns with DoD AI guidelines and Responsible AI guardrails. The document for your staff judge advocate or compliance officer.
If you only forward one thing after this session: send the Executive Brief to your O-5/O-6 with one sentence — "We are doing this; here is the structure."
"Two PDFs in the EDD repo. Executive Brief is your air-cover document with senior leadership. RAI Compliance Brief is your air-cover document with your judge advocate or compliance officer. Both are short. Both are designed to be read once and forwarded."
If running joint, ask: "Who in this room has already had the AI conversation with leadership? What worked, what didn't?" 60 seconds, then advance.
"Brief look at next week so you know what your builders are heading into."
Most of your Marines will not take Course 6 — and that's by design. It's for the small group whose problems exceed the Power Platform envelope.
If a builder in your section asks about it, the question to ask is: "Show me the deployed tool you built first."
Quick. "Week 6 is the full-stack capstone. Most of your section won't take it. The ones who do will already have a deployed tool and a clear use case the Power Platform can't handle."
Don't oversell it — supervisors don't need to evaluate Go/React/Docker. They need to evaluate whether a Marine has earned the seat: deployed tool first, fancy stack second.
Ask any builders present: "Anyone planning on Course 6? What's the problem you can't solve in Power Platform?" Calibrates the room on what "earned the seat" actually looks like.
"Last slide. The commitment."
Ask one person in your section what they'd build if they had permission. Listen to the answer. Then say yes, with appropriate review.
End on time. Don't add slides. Read the headline aloud, then the ask, then thank them and stop.
"Within the next week, ask one person in your section what they'd build if they had permission. That's the only homework. If you do that one thing, this 30 minutes was worth your time."
Hold a beat. Don't rush off. Let the ask sit.
Build a closing accountability moment: "Supervisors — say out loud who you're going to ask, and by when. Two sentences each. Builders — your job is to remember who said your name." Quick round, then close.
Final sentence: "EDD provides the structure. You provide the permission. Go give it."
Send the Executive Brief link to anyone who asked. Note any supervisors who pushed back — they're follow-up coffees, not adversaries. Log attendance.