WHY ATSCALE
AtScale: The engine that makes AI accurate & efficient.
AtScale is THE leading universal semantic layer and computation engine for AI. It helps you model your business semantics, measure if they’re computed right, and compute them every time an analyst using an LLM, an AI agent, or BI tool asks a business question. Then repeat so the model never drifts.
Model, measure, compute, repeat. Get the right answer ever time. That’s the AtScale way. That’s the AtScale flywheel.
Model, Measure, Compute, Repeat: The AtScale AI Accuracy Flywheel
five reasons AI GUESSES
AI gets business questions wrong; AtScale get them right.
Your AI never says “I don’t know,” even when it should.
Three are five reasons LLMs get enterprise answers wrong. AtScale addresses all five.
| Why AI gets it wrong | What happens |
|---|---|
| Vague prompts and agentic computing | Analysts prompt, “What was revenue last quarter?” but don’t specify which view of revenue they mean: Bookings? Net sales? Recognized revenue? Gross revenue? They’re all correct answers, and each has a different number. The model picks one and doesn’t tell you which. AtScale closes this accuracy gap with Ava, which uses its knowledge of your metrics to disambiguate prompts and supply only attested metrics to automated AI agents. |
| Vague semantics | Even when you mean recognized revenue, the model doesn’t know which column it needs, or how to compute it, merely by inspecting your warehouse, or that finance excludes “intercompany sales.” A properly crafted semantic layer and attested, measured metrics close this accuracy gap. AtScale CoModel and Canary, which measures LLM responses based on a Golden SQL definition, closes this accuracy gap. |
| Inconsistent and complex data | Your gold data layer is clean, but it wasn’t built for AI to read. One customer is spelled three ways across source systems, a column called amt_net holds recognized revenue, and each row in the orders table is a line item, not an order. The model guesses what each one means, and revenue gets counted twice. AtScale resolves this at the data layer; as you build your layer, you specify the precise way AtScale should compute metrics under the guidance of your data engineers that know your data. AI gets the right answer, every time. And when schemas change, AtScale Canary keeps measuring accuracy and flags it to you, so computations don’t drift. |
| Multi-pass calculations | Recognized revenue takes several passes: pick the contracts that count, spread each one over its service period, convert currencies, then roll up by region and quarter. Get one pass wrong and you still get a number, just the wrong one. AtScale’s multi-pass computation engine has 13 years of computing these real-world, complex, multi-pass metrics. |
| Unchecked answers | Do you measure the accuracy of your AI layer? The AtScale family does. Our new, innovative AtScale Canary AI accuracy and cost system flags when recognized revenue comes back as $41.8M, because it counted intercompany transfers, and shouldn’t. AtScale Canary is the canary in a coal mine, ensuring AI is safe, accurate, and don’t drift. |
HOW ATSCALE gets IT right
How AI with AtScale gets answers right.
For every reason AI gets it wrong, there’s a part of AtScale that makes sure you get them right.
| Why AI gets it wrong | How AtScale gets it right |
|---|---|
| Vague prompts | AtScale Ava asks, “Do you mean recognized revenue, bookings or net sales?” before anything computes. |
| Vague semantics | AtScale Semantic Modeling defines recognized revenue once and maps it to the right columns, and a person on your team approves it. |
| Inconsistent and complex data | AtScale Semantic Modeling reconciles the messy schemas, values, so every revenue row counts once. |
| Multi-pass calculations | AtScale ACE runs every pass and returns one number for recognized revenue. |
| Unchecked answers | AtScale Canary (coming this fall) tests the answer against the number finance closed the books on, and AtScale Governance versions every change to the definition. |
THE BENCHMARK
A semantic layer makes AI 2X more accurate and the flywheel paves a path to 100% accuracy.
On BIRD-Interact, the public benchmark built to test realistic enterprise questions, adding AtScale made an LLM twice as accurate. A better model alone won’t do that, and neither will an ontology or a skill.
LLM alone
LLM With AtScale
2x
Accuracy on BIRD-Interact answerable questions, Clause Sonnet, with and without AtScale. Preliminary
ATSCALE SEMANTIC MODELING
AI-assisted semantic modeling.
Weeks of modeling, done in minutes.
AtScale CoModel reads your warehouse schema, Power BI dashboards, query history and catalogs, then drafts the semantic model. Data analysts make calls only a person familiar with your business and data can make, and approves every metric before they ship.
That’s the AtScale flywheel: Model. Measure. Compute. Repeat. “Model” shown below:
ATSCALE AVA (ASK, VERIFY, ANSWER)
Prompt disambiguation for humans.
Make our day. Ask a vague question. AtScale Ava (Ask, Verify, Answer) clarifies what you meant.
AtScale Ava uses it’s knowledge of your business semantics to check every question before the LLM can invent an answer.
Ask “What was revenue last quarter?” and Ava checks your layer. It finds multiple definitions of “revenue,” from GRR to ARR. Instead of guessing, Ava stops and tells your LLM to ask: “Do you mean recognized revenue, bookings or net sales?” Putting the human in the loop is the key to effective AI prompting.
ATSCALE MCP SERVER FOR AGENTS
Programmatic prompt disambiguation for AI agents.
Give an agent a vague task and it still gets the right number.
Agent need to think, too. A finance agent told to “draft the quarter-end revenue summary” asks AtScale via Ava’s MCP server to find approved metrics that match. If it’s unsure, it can bring the human into the loop. Once disambiguated, AtScale’s ACE (the AI Computation Engine) compute the metric. Agents sign in through authentication schemes like OAuth, so it only sees what the person it’s acting for is allowed to see. See how AtScale powers AI agents →
ATSCALE ACE (AI COMPUTATION ENGINE)
Multi-pass metric computation.
ACE does the math, so the LLM doesn’t guess.
Most business metrics take more than one pass. Recognized revenue means picking the contracts that count, spreading each over its service period and converting currencies, then rolling up by region and quarter. Get one pass wrong and you still get a number. It’s just the wrong one.
ACE reads those rules from your semantic model and runs every pass the same way each time. It also builds aggregate tables in your platform, so the many queries an AI agent sends don’t all scan raw data.
ATSCALE CANARY / THIS FALL
AI accuracy an cost measurement and attestation.
AtScale Canary will measure how accurate your AI is on your own questions.
AtScale Canary is the industry’s first AI accuracy and cost attestation system. It runs business questions against golden SQL answers your team verifies, scores the results, and sends every miss back to AtScale Semantic Modeling with suggestions to fix it.
Run Canary every day, fix errors, add use cases. Your model never drifts, your metrics stay computed right, and do more with AI. Your AtScale flywheel keeps you on course to enterprise AI that works, and is trusted.
ATSCALE GOVERNANCE
Your source code is under version control; why aren’t your metrics?
AtScale keeps your metrics and your data in step. Change a definition on a Git branch and open a pull request. Your CI/CD pipeline tests the change before it ships. When a column’s renamed or dropped in the warehouse, CoModel flags it, so the model doesn’t break without anyone noticing. Your warehouse permissions still apply to every query, from every person and every AI agent.
WHERE ATSCALE FITS IN YOUR STACK
Freedom of choice: Keep the tech stack you invested in. Be free to change your mind.
Keep your catalog, your ontology and your skills. They describe your data and tell the model how to behave. AtScale computes the number your AI uses, a job none of them were built for.
Book a workshop.
The workshop takes thirty days from kickoff to results.
- Bring us your most controversial metric and the business questions your team and your agents ask about it every week.
- We benchmark your AI on them, with AtScale and without it.
- You keep the semantic model we built to run the test.