AtScale for AI Gives Enterprise AI Agents Governed Business Metrics

Enterprise AI agents just got a new accuracy layer. AtScale for AI, a semantic-layer suite announced on October 15, 2026, is designed to sit beneath AI agents and give them the same governed business metrics that human analysts already use. The pitch from Boston-based AtScale is blunt: buying a smarter model cannot fix an enterprise's definitions problem, because no model can reason its way to what "revenue" means inside a specific company.

According to the company's announcement, carried by Business Wire, AtScale for AI helps enterprises define what business questions mean, compute trusted answers, and continuously test and improve AI accuracy. AtScale for AI is powered by the company's AI Computation Engine (ACE) and combines four connected capabilities: AtScale CoModel, AtScale Ava, ACE itself, and AtScale Canary. The launch coincides with a relaunched AtScale website reflecting an expanded focus on trusted enterprise AI.

The accuracy gap the launch exposes

Alongside the product launch, AtScale published research that quantifies how bad the definitions problem really is. In tests using BIRD-Interact, an independently developed benchmark with 600 tasks across 22 databases, leading AI models answered fewer than four in ten business questions correctly when querying the database directly — even with the schema, column documentation, and a business glossary provided, according to AtScale's research as reported by BigDATAwire. Adding the AtScale for AI semantic layer doubled accuracy on average across all seven models tested, with gains ranging from 1.7x to 2.4x on the 410 business questions in the set.

One finding stands out for agent builders: an open-weight model kept pace with frontier models once the AtScale for AI semantic layer was in place, suggesting that the binding constraint on enterprise accuracy is context, not raw model capability. The benchmark's design mirrors real enterprise workflows, with ambiguous questions, messy data, and simulated-user conversations, as reported by BigDATAwire.

Early enterprise adopters describe meaningful jumps. In banking, where the company's data leadership says an answer that is almost right is still wrong, HSBC reported that accuracy on AI-driven deterministic analytics on a 10 TB table in Google Cloud leapt from 65 percent to better than 99 percent once AtScale served as the semantic layer, per the announcement. At Carrefour, where a semantic layer already governs thousands of KPIs, the retailer's data and analytics director said extending that foundation to AI means agents now use the same definitions as analysts and BI tools instead of producing different answers depending on how a question is asked.

Four components, one feedback loop

AtScale for AI is built as a continuous improvement loop, according to the announcement. AtScale CoModel uses AI to help teams create and maintain semantic models, translating business context into shared definitions of metrics, dimensions, and relationships, with proposed changes subject to human review so definitions stay current as the business evolves.

AtScale Ava — short for Ask, Verify, Answer — connects natural-language questions to the intended governed metrics before computation begins. When a user asks about "revenue," Ava can clarify whether the intent is bookings, recognized revenue, or contract value, then pass the resolved request to ACE for computation.

ACE then applies the governed metric definitions and calculation logic to enterprise data, producing consistent results across AI agents, applications, and analytics tools while optimizing query performance and computational cost. AtScale Canary closes the loop by evaluating AI answers against validated results, using a framework based on the BIRD-Interact benchmark and the enterprise's own question-and-answer pairs, to identify where answers go wrong and track regressions over time.

Questions handled by Ava and errors flagged by Canary can feed back into improvements through CoModel, with human review governing every change to business definitions and logic. AtScale Ava and ACE are available immediately; CoModel and Canary are listed as coming soon. The company is hosting a webinar on the BIRD-Interact results, titled "Make Your AI Show Its Homework," on October 28 at 2 p.m. Eastern.

Why semantics is becoming agent infrastructure

The launch lands in the middle of a broader enterprise-agents reliability push. Agents that can act in production need two things: verified permission to act, and shared definitions of what is true. The first half of that equation is being built through identity and governance tooling — signed agent credentials such as ServiceNow's AI Passport now verify what each agent is authorized to do. AtScale for AI takes on the second half: when agents reason over financial, operational, and sales data, they need the numbers they produce to match the numbers the business runs on.

The same logic applies to the fast-growing layer of data-science agents moving into enterprise analytics teams. A no-code agent that builds machine learning workflows is only as trustworthy as the definitions beneath it, and a semantic layer that turns accuracy into something measurable — and regress-able — is the infrastructure that makes that trust verifiable.

AtScale is also committing AtScale for AI to open standards. The announcement notes support for Apache Ossie, formerly known as Open Semantic Interchange, a shared semantic standard intended to help organizations preserve business logic as they adopt new AI models, agents, and data platforms. Full results and methodology for the benchmark are available in the accompanying AtScale research announcement, and the launch details are covered in full at BigDATAwire. One caveat worth keeping in mind: the benchmark research was company-run, so independent replication will be the real test of the headline gains. But the direction of travel is clear — the agent stack is getting a shared brain for what the business means, and accuracy, not just capability, is becoming the product.