How do you decide what you should actually cut?

The AI line on the technology bill grows every period, finance tracks it the way it arrives, by provider and by model, and the plan to bring it down starts at the top of that list. A cost-reduction plan reaches the budget review with the largest lines ranked first and a share to take out of each, which is a sensible way to cut almost any other cost.

AI features have left the pilot and moved into the product, so the token line now moves with customer use. The cutting has already started, Cloudera's enterprise AI survey, reported by IT Brief, found 89 per cent of organisations globally have limited or redesigned AI use cases because of token consumption, and only 21 per cent of Australian respondents said AI currently returns more than it costs. That second figure cannot be answered in aggregate, a business can say whether AI pays only once it can say which feature pays, and the token bill is not built to tell it.

The CTO who throttles a feature the week its bill doubles is right to do so, that is incident response, and nobody should wait on a model while spend runs away. The question is what happens after the spike, when the throttle becomes a line in next period's budget and a temporary measure turns into a permanent decision about which features the product keeps, which is exactly the point where precision is worth paying for, because it can still change the outcome.

That decision gets made on the bill, because the bill is the only cost record finance holds, and the bill knows provider, project, model, API key and token type, never feature. An inference call is an event triggered by a feature, with no resource behind it to tag, several features often call through one key or gateway so their tokens land in a single pool, and what each feature earns sits in billing and subscription data that is never joined to its token cost. Nobody was careless, the split has to come from how the product is built, and finance was handed the one record that leaves it out.

Cutting AI spend well means cutting it in the product's units rather than the provider's. A provider's unit is a token, the slice of text a model reads or writes, priced by model, a product's unit is a use of a feature, an image generated, a presentation collated, a block of copy written, each of which costs something to serve and earns something when a customer pays for it. The unit economics of an AI feature are what it costs per use set against what it earns per use, and that is the ranking a cut has to be made from.

Take a product carrying four AI features, over one month.

On the bill the copy generator is the obvious target, it is more than half the AI spend, so the plan cuts its per-request token budget by 35 per cent, shorter context and cheaper calls. Shorter context makes weaker copy, and if 40 per cent fewer generations are run and billed as a result, token cost falls to $14,040 and metered revenue to $36,000, the plan saves $21,960 of tokens, gives up $24,000 of revenue, and contribution falls by $2,040. The copy generator is the largest line because it is the most used, 600,000 generations in a month, each one a use a customer paid for. The long-form writer, third on the bill at $9,000, costs $1.80 a document and earns $1.20, and nothing in the plan touches it.

Cut on the bill and you cut volume, not waste.

The plan on the bill still saves more spend, $21,960 of tokens against the $9,000 that retiring the long-form writer would remove, and a CFO with a number to hit may prefer it for that reason. A spend target met by giving up more revenue than it saves is not a saving, it is a smaller business with a smaller AI line, while retiring the long-form writer lifts contribution by $3,000 and removes the only feature that loses money each time a customer uses it. Whether to reprice it instead is the pricing leader's call.

A product leader will object that AI features are table stakes, customers expect them in the product and nobody buys them separately, so revenue per feature is a fiction. Some of it is priced directly, the copy generator is metered and the long-form writer is a seat add-on, which is also why the long-form writer ranks last on its own price against its own cost, whatever its usage. Where pricing attaches nothing, as with template search, usage weights the share of subscription revenue left once the priced features have taken theirs, declared as an apportionment and inspected like one, and an included feature that ranks badly is flagged for review before anyone cuts it, since what it contributes to renewals sits outside the ranking. That is an estimate, it is a far better one than a token count, which carries no revenue at all.

Getting there takes two joins the bill cannot make. Where a feature has its own project, workspace or key, its spend can be attributed to it directly, where features share a key the pooled spend is split by what each one used, which needs request logs carrying a feature identifier, something engineering adds once and finance reads every period after. OPTIMAZE ingests AI spend as the provider reports it, by project, model, key and token type, attributes it to the features that caused it, and reads each feature's unit economics, what it costs per use against what it earns. Caching, routing and shorter prompts lower what a use costs and are worth doing, they move the numbers in the ranking and leave the ranking in charge, and which model a single feature should run on is its own decision.

Before the plan goes back to the budget review, rebuild its ranking in the product's units, each feature's cost per use against what it earns per use, and cut from the bottom of that list. Spikes still go to the throttle list. The question to ask of any proposed cut is the one the bill cannot answer, what does this feature earn each time it is used, against what it costs each time it is used.

That is Technology Capital Performance applied to the fastest-growing line on the bill, AI spend treated as deployed capital and judged on what each feature produced rather than on how many tokens it consumed. The token bill tells you what was supplied, the product tells you what it was worth.

If this sounds like something you or your business needs, get in contact with our sales team.