The Wall Street Journal recently described how CIOs are managing a fast-rising AI bill by reaching for the cost-control playbook they sharpened during the cloud era: usage dashboards, monthly reports to the CFO, showback and chargeback to departments, token caps per team, model routing, smaller and open-source models, and a dedicated FinOps owner.
Every one of those is sensible. Together, they will not solve the problem.
The reason is buried in the article's own numbers. According to research it cites from Bain & Company, the price of a model token fell by roughly half between December 2024 and December 2025 — while the number of tokens consumed grew 4.5×. Prices went down. The bill went up.
And that is the calm before the agentic wave. Goldman Sachs, also cited, estimates that asking an AI agent to complete a task can take up to 50× the compute of asking a chatbot a question, and projects agent token consumption rising 24× over four years and 55× by 2040.
You cannot dashboard your way out of a curve like that.
The cloud toolbox meters the spend — it doesn't change the workload
Look closely at what each tactic in the story actually does.
| Tactic (from the WSJ) | What it does | What it can't do |
|---|---|---|
| Dashboards, showback, chargeback (OpenText reports 20–30% savings) | Visibility + accountability — teams see and own their spend | Change what the spend is for |
| Per-team token caps (Qualcomm) | Throttle consumption | Separate waste from value — a cap on tokens is also a cap on output |
| Model routing, smaller / open-source models, own hardware | Lower the price per call | Out-run 4.5× volume growth by getting cheaper per call |
| A FinOps owner (Smartsheet; CVS is hiring for it) | Someone owns the budget and the alerts | Own the architecture that generates the calls |
All of it operates on the meter — the price per token, the tokens per user, who gets billed. None of it asks the prior question: should this token have been spent at all?
The lesson the cloud era actually taught — and the story left out
Cloud FinOps did not get its biggest wins from dashboards. It got them from re-architecture. The dashboard found the waste; deleting idle instances, right-sizing over-provisioned services, and moving batch jobs off always-on compute removed it. Visibility was step one. Changing the workload was step two — and step two is where the money was.
AI cost control is stuck on step one.
The AI equivalent of “delete the idle instance” is blunt: stop running a language model for decisions that never needed one.
Where the repeatable spend hides
A large share of agent tokens goes to re-deciding the same thing. Is this customer eligible? Which team should this contract route to? Does this discount need approval? Is this transaction taxable here? Does this claim meet the policy?
These are not open-ended reasoning problems. They are policy, applied. But in a typical agentic design the agent retrieves the policy, reads the relevant passages, reasons through the criteria, produces a result, and explains it — and the meter runs for every step. Then the next case arrives, and it pays to do the identical work again.
That is the 4.5× compounding, one workflow at a time. It is the most expensive way imaginable to apply a rule you already know.
Sertainly: take those decisions out of the token economy
Sertainly draws a hard line between two kinds of work.
Interpreting a policy — reading a regulation, structuring the rules, drafting tests, handling a genuine edge case — is exactly where AI earns its cost. So Sertainly uses AI at build time to help turn source material into a tested, versioned, deterministic decision package.
Applying that policy at runtime is not open-ended, and Sertainly does not spend a model on it. The agent calls a decision API and gets back a structured result — classifications, routes, required evidence, calculated values, reason codes, and a trace — with no LLM inference in the path.
Reason once. Decide millions of times.
The cost consequence is the part FinOps can't get from a cheaper model: the spend stops scaling with volume. A compiled decision executes at ordinary software cost — flat whether it runs once or a billion times. On a showback report, the single largest repeatable line item, “policy decisions,” drops toward zero and stays there as usage grows.
This isn't instead of FinOps. It's the highest-value move in it.
Nothing here argues against the CIOs in the story. Dashboards, ownership, routing, and showback are real discipline, and they should stay. Sertainly is the step after the dashboard: once showback tells you where the tokens go, it lets you remove the biggest repeatable line rather than just watch it.
It also answers the hardest question in the piece. Priceline's Chris Reed said the difficult part isn't tracking the spend — it's knowing the business outcome attached to it. A variable-cost LLM call gives you neither a knowable unit price nor a governed, inspectable output. A deterministic decision API gives you both: a fixed cost per call and an auditable result you can point to. You stop guessing whether the spend was worth it.
The bonus the meter can't price
The 4.5× is not only volume. Part of it is agents talking to agents, retrying, and re-explaining — variance a dashboard records but cannot fix. Deterministic decisions don't drift between prompts, model versions, or vendors. They're versioned, tested, and auditable, so the same case yields the same answer tomorrow. Cheaper and governed — not one at the expense of the other. (We cover how a governed decision survives a rule change in When the Rules Change.)
The question the story doesn't ask
The executives quoted in the Journal are right to bring cloud discipline to AI. What's undersold is that the deepest cost lever isn't a smaller model or a tighter cap. It's architectural: deciding which work belongs on the meter at all.
For open-ended reasoning, keep paying for intelligence — it's worth it. For the repeatable application of policy, stop metering it. Compile it once, and let it run.
Reason once. Decide millions of times.
We made the broader architecture case in The Coming AI Cost Explosion; this is the answer to the specific tactics enterprises are deploying right now.