Ask an enterprise assistant for last quarter’s revenue and the grammar is easy. The accounting question is not. One system may store signed contract value, another invoices, another cash receipts and the ledger recognised revenue. All four can be attached to the word “revenue” in ordinary conversation. A fluent model can choose the wrong one, generate a valid query and return a number that looks professionally finished.[1]
AtScale’s new AtScale for AI suite is built around the proposition that this is not primarily a model-intelligence problem. It is a business-definition problem. The company combines a conversational clarification layer, a governed computation engine, an AI-assisted semantic-model builder and an evaluation system intended to compare answers with validated results.[1]
Four numbers can all be true
A hypothetical £12 million software contract
| Measure | What it records | When it changes | Why an AI could confuse it |
|---|---|---|---|
| Bookings | The signed commercial commitment | When the contract is executed | Sales teams may call this revenue before accounting recognitionSource 1 |
| Annual contract value | The contract value normalised to one year | When term or recurring value changes | Useful for recurring-business analysis but not a ledger measureSource 1 |
| Recognised revenue | The amount earned under accounting rules during the period | As obligations are satisfied | This may be far below the total contract value in an early quarterSource 1 |
| Cash collected | Money received from the customer | When payment clears | Timing can lead or lag recognised revenueSource 1 |
A semantic layer records which measure is meant, how it is calculated, which dimensions may be used and how time, currency, hierarchies and access rules behave. AtScale’s ACE then executes that governed logic rather than asking each model to invent a fresh query. The distinction is important: metadata that says “retention is a percentage” does not prevent a tool from joining the wrong population or counting duplicate rows.[3]
The dangerous errors do not crash
AtScale reports that an AI-drafted semantic model, followed by engineering review and tuning, enabled Claude Sonnet 5 to answer 296 of 317 answerable questions correctly in the best of three runs. The company’s accompanying analysis emphasises that the misses often executed without an error. That is precisely why semantic mistakes are difficult: a database can return a perfectly valid result for a badly specified business question.[2]
The new suite assigns a different job to each component. Ava is meant to clarify the question. CoModel drafts and maintains definitions for human review. ACE applies the governed calculation. Canary is intended to test answers against validated question-and-answer pairs and track regressions. Treating these as separate stages makes it easier to locate whether a failure came from language interpretation, business logic, data computation or evaluation.[1]
What is available now
| Measure | Value | Context |
|---|---|---|
| Ava | Available | Conversational clarification and access to governed definitionsSource 1 |
| ACE | Available | Execution of governed calculations and query optimisationSource 1 |
| CoModel | Coming soon | AI-assisted drafting and maintenance of semantic modelsSource 1 |
| Canary | Coming soon | Repeatable evaluation against validated enterprise answersSource 1 |
The broader lesson survives any one vendor. Enterprise AI needs a place where business meaning is explicit, versioned and reviewable. The language model can help draft that meaning, ask for clarification and explain the result. It should not be allowed to silently decide whether “revenue” means the sales target, the contract, the invoice, the cash or the ledger. Intelligence can improve the conversation. Governance determines what the number is allowed to mean.
Sources & further reading
- AtScale launches AtScale for AI to close the enterprise AI accuracy gap AtScale · 15 Oct 2026 · Accessed 16 Oct 2026
- Stop writing your semantic model. Start reviewing it. AtScale · 14 Oct 2026 · Accessed 16 Oct 2026
- You defined your metric once. So why do not your AI tools agree? AtScale · 14 Oct 2026 · Accessed 16 Oct 2026