From November, IBM will pass token consumption into the client's bill of materials. The logic, from IBM Consulting's Andy Baldwin, is hard to argue with: you can absorb $100 million of cloud cost, but not $600 million. Somebody has to carry the variable cost of AI, and from November, IBM has decided it won't be IBM.
That sentence should stop every carrier executive cold. Not because IBM is wrong, but because every carrier buying agentic AI is about to face the same decision, and almost none of them have priced it.
A pilot hides consumption. Volumes are small, the vendor absorbs the tokens, a handful of users poke at a handful of workflows. The business case gets written on that picture: fixed effort, fixed price, a clean ROI slide. Production destroys the picture. Every claim, every submission, every broker question runs through models, and every run burns tokens. The cost that looked fixed turns out to be variable, and it scales with your success. The more the system gets used, the bigger the bill — which is exactly what you wanted, until the bill arrives.
Tokens behave like claims. That is the sentence that makes the whole problem governable, because carriers already know how to govern a variable cost. A loss ratio is frequency times severity, reserved for, watched monthly, repriced when it moves. Token cost has the same two levers. Frequency is how many decisions run through models. Severity is how many tokens each decision burns: long context windows that re-read the entire file at every step, agents that retry instead of stopping, the flagship model doing a routine job a small model could do for a tenth of the price. Nobody has reserved for either lever, because nobody has been asked to. The meter was never part of the pilot.
Here is the other side's ledger, stated fairly. Vendors have a real reason to move you to consumption pricing. It aligns what you pay with what you use, and it scales with the value you get. That is not villainy, it is economics. But it means the buyer has to bring the counterweight, and the counterweight has a name: FinOps guardrails, sold with the consumption model, not after it. That is HFS's own warning about the IBM move, and it applies to every carrier deal, not just IBM's. A consumption model without forecasting, caps, and routing rights is not a pricing model. It is an open tab.
Model routing is the cost lever hiding inside the technology story. If around 80% of tokens run on open-weight models by 2030, as Arvind Krishna expects, then the cheapest adequate model — not the flagship — does most of the routine work. Routing is not an architecture detail. It is the difference between a token bill you can forecast and one you can't. The carriers who treat model selection as a cost decision will have a different P&L from the ones who treat it as a capability decision.
So who carries the token risk today, in your organization? Ask the question in a room and watch what happens. The CIO points to the vendor. The vendor points to the contract. The CFO points to a line item that doesn't exist yet. If no one owns it, the answer is that the enterprise carries it by default, and discovers the size of it in the bill.
The operator moves
Meter the pilot. Instrument tokens per decision and per workflow before anything scales. A pilot without a meter is a press release. You cannot govern what you never measured, and you cannot renegotiate a price for consumption you never counted.
Write the token clause. Caps, overrun ownership, and the right to route work to cheaper models — negotiated before signing, not after the first bill. If the contract prices the outcome but not the consumption that produces it, you have bought the upside and rented the risk.
Name an owner. Someone reconciles AI consumption to the ledger every quarter, the way IBM reconciles transformation value to the books. Token cost is a financial metric now. It needs a financial owner, not a technology owner who sees it as someone else's budget line.
Design for cost. Route routine steps to small, cheap models. Cache what repeats. Make retries a budgeted exception, not a default behavior. The most expensive token is the one an agent burned re-reading a file it had already read.
What carriers can actually apply
The moves above are generic on purpose — token risk hits every industry the same way. Inside a carrier, they land on specific workflows:
- Put a token budget on each workflow, not on AI in general. Cost per claim touched, cost per submission cleared, cost per broker call handled. A single enterprise token budget tells you nothing; a unit cost per decision tells you where the money goes — and which workflow to fix first.
- Tier your models by decision, not by habit. Routine extraction and summarization go to small or open-weight models; only the genuinely judgment-heavy steps — coverage questions, large-loss triage, borderline underwriting — earn the flagship. Your decision inventory already classifies decisions; add a token tier to each one.
- Cap the agent loop. Every agentic workflow gets a maximum step count and a retry budget. When the agent hits the cap, it hands to a human with a clean summary instead of burning another thousand tokens re-reading the file. Retries are a cost center with a human fallback, not an infinite loop.
- Demand an itemized meter from vendors. Not a monthly lump sum — tokens per workflow, per model, per decision type, reconciled quarterly to the outcomes ledger. If a vendor can't show you the meter, you can't govern the cost, and that belongs in the contract before signing.
- Make token spend a standing item in AI governance. The same council that approves agents for production reviews consumption monthly: which workflows grew, which models got more expensive, which routing change saved money. Token cost is a standing agenda item, not a year-end surprise.
- Gate scale-up on a token budget. No workflow moves from pilot to production without a forecasted token cost per unit, a named owner, and FinOps sign-off. That gate is what keeps a successful pilot from becoming an unbudgeted production bill.
Every carrier knows what happens when a variable cost runs without a reserve. It runs fine until it doesn't, and then it runs through the quarter. Tokens are about to become the first variable cost of the AI era. The only question is whether anyone is watching the meter while it runs — or whether the meter is watching you.
— Krishna