We Replaced a $150K AI Training Program With LEGOs
“Jay, we’re not seeing any ROI with AI.”
Michelle threw her hands in the air.
“Yesterday in a meeting, my CEO saw the $150,000 invoice for renewing our Microsoft CoPilot licenses. He. Flipped. Out.” She wasn’t done. “Adam barked at all of us, questioning ‘Why are we not gaining any traction?!’”
“We gave our teams LinkedIn Learning modules on AI, but nobody is making time to watch them.” She shook her head. “I don’t think it’s a learning problem. I think people are struggling to integrate AI meaningfully in their workflows.”
She was right.
Six weeks later, I was standing in front of her team.
• • •
Twenty-three people in a conference room on the fourth floor. Fluorescent lights.
Two boxes of donuts on a folding table in the back that nobody had touched yet. A pile of LEGO bricks in the center of the room that was getting suspicious looks.
Half the room was on their laptops.
A few were whispering to each other with the energy of people waiting at the DMV. One guy in the corner (I’d later learn his name was Derek) had his arms crossed and his chair pushed back from the table a solid six inches further than everyone else’s.
The universal body language of I have been voluntold to be here and I want you to know it.
Michelle stood in the back, watching me with the expression of someone who’d bet her credibility on this morning going well.
I had three hours.
• • •
“Before we go into that room, here’s what you need to know.
Rich and I have run this exact workshop over 50 times. The first version was a full-day event — eight hours of AI history, ethics, prompt engineering theory, tool demos, hands-on labs. It was thorough. It was well-produced. And the feedback gutted us: “Great session, but I still don’t know what to do on Monday.”
So we did what any reasonable people would do. We cut it in half. It was still too long. We cut it again. We killed the AI history lesson. We killed the prompt engineering lecture. We killed the 45-minute tool demo where people nodded politely while checking their phones under the table. We killed every section that made us feel smart but didn’t change a single behavior.
What survived was three hours with one rule: nobody watches anyone else do AI. Everyone builds something themselves.
This is what that looks like when it works.”
9:00 AM
I didn’t show an agenda.
I didn’t introduce myself with a bio slide.
I picked up a donut, took a bite, and asked a question.
“How many of you use AI on a daily basis?”
Most hands went up.
A few tentative ones. Derek’s stayed in his lap.
“Okay. How many of you feel like you’re using it to its full potential?”
Almost every hand dropped.
A woman in operations laughed out loud - the uncomfortable kind, the kind that means you just named something I’ve been thinking about for months.
I let the silence sit for a beat. That gap between those two questions - between I use it and I’m not getting anything from it - is the entire reason 23 people were in the room on a Tuesday morning instead of at their desks. Now they could feel it.
“I’m going to ask three things of you today,” I said. “Be present. Play outside your comfort zone - especially if you think you’re not a tech person. And dial a friend sooner rather than later. If you’re stuck, ask me, ask your neighbor, or hell, ask ChatGPT.”
Nods.
A couple of “yeahs.”
Derek uncrossed his arms long enough to reach for his coffee. I counted it as progress.
• • •
9:10 AM
“Imagine you’re Bond. James Bond.”
A few smirks.
“Or Jane.”
A real laugh from the left side of the room.
“If you could build a magical invention that solves your single biggest problem at work — any problem, no constraints, laws of physics don’t apply — what would it be?”
I gestured toward the LEGOs.
“Pair up. You have ten minutes. Build it.”
Silence for exactly four seconds.
I know because I counted.
Four seconds is how long it takes a room of adults to decide that playing with LEGOs during work hours is allowed.
Then the room cracked open.
Two women from operations dove into the LEGO bin like it was Black Friday. A guy from finance was sketching on a napkin before his partner even pulled her chair over. A quiet developer in the middle of the room started building something intricate without saying a word to anyone - just immediately locked in.
I looked at Derek. He’d picked up a blue brick. He turned it over in his hand. Then, grudgingly, almost irritably, he started stacking.
Within sixty seconds, every laptop in the room was forgotten.
• • •
When the timer went off, we went around the room.
The operations pair had built something they called “The Auditor.” It was a lumpy construction of red and yellow bricks that they described with absolute conviction. “It scans every vendor invoice against the contract terms before anything gets paid,” one of them said. “We lose two full days a month checking this stuff manually. Two days.” Her partner jumped in: “And we still miss things. We caught a $40,000 overbill last quarter by accident. By accident.”
A product manager had built something she called “The Translator.” It would take customer support tickets (messy, emotional, jargon-filled support tickets) and turn them into structured product requirements her engineering team could actually act on. “Right now, I’m the translator,” she said. “It’s about 30% of my job.” She paused. “And I hate it.”
The quiet developer had built an elaborate machine he said could detect bugs in code before they reached production. No one fully understood how his LEGO model represented that, but his enthusiasm was infectious.
Then it was Derek’s turn.
He’d built a tower. Just a tower. Tall, narrow, slightly crooked.
“What does it do?” I asked.
He shrugged.
“It finds every status report I have to fill out and sets them on fire.”
The room erupted.
Michelle, in the back, put her face in her hands laughing.
Even the quiet developer was grinning. And Derek (arms-crossed, chair-pushed-back, voluntold Derek) let a smile crack through.
I’d been doing these workshops long enough to know what I was looking at. That wasn’t just a joke. That was a man who spends hours every week on a process he finds completely pointless, and this was the first time anyone had asked him about it.
• • •
“The facilitation move that makes the rest of this workshop work:
While people present their inventions, write every use case on an easel sheet. “Invoice checking — got it. Ticket translation — noted. Status report elimination —” Derek grinned at that one. You’re building the use case library for the hackathon, and the room doesn’t know it yet. Two hours from now, when you tell people to pick a problem to solve, they won’t stare at a blank screen. They’ll look at that wall.
We’ve tested this workshop with and without the 007 Invention. Without it, teams burn the first 15–20 minutes of the hackathon trying to agree on what to build. With it, they’re building in under five. This one activity eliminates the cold-start problem that kills most hackathons. Don’t think of this as an icebreaker, think of this as the foundation.
”
9:40 AM
I pulled up five slides. Ten minutes. Not the history of AI. Not a tour of every new model release.
Five concrete things relevant to the work these specific people did every day:
CoPilot agents with shared org-wide context
Claude building financial models from a single prompt
AI browsers handling full research workflows.
I skipped everything that would make me sound smart but wouldn’t change what they did on Monday.
Then I split them into small groups with one question: What do you think this means for your team?
The operations pair started talking immediately. The product manager was already on her phone Googling something. But I was watching Derek. He’d leaned forward - actually leaned forward, chair now at the table, elbows on the surface - and said to his group, loud enough for me to hear: “If that agent thing is real, we could probably kill the Monday reporting process entirely.”
I walked to the easel and added it to the list.
• • •
9:50 AM
“Alright. You’re about to build your first AI agent.”
I said it casually. Like it was no big deal.
Like of course they were going to do this.
We built it together, step-by-step, everyone on their own laptops.
A simple CoPilot agent called Policy Polly - designed to answer questions about the company’s travel, HSA, and vehicle use policies.
We uploaded a pre-built knowledge base so the barrier to entry was as close to zero as possible. No one got stuck on file formatting. The point wasn’t the agent. The point was the moment it worked.
But the real unlock wasn’t the tool. It was the instructions. We’d written a tight set of constraints for the agent - respond only to policy queries, avoid speculation, guide users to resources when you can’t answer directly.
When I asked the room what they noticed about those instructions, the product manager said: “They’re really specific. Way more specific than how I usually talk to ChatGPT.”
And there it was.
That moment happens in every workshop, and you can’t lecture someone into it.
They have to feel the contrast between their vague “help me with this” prompts and a set of instructions that actually constrain the AI’s behavior.
When they feel that gap, prompt quality stops being an abstract concept and becomes something they understand in their hands.
Within five minutes, everyone in the room had a working agent.
People were testing questions, laughing at responses, leaning over to show their neighbor what it said.
Someone asked Policy Polly about the dress code and it pulled the exact paragraph from the employee handbook.
Someone else asked a trick question about a policy that didn’t exist and it correctly said it didn’t have that information.
“Oh, that’s because of the ‘avoid speculation’ instruction,” someone said. Now they were reverse-engineering the prompt. Without being asked.
Derek leaned over to the guy next to him. I watched it happen from across the room. “I’ve had CoPilot for eight months. Eight months. And I never understood what this thing could do.”
• • •
10:30 AM: Break
Ten minutes. If you don’t give people a window to clear the inbox anxiety, they’ll be half-present during the hackathon. Ten minutes now buys you an hour of focus. I used the break to grab a donut. The boxes were nearly empty. That’s always a good sign!
People eat when they’re comfortable.
10:40 AM
“You built your first agent. That was the warm-up.”
I pointed to the easel sheet.
“Now you’re going to build the one that actually matters.”
Every use case from the 007 Invention was still on the wall. The Auditor. The Translator. Derek’s status-report incinerator. Nine total.
“Groups of three. Your problems are on that wall. Pick one, combine two, or riff on something new. Fifteen minutes to decide. Not sixteen.”
That deadline is deliberate.
We learned it by watching teams burn 20 minutes debating use cases in open-ended time. They never start building. Discomfort with a deadline produces decisions. Comfort with open time produces committee meetings. A $50 gift card for the best solution doesn’t hurt either - we’ve watched output quality double when something is on the line.
Groups formed fast. The operations pair grabbed the accountant who’d been sitting near them all morning. The product manager recruited the quiet developer. And Derek’s group — three people who I would have bet money were the least engaged humans in the building at 9:00 AM — were huddled around a laptop trying to figure out if they could build an agent that auto-generated status reports from Slack messages.
Not everyone hit the ground running.
One group spent their first ten minutes trying to get their agent to stop hallucinating contract terms that didn’t exist. Another group’s process map had seven boxes and they couldn’t figure out which two to cut. I walked them through the constraint: three questions.
What is your normal process today?
What would you want AI to handle?
What steps still need a human?
Sketch it on paper. Five boxes with arrows between them. If you can’t explain your process in five steps, you don’t understand it well enough to automate any part of it.
The paper forces the clarity that tools can’t.
Twenty minutes to map. Twenty minutes to build.
I walked the room, answered questions, pointed people to each other when someone had already solved the problem another group was stuck on. Resisted - physically resisted - the urge to take over anyone’s keyboard.
Then the demos happened, and something shifted.
The operations team had their agent cross-referencing invoice line items against contract terms.
They tested it on a real invoice from last month. It flagged two discrepancies they’d actually missed.
The woman who’d found the $40,000 overbill “by accident” turned to her partner with her mouth open.
“That took us a day and a half last time,” she said quietly.
The product manager’s group had the ticket translator converting a real support email into a formatted product requirement.
The developer had done something with the prompt structure that made the output cleaner than what the PM had been writing by hand. “Wait,” she said, reading the output. “This is better than what I’d write.”
And Derek’s group.
They were laughing. Actually laughing - the deep, involuntary kind that makes other groups look over.
Their agent had just generated a perfectly passable weekly status report from a Slack thread that contained three memes, a GIF of a cat falling off a table, and one genuine project update buried in seventeen messages.
Derek was staring at the output like he was seeing a magic trick. “This is... this is actually good,” he said. He sounded offended by how good it was.
• • •
11:40 AM
I brought the room back together and asked three questions. I’ve asked these three questions at the end of every workshop for a year, and the answers always surprise me.
“What worked well?”
Mostly what you’d expect — the energy, the hands-on format, the LEGOs. But then the quiet developer raised his hand. “I liked that nobody told us what to build,” he said.
The room went still for a second.
“Every other training I’ve been to, someone shows you a demo and says ‘now you try.’ This was the first time someone said ‘what’s your problem?’ and then got out of the way.”
“What ideas did today spark for you?”
The product manager raised her hand. “I have a list,” she said. “Like, a long list. Client onboarding. QBR prep. The competitive analysis we do every quarter that takes two weeks and nobody reads.” She paused. “Can we do this again?”
That’s the sentence. When someone asks to do it again, you’ve won.
“Last question. What did you struggle with — and does anyone in this room know how to solve it?”
One group had struggled with knowledge base formatting. Another group had already figured it out and explained their fix in thirty seconds.
A different group didn’t know how to set guardrails for when the agent didn’t have an answer - someone pointed back to the “avoid speculation” instruction from Policy Polly.
You’ve just turned the room into a peer support network that will outlast the workshop. They don’t need you anymore. They have each other.
• • •
12:01 PM
People were packing up slowly. Lingering. Pulling out phones to take pictures of the easel sheet. One group was still at their laptop, tweaking their agent.
That’s the tell.
When people bolt for the door, you’ve lost them.
When they don’t want to leave, you’ve built something they weren’t expecting.
Michelle caught me by the coffee table. The donut boxes were empty. The LEGO bricks were scattered across three tables.
“Derek stopped me in the hallway,” she said.
I waited.
“He asked if we can get budget for a dedicated AI sprint next quarter.”
I laughed. “Derek? Arms-crossed, chair-pushed-back Derek?”
“That Derek.” She shook her head, but she was smiling. “Three hours, Jay. Three hours.”
Three hours.
Not a LinkedIn Learning module. Not a tip sheet emailed to 200 people who never opened it. Not a $150,000 invoice and a prayer.
A room. Some LEGOs. And permission to build something real.
• • •
Two weeks later, Michelle forwarded me an email.
Her team had run a second workshop on their own — same format, same LEGO activity, same three questions at the end. They didn’t need us. They used the easel sheet from the first session as their starting point and built five more agents in an afternoon.
Her CEO — Adam, the one who flipped out about the CoPilot invoice — sat in on the demos.
He didn’t flip out this time.
He asked when they could do it again.
———
We’ve run this workshop with 50+ teams and we’re booking for 2026/27. If you want us to come in and facilitate, book a discovery call with Rich. If you’d rather run it yourself, you just read the playbook.
