Ask a language model what an instance costs and it will answer with confidence and, often, with a number that is close but wrong. That is not a flaw you can prompt away; it is what the tool is. A model predicts likely text, and a price that looks right is likely text. For cost optimization, where the whole point is a figure a CFO will act on, close but wrong is worse than useless, because it is convincing. The fix is architectural, not clever prompting: never let the model generate a number. A pricing engine computes every figure from your usage and a dated catalog snapshot, and the model is confined to explaining what the engine found.
What deterministic pricing means in practice
Deterministic means the same inputs always produce the same output, and every output traces to a source. A pricing engine holds catalog snapshots, list prices captured on a specific date with the source recorded, and applies them to your resources by rule. Ask it the cost of a workload and it does arithmetic you can reproduce: this shape, in this region, at this rate captured on this date, times this quantity. There is no sampling, no temperature, no drift between two identical questions. When a price is wrong, it is wrong in a way you can find and fix, because you can point at the catalog entry. That auditability is the entire value.
So what is the model good for?
Plenty, as long as it never touches the arithmetic. The model is excellent at the work around the number: reading a messy CSV and mapping its columns, writing the one paragraph explanation of why a resource is wasteful, summarizing a run for an executive, answering a plain language question over data the engine has already priced. These are language tasks, and language is what the model is for. The discipline is a clean division of labor. The engine owns every digit. The model owns every sentence. A verdict from the model is machine validated before it is shown, so even its qualitative judgments are checked against the data rather than trusted on faith.
How to tell a real number from a plausible one
When you evaluate any AI cost tool, this is the question that matters most. Ask the vendor to show you where a specific figure comes from. A tool built the right way can trace it: this rate, this snapshot date, this quantity, this arithmetic. A tool that lets the model generate numbers cannot, and will instead talk about accuracy in the aggregate. In a comparison across clouds the difference is stark, because small rate errors compound across hundreds of line items into a ranking that is simply wrong.
Point at any number on the screen and ask the tool to show its source. If the answer is a catalog entry and a date, trust it. If the answer is a confidence claim about the model, do not put it in front of your board.
Why this is a trust differentiator, not a technicality
Buyers are right to be wary of AI in finance, and blanket claims about accuracy do not earn trust. A traceable number does. When every figure on a page can be followed back to a dated list price, the conversation with finance changes from do we believe the tool to do we agree with the assumption, which is the conversation you actually want. This is the same reason we take zero provider commissions: the output is only worth something if its independence and its arithmetic are both beyond dispute. Prices here are list prices from dated snapshots, informational rather than financial advice, and that honesty is the point.
Frequently asked questions
Can a language model calculate cloud costs accurately?
What is a dated catalog snapshot?
Where should AI be used in a cost tool then?
Numbers you can take to the board
Datum is built on one rule: the engine computes every figure, the model only explains it. Tour the platform to see traceable pricing in action, or bring us your estate.
Put a defensible number on your cloud spend.
No provider in the room, no published price list. Tell us your footprint and we will scope the savings against your billing data — we reduce your cloud spend or we reimburse our service fee.
The Cloud Spend Navigator: what changed in cloud pricing, commitments, and FinOps — no vendor spin.