Jul 1, 2026
Stop Blaming Your AI Bill. Start Educating Your Team.
A $40K AI bill isn't a spending problem, it's an education problem. Here's how to teach your team the real cost of chat vs. agents vs. automations.
Last week, a CEO called me with a $40,000 monthly AI bill. The month before it was $22,000. Six months ago, nobody was tracking it at all.
My own reaction, watching our team’s bills climb every month? Kill the tools. Move to open source. Tell people to pick one AI instead of the four we provision today.
Bad idea. I’ve since realized the real question is what is the right dollar amount of AI cost as a percentage of my team’s total salaries.
We Don’t Have a Budget Problem
The CEO thought they had a spending problem. They had an education problem.
The people running up those bills weren’t being reckless. They were doing exactly what anyone does when handed a powerful tool without a manual: they figured it out themselves. They just had no idea that running 5 agents on Opus, each with two subagents, to move files around costs real money. Nobody told them.
AI = intelligence. We all want more of it, faster, and as smart as possible. Our spend is high because we handed the team AI and never taught them how to use it to ship business value.
And if that’s my real problem, downgrading to open source models won’t fix it. You’re swapping the instrument when the problem is the player. This is the same trap I keep coming back to in AI marketing ROI: teams optimize the wrong lever because nobody defined what return actually means for the tool in question.
The 30x Problem Nobody Explained
In 2023, the average cost per AI interaction was $0.04. In 2026, it’s $1.20. That’s a roughly 30x increase. Your IT budget almost certainly didn’t grow 30x. And here’s what matters: the model prices didn’t cause that jump (they mostly went down). The usage did.
Chat is cheap. A single back-and-forth with a language model costs fractions of a cent. An agentic workflow running many subagents is a different animal. The model reasons across multiple steps, calls external tools, and loops back to QA its own work. That can burn 100x the tokens of the same chat exchange. A scheduled automation runs on a clock whether anyone needs the output or not. That’s a meter you forgot you turned on. I found this out the hard way when I burned $700 on a runaway loop testing Gemini. Daily spend caps matter.
Nobody is teaching the team the cost shape.
KPMG’s Q2 2026 survey of 204 executives at $1 billion-plus firms confirmed what I see in practice: only 36% of enterprises have any token or usage controls in place, and only 26% have real-time visibility into what their AI actually costs to run. You can’t manage what you can’t see. And you can’t teach what you haven’t measured.
The Gap Is Larger Than You Think
Only 35% of employees have had any formal AI training. Meanwhile, 94% of CEOs say AI skills are a top priority.
The gap between what leadership wants shipped and what your team actually knows is where your bill lives.
KPMG found that 35% of $1B+ enterprises name “AI cost management and economic literacy” as a top barrier to AI adoption. Not model quality. Not data privacy. Literacy. What’s slowing cost control isn’t a missing tool. It’s missing understanding.
The ROI case for fixing this: formal AI training returns $3.70 for every $1 invested (Microsoft-IDC). Trained employees are 2.7x more proficient than those who figured it out on their own. Training isn’t a soft people-ops line item. It’s one of the cheapest ways to bring your AI spend down.
My AI Teaching Framework: Mechanics, Measurement, Outcomes
Step 1: Teach the mechanics (in plain business terms, not a tutorial)
Your team doesn’t need to understand transformer architecture. They need to understand cost shape. Three concepts cover 80% of what they need:
Chat vs. agent vs. scheduled automation. A chat is a single conversation turn: cheap, discrete, low-stakes. An agent is a multi-step reasoning loop with tool calls that can run 50–200+ exchanges to complete a single task. A scheduled automation runs on a clock whether the output is needed or not. Three completely different cost profiles. Most teams treat them as the same thing.
Using cache. When the same context (the same system prompt, the same background document) gets loaded repeatedly, the model can cache it and read it back at a fraction of the cost. Teams that know this build workflows with stable context up front. Teams that don’t will reload everything from scratch on every run.
Why “running in the background” isn’t free. The most expensive AI costs are the invisible ones. Jobs nobody’s watching. Agents without kill switches. Gemini loops that don’t stop. This is the new shadow IT. Except instead of unsanctioned software licenses, it’s unsanctioned token burns.
Step 2: Show real numbers
Abstract statistics don’t change behavior. Real ones do. I showed a manager his team’s usage numbers and he was floored. He thought everyone used AI like him. They were doing 3 chats a week.
Pull three workflows your team is actually running. Show what each one costs per run. Show the monthly total at current usage. Let people see it.
This isn’t a gotcha. It’s education. Most people are genuinely surprised by the numbers.
If you don’t have this visibility yet, fix that first. You can’t teach cost literacy with costs you can’t see. Stand up an LLM gateway, an observability tool, or even a simple shared dashboard before you run the session. Without attribution, you can’t prioritize. And you can’t educate.
Step 3: Tie it to business outcomes, not compliance
“Be more careful with AI” shuts people down. “Here’s how to build AI workflows that are faster, more reliable, and don’t produce surprise bills at month-end” opens people up.
Show the team what optimized looks like. Walk through a workflow you redesigned for cost, and show that the output quality held. Run the before/after numbers. When people see they can do better work at lower cost, behavior follows. The frame isn’t frugality. It’s mastery.
Where to Start
Pick three real use cases from your organization. Measure what each one costs per run today. Run one working session with the people who own those workflows. Cover the three mechanics, show them the numbers, and walk through how you’d redesign each one.
Measure again after 90 days.
No vendor. No LMS. No culture initiative. Three workflows, one session, one quarter.
Open source can be part of your eventual stack. It’s 50% of my personal AI usage but only 2% of my team’s. But it’s not the first lever to pull, and swapping models without teaching your team changes the price of the problem, not the source of it.
Your inference bill isn’t a model problem. It’s not a CEO problem. It’s an education problem. And that one, you can fix.
If you want the wider map of where AI actually pays for itself across your marketing function, that’s what I keep current on the AI marketing hub.
Hit reply and tell me where you’re stuck.
Alec
Stop paying for AI nobody understands
Once a week I send the real cost breakdowns: what a chat costs versus what an agent costs, which teams are burning budget on jobs nobody’s watching, and the training moves that actually bring the bill down. No hype, no upsell.