Intercept.
The question arrives in plain language, from an AI assistant or from an agent over MCP, before it reaches any query engine.
PRODUCT : semantic layer : Atscale AVA (ask, verify, answer)
Clarify the question before we compute the answer.
AMBIGUOUS BUSINESS QUESTIONS
How does Ava resolve ambiguous business questions?
AtScale Ava (Ask, Verify, Answer) maps each question to the certified calculation before answers are given, and prompts the user when a term could mean more than one thing. Humans reach Ava from any AI client, and AI agents reach it through MCP.
THE PROBLEM
“What was our revenue in Q3?”
doesn’t have one answer.
It could mean bookings, recognized revenue, ARR, TCV or net sales, and each is a different number.
Every one of those metrics is correct for someone. Finance closes the books on recognized revenue, sales runs on bookings, and the board watches ARR. Hand an LLM that ambiguity and it’ll pick one without asking which you meant or telling you which it chose.
The answer looks right, but it answers a question you didn’t ask.
LLM accuracy drops 39% when a prompt is underspecified.
Prompts are often short and vague, like “show me revenue.” Research from Microsoft and Salesforce found that LLM accuracy drops 39% on average when a request arrives underspecified, and models rarely recover from the first wrong reading. Ava exists to fix that. It asks people one clarifying question before anything runs, and it gives AI agents certified metrics to call by name through its APIs.
how ava works
Ava asks which revenue you meant.
When a question could mean more than one thing, Ava shows you the matching metrics and asks which one you mean before ACE calculates the answer. That choice can then be approved into the semantic model, so the next user gets the right answer automatically.
Intercept, resolve, hand off
Resolve.
Ava matches it to one metric and its dimensions in the semantic model. When more than one metric fits, it asks.
Hand off.
ACE computes that metric in your warehouse, and the answer comes back with the name of the metric that produced it.
FOR AGENTS
Ava for AI agents
An agent can’t stop and ask a person which revenue it means. Instead, it calls Ava’s APIs through the AtScale MCP server. One call finds the certified metrics that match its terms, by their business descriptions. Another sends the query to ACE to compute. Each agent runs under its own identity and never writes SQL against raw tables.
the proof
How a clarifying question fixes a
BIRD-Interact answer.
On BIRD-Interact, the public benchmark, a user prompts for “a ranked list of stars in the galaxy.” Without AtScale, the model guesses alphabetical, then size, and runs out of budget. With AtScale, Ava asks which ranking the user means, and a single query gets it right. Across the benchmark, adding an AtScale semantic layer made an LLM twice as accurate, and most of that lift came from asking before computing.
FAQs
Is this text-to-SQL?
No. Ava resolves a question to one defined metric, and ACE, the AI Computation Engine, computes it. Text-to-SQL has a model write a new query against raw tables for each question.
What makes a question ambiguous to Ava?
It’s ambiguous when more than one metric in the semantic model matches the word someone used. “Revenue” can match bookings, recognized revenue, ARR, TCV and net sales. Ava lists every metric that fits and asks which one you’re after.
What happens when Ava can't tell?
It asks which metric you meant before ACE computes anything. A clarifying question takes a few seconds, and a wrong number can cost far more.
Can I see which metric Ava picked?
Yes. The answer comes back with the name of the metric that produced it, so a person or an agent can see exactly what’s been computed.
Does Ava remember my answer?
Yes, once your team approves it. Each clarification goes back to Semantic Modeling, where your team decides whether it becomes a rule. Approved rules apply to every person and agent after that.
What does an agent do with a vague prompt?
It applies the rules your team already approved through Ava. Through the AtScale MCP server, the agent finds metrics by their business descriptions, requests one metric by name, and ACE computes it.
Does it work with my agent?
Yes. It works with any client that supports remote MCP servers, including Claude, ChatGPT, Google Gemini, Databricks agents, Codex, Cursor and agents you build yourself. Agents sign in with OAuth, and your access rules apply to them the same way they apply to people.
Do I buy Ava separately?
No. It’s part of AtScale, alongside ACE, the MCP server, Semantic Modeling and Governance.