Seven million dollars. That is the average enterprise AI budget in 2026, up from $1.2 million in 2024, an increase of 483%, according to the FinOps Foundation's 2026 survey of more than 1,100 practitioners. The same survey found that 73% of enterprises exceeded their AI cost projections in the past year.
If your bill is moving that way, you need to know which lever to pull first, and the usual advice ("switch to a cheaper model") is the wrong answer to that question. This guide ranks the levers by what they cost you in accuracy and effort, and says what to check before you trust each one.
First, change the number you watch
Most dashboards report cost per token or cost per call, and both reward the wrong behavior. A cheap model that fails the task and triggers a retry, or hands off to a person anyway, can cost more per finished job than an expensive one that gets it right first time. The FinOps Foundation recommends cost per successful output instead, and we agree with that completely. Track only tokens and the workflows burning money on rework look efficient.
This is why AI spend belongs inside the reporting you already run. If your business intelligence already tracks cost per order or cost per ticket, put cost per completed AI task beside it, and the finance lead can see whether the automation is cheaper than the process it replaced. The API invoice cannot tell them that.
Which lever, given your workload
The table follows Anthropic's and OpenAI's pricing documentation as compiled by Digital Applied. The savings are best cases for teams calling managed APIs, so measure them on your own traffic before they go into a forecast.
| If this describes your workload | Lever | Typical saving | Effort | Check before trusting it |
|---|---|---|---|---|
| The same system prompt or reference document goes out on every call | Prompt / KV caching | Up to 90% on cache-hit tokens | Low | Your actual cache-hit rate |
| Many near-identical questions arrive | Semantic caching | 61-69% fewer API calls | Medium | That similar questions really share the right answer |
| Nobody is waiting for the result (reports, backfills, classification) | Async batching | 50% off sync rates | Low | That the turnaround window is acceptable |
| Most requests are easy and a few are hard | Model cascade / routing | 30%+ | Medium | The router against your own prompts, because one that misjudges difficulty sends hard questions to a weak model |
| You self-host at scale | Quantization | Not covered by the sources above | High | This is an infrastructure decision, not a first move |
Read the table top to bottom, because the order is the advice. Caching goes first because it costs no accuracy: a workflow that sends the same prompt on every call pays full price once and a fraction after. Batching is next for anything that can wait. Routing comes later, because you cannot trust it until you have benchmarked it, and it is the first lever that can make answers worse.
Why waiting for cheaper models is not a plan
Finance teams assume AI gets cheaper. Axis Intelligence Research's tracking of 130 LLM APIs found that frontier model pricing has doubled since January 2026, while the cheapest models matching 2023-era quality fell more than 200x in price over three years. Both hold at once: a floor that keeps dropping and a ceiling that keeps rising, and your bill depends on which one your workflow touches.
That is also how total spend can climb 483% while token prices deflate. The survey does not say why. The likeliest explanation is volume: cheaper tokens put models in more places, and an agent that chains several calls per task multiplies the tokens behind each outcome.
What we would do on Monday
Give one person ownership of the inference line, as you would for any vendor cost, and review it against outcomes on a schedule. Turn on prompt caching wherever it exists. Pick one workflow and compute its cost per successful result, including retries and human handoffs. That number will tell you more than the invoice. For agent workflows, measuring ROI before scaling a pilot belongs in the same conversation.
ThinqHub implements Odoo and ERPNext, and the place we would want that metric to live is next to the operational data in your reporting. If AI spend is growing faster than anyone can explain, talk to us about getting that visibility built before the next overrun.
