An HVAC owner I work with sent a company-wide email about AI last month. It wasn’t a sales pitch — it was a real announcement. “We’re using this now. Dispatcher, you’re running route optimization through it. Office manager, you’re drafting customer follow-ups with it. Everyone else, use it where you can.”
Clear. Direct. He’d bought the team AI subscription, granted access, dropped the email on a Tuesday, and walked away.
Six weeks later, dispatcher was still building routes on a spreadsheet. Office manager was still writing follow-ups in Gmail. Nobody had used the tool in any meaningful way. Dispatcher had tried it once, didn’t understand what he was supposed to do with it, and gave up. Office manager had literally forgotten the password.
When he called me, he didn’t say AI isn’t ready. He said something sharper — “I think I announced it. I didn’t deploy it. Totally different.”
Yeah. That word pair — announced versus deployed — is the whole diagnosis. If your AI tool is sitting unused in your business, you’re probably here.
The trap has a shape
Here’s what announcement looks like.
- Owner identifies a problem (routing is manual, proposals take too long)
- Owner researches the tool (ChatGPT, Claude, Perplexity)
- Owner sets it up (pays for Teams, creates the login)
- Owner broadcasts it (all-staff email, quick Slack message)
- Owner assumes deployment happened
- Everyone ignores it
- Owner concludes AI isn’t ready or my team won’t adopt it
Trap is step 5. You feel like something happened. You made a decision. You communicated it. A reasonable person would think that’s a deployment.
Right? It’s not. It’s the announcement phase of a deployment. The announcement is 10% of the work. The other 90% is still sitting on the table.
Here’s what you confused
You treated AI the way you’d treat a new hire if you were in a hurry.
If you hired a dispatcher, you wouldn’t just send a company-wide email — “New dispatcher starts Tuesday. Dispatcher, you’re responsible for route optimization. Everyone else, work with him where you can.” Then walk away.
You’d do this instead.
- Training — Show him the business. How does dispatch actually work today? What’s a good route? What makes a bad one? What does the owner care about?
- Context — Before his first shift, walk him through an actual job list. “Here’s what a Tuesday morning looks like. Here are the constraints.”
- Guardrails — Tell him the rules. “Never commit to a time you’re not sure about. Never send a customer a quote without running it by me. If a customer complains about timing, escalate it.”
- Examples — Show him five routes you’re proud of. “These work. Here’s why. This one didn’t work. Here’s what went wrong.”
- Specification, not procedure — Tell him the outcome you want. “Get the crew on the job as early as possible, minimize dead time, respect customer windows, don’t double-book.” Don’t give him a 12-step checklist.
- Measurement — Check in. “How many routes are you running? How fast? Any customer complaints?” You want data, not vibes.
- Feedback loop — When something breaks, update the system. Add it to the training. Update the examples. Tighten a guardrail. Don’t just fix it and move on.
A dispatcher who gets the seven things on that list is a working dispatcher.
A dispatcher who got an email and nothing else will sit at the desk confused and eventually leave.
Same is true of AI. The announcement is step 4 (communication). It’s not the other six. Does that make sense?
What you’re actually seeing
If your AI tool is unused or mostly unused, one of these is true.
Scenario 1 — No Training. Your team knows AI exists. They don’t know your business the way the AI needs to know it. Tool doesn’t understand your customers, your pricing, your voice, your rules. It talks like a stranger. Nobody recognizes the value.
Scenario 2 — No Context. Team knows the tool exists, but they don’t know when to use it or what situation to feed it. So they use it randomly, get mediocre results, and stop bothering.
Scenario 3 — No Guardrails. Tool ran somewhere and broke something. Quoted a customer without your sign-off. Sent an email in a tone that felt wrong. Made a commitment the company can’t keep. Now people don’t trust it.
Scenario 4 — No Examples. Team tried it, got output that sounds generic and kind of useless. Why would I use this when I can just do it myself? They stop.
Scenario 5 — Wrong Instruction. You told them the process — “Step 1 paste the customer info. Step 2 ask for a route. Step 3 copy the route to the spreadsheet.” The AI followed your process and produced mediocre work that sounds caged. Nobody sees the value.
Scenario 6 — No Measurement. You don’t know if it’s working because you’re not tracking anything. Team uses it or doesn’t. You have vibes, not data. So every conversation about whether to keep the tool is opinion-based.
Scenario 7 — No Feedback Loop. When the tool screwed something up, the team fixed it in the moment and didn’t tell you. Same mistake happens next week. And the week after. Team keeps fixing it privately. You think the tool is dumb. Actually, the system is unmaintained.
Most underperforming AI deployments have all seven gaps. But you only need to close two or three to see dramatic improvement.
What deployment actually looks like
Here’s the thing. What the HVAC owner should have done instead.
Step 1: Pick one person. Pick one job. Not the whole company. Not the whole dispatcher workflow. One person (the dispatcher). One specific job (optimizing a route for Tuesday morning).
Step 2: Sit down with them. Run it once together. Not a training video. Not an email. You and the dispatcher at his desk, pulling up the route tool, feeding it today’s jobs, watching the AI suggest an order, comparing it to what the dispatcher would do. Talk through it. “I see why it put this one first. That makes sense. But this job’s customer hates waiting, so we’d actually do it second.” Update the prompt.
Step 3: Measure it. Does the dispatcher use the tool on his own now? How often? Did the routes get faster or slower? Did customers notice? Get a yes or no.
Step 4: Close the loop. If something doesn’t work, update the tool. Add the lesson to the system prompt. Try again next Tuesday.
That’s not six weeks of rollout. That’s one hour on Tuesday morning. Then you have real data about whether this actually helps.
If it does, you can roll it to the office manager’s workflow. Same pattern. One person. One job. One hour.
If it doesn’t, you know specifically why (which of the seven gaps killed it) and you can fix it or kill it with honesty.
The Monday Move
So. Stop planning a full rollout of AI tools. Start deploying them one workflow at a time.
Pick one person in your business whose job would actually be easier if they had good AI support. Dispatcher routing jobs. Office manager drafting follow-ups. Estimator writing proposals. Anyone.
Block 90 minutes this week. Sit down with them at their desk. Bring the tool. Bring one actual scenario from their job — a real route, a real customer, a real estimate.
Feed it into the AI. Look at the output together. Ask — “Does this help? What’s missing? What’s wrong?” Update the tool based on their feedback.
Now ask them to use it on their own for one week. Give them permission to adjust the system prompt themselves if something’s off. (Don’t require them to ask you first — that kills iteration velocity.)
Check in a week later. Data point — did they use it? Did it actually save time or make the work better?
If yes, that’s a real deployment. You can expand from there.
If no, you have a diagnosis. Not AI doesn’t work. One of the seven ingredients is broken. Fix it.
Difference between this and what you did with the all-staff email is that this one is work, not communication. This one requires you to actually understand their job, not just announce something and hope.
But it takes a few hours, not months. And it gives you real data.
The shift
Deployment and announcement look similar from 30,000 feet. They both involve a tool, access for the team, and communication. But they’re different jobs.
Announcement is passive. You’re telling people a thing exists. They can take it or leave it. Most leave it.
Deployment is active. You’re integrating a tool into a specific workflow, training the person in that workflow to use it, measuring whether it works, and fixing it when it doesn’t.
You’re good at deployment. You’ve done it with people — trained dispatchers, office managers, estimators. You just treated this tool like an announcement instead of an onboarding project.
So. Stop announcing AI to your business. Start deploying it to your workflows. Pick one. Run it right. Measure it. Expand from there.
Tool you’ve had sitting unused for six weeks isn’t broken. Your deployment was incomplete.
Framework: The Professional Recipe — all seven ingredients, which define the work announcement skips. Related failure mode: The Announcement Trap (#1).
Companion piece: Stop Hiring AI. Start Building It. — the parent principle (building, not hiring) that makes this deployment model work.
~ source material · The Professional Recipe (all 7 ingredients)
