I was working with a client earlier this year who had budgeted a specific number for AI spend for the year. Reasonable number. Built off real usage. Then we went through the exercise of upgrading them from a team plan to an enterprise plan, and after a single month of usage data it was obvious we were going to land around 3x the original budget.
They didn’t add a single user. Not one. The only thing that changed was that the pricing model moved from seats to tokens.
So when 1Password’s CFO published a post admitting his company got an AI bill 5x the value of the contract they signed, I recognized it immediately. It’s the same failure, just further down the road.
How they get you in
The path is pretty consistent. Somebody starts with an individual license at $20 a month. A heavier user upgrades to a max plan at $200. Then the organization notices people are using it, so they move to a per-seat team plan — $20, $30, maybe $40 a person. All of this is still budgetable. If it’s $20 a month and you have a thousand people, you know your number.
Then you hit the ceiling of that plan, or someone sells you the enterprise tier, and you go from seat-based pricing to token-based pricing. That’s the moment the math changes, and most finance teams don’t register it as a change at all. It looks like a routine upgrade on a bigger contract.
What actually happened is that your contract value stopped being a ceiling and became a floor.
The gate is measuring the wrong thing
Go look at your approval matrix. Almost every one routes a contract to Finance based on the dollar figure on the signature page. Under some threshold, a manager signs. Over it, the CFO does.
At 1Password, the original AI contract came in under that threshold. It cleared without ever reaching the CFO. By the time it did reach him, it was in the form of a follow-on procurement request for a million dollars and the tool was already so embedded in how people worked that turning it off wasn’t a real option.
That last part is the one people underestimate. Once AI is woven into daily procedures, telling someone they’re out of tokens on the 15th of the month is not a budget conversation. It’s a fight. I’ve been on calls where employees accused the company of being cheap because it wouldn’t give them unlimited tokens to keep regenerating slide decks.
Your pilot data is lying to you
The input that drives consumption spend isn’t headcount. It’s the adoption curve and the adoption curve has two dimensions, not one. It’s how many people have access, and how heavily each of them uses it during the day. I’d expect both to keep climbing for the next couple of years.
There’s a version of this showing up in the burnout conversation right now. Before AI, a recruiter might do four interviews a day. They had to prep, take notes, write everything up. AI absorbs a lot of that admin work. So now they’re doing seven. Their AI usage didn’t stay pinned to the original four interviews. It scaled with the new workload.
Which means if you’re extrapolating your annual budget from pilot-program usage, you’re not forecasting. You’re reading the floor.
Why the cap is the wrong instinct
When the bill spikes, everyone reaches for the same lever: cap it per user. I understand the impulse. If you can’t see usage, a cap is the only tool you’ve got.
But it’s a meter you don’t really control. Most of these tools only meter to a monthly limit, they can’t do daily or four-hour throttles the way seat licenses can. So nobody knows the allowance is gone until it’s gone, unless you’ve built real monitoring and actually communicate it. And cost is disconnected from experience: burning three times the tokens feels exactly the same to the user as burning one.
More importantly, capping a tool people depend on doesn’t make them more efficient. It tells them to stop working, and nobody stops working. They expense a personal subscription, or move the task to a free tier nobody approved. Your spend line goes down. Your actual spend doesn’t, it just left the perimeter where you had visibility and a data processing agreement.
What actually works: monitoring, then education
Monitoring isn’t the control by itself. Monitoring is what makes education possible, and education is where the savings live.
The biggest single lever is model selection, because almost nobody chooses a model. They use whatever the interface hands them. At one client we spent serious time getting people off Opus and onto Sonnet for work that never needed the frontier model. They were spending far more without knowing it, purely because of a default. Claude even made this harder than it needed to be: the default was set separately in the desktop app, the phone app, Excel, and Word, so changing it one place didn’t change it anywhere else.
Now we segment. Separate sessions for people still on Opus, for people not using the tools at all, and for people who look like they’re using them well. Every session runs both directions. We teach, and we ask why they’re doing what they’re doing so the training gets better. Spend comes down measurably, because people work a little smarter.
One honest note on the source: 1Password sells a tool that solves the problem the post describes. There’s marketing in there…we made a mistake, we built a product so you don’t have to. Take the diagnosis, which is credible and comes from a sitting CFO. Discount the remedy.
Key Takeaways
- On a consumption contract, the signed dollar value is the entry fee, not the commitment.
- Route contracts to Finance and IT based on pricing structure, not dollar amount. If it bills by token, credit, or usage, it gets reviewed — at any size.
- Model expected usage before signing: 30 days, and 90 days at full departmental adoption.
- Hard caps don’t reduce spend. They relocate it somewhere you can’t see or govern.
- Model selection is the highest-value thing to teach. Most people are on a default they never chose.
- Expect the curve: a spike when a new group gets access, a drop as education lands, then a rise again as adoption deepens.
- Whatever you’ve budgeted, it’s low. On consumption pricing, your budget is a function of how well the rollout goes.
If you change one thing this week, change the trigger. Route on how it bills, not on how big it is. That single line would have caught 1Password’s invoice before anyone signed.
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: The Invoice Nobody Approved
Leave a Reply