Why AI spend is so hard to explain, what it would take to run it like any other operating cost, and how we answer it in three weeks.
The challenge
Ask most organisations what they spend on AI and you will get a number. Ask what drove it and the room goes quiet.
AI spend arrives as a line on an invoice, and usually several invoices, because many firms now buy from more than one vendor. Nobody can say which piece of work drove the total. Nobody can say whether each job is running on the right model, or whether the seat licences on the books are paying for themselves.
Then finance is asked to approve next quarter's figure. It gets set from last month's invoice plus a margin for growth, because there is nothing better to set it from.
The usual reaction is to hunt for a cheaper model. That rarely fixes it, because model price is only one of the levers behind the bill:
- Volume and size. How often each job runs, and how much text goes in and comes out every time.
- Reasoning. Reasoning models think before they answer, and that thinking is billed even though it never appears in the response. OpenAI bills reasoning tokens as output tokens.
- Caching. Many jobs send the same long instructions with every request. At Anthropic, reading that repeated text from the cache costs a tenth of the normal input price.
- Batch. Work that can wait hours rather than seconds can go through a batch route. Anthropic, for example, charges half price for batch work.
- Seats against tokens. A seat is a fixed monthly cost per person. Whether it beats paying per use depends on how heavily each group of people actually uses it.
On top of that, prices move every month. A forecast without the date of the price list behind it is a forecast nobody can check.
The vision
AI should be run like any other operating cost. Every workload has an owner and a reason for the model it runs on. Every price carries the date of the list it came from. Every figure says whether it was measured from your bills or estimated from an assumption, and the assumption is written down.
That is the standard a finance team already applies to everything else it signs. A budget for AI should show what it becomes at half the expected volume and at double, and what a spending ceiling would buy.
Above all, it should be a figure the person signing it can explain to the person who asks. A number you cannot check is an opinion, and AI spend is no exception.
How Analect answers: the AI Spend Forecast
The AI Spend Forecast is a three-week, fixed-price service from ETT, run on Analect. It answers four questions:
- What are we spending, and on what? Your current spend broken down by workload and by model, with what you paid against list price and your effective discount.
- Is each job on the right model? Every workload priced across around 200 models from 27 providers, filtered by your rules on data residency and approved vendors, and by what the job needs. Each pick comes with the reason it won and the models it beat.
- Are the seats paying for themselves? For every seat licence, a verdict against the same vendor's pay-per-use pricing, with the break-even point. We hold the vendor constant, so the comparison is a fair one.
- What should we budget? The recommended monthly and annual run-rate by workload and by vendor, what it becomes at half and double the volume, and what a ceiling buys if you set one.
It runs over three weeks:
- Week 1, gather. We take in your provider usage exports, your seat licences and the workloads you want priced, and agree your rules.
- Week 2, model. A working session where we walk the forecast with your team and push on the numbers live. Your people get their own logins.
- Week 3, report. The report and the budget, presented to your sponsor and your finance lead.
Your billing files are read in your browser and never uploaded. We need no documents, no customer data and no prompts.
Every figure in the report is labelled as measured or estimated, and the date of the price catalogue is printed on it. Where a cheaper model fits a job, the report says what to test before you move anything. It is a forecast, not an instruction to switch.
If you want to try the seats question on your own numbers first, our free seat break-even calculator does it group by group.
If AI spend is a number you are being asked to defend, the AI Spend Forecast is built for exactly that.