Fixing AI's Math Problem: AtScale Launches Semantic Layer for Enterprise Agents

BOSTON, MA – October 15, 2026 – The generative artificial intelligence boom has a well-documented, multi-billion-dollar bottleneck: large language models are fundamentally bad at basic business math. While today's most advanced AI agents can write complex code or draft eloquent marketing copy, they routinely stumble when asked a seemingly simple question like, "What was our recognized revenue last quarter?"

The failure does not stem from a lack of computational intelligence, but rather a lack of context. A model cannot reason its way to a company's bespoke, internal definition of revenue. To bridge this critical enterprise accuracy gap, Boston-based AtScale today announced the launch of AtScale for AI, a comprehensive suite of capabilities designed to connect natural-language questions directly to governed, shared business definitions.

Powered by the company's AI Computation Engine (ACE), the new suite aims to reposition the semantic layer from a useful business intelligence middleware tool into an absolute architectural requirement for reliable AI deployments. The launch, which coincides with a broader corporate rebranding and expanded focus on trusted enterprise AI, signals a major shift in how organizations are attempting to rein in AI hallucinations and deploy autonomous agents at scale.

The Missing Link: Business Logic Over Model Intelligence

For years, the enterprise AI industry has operated under the assumption that scaling up model size and training data would eventually smooth out accuracy issues. However, enterprise data architectures are notoriously messy, siloed, and highly specific to individual corporate cultures.

"The AI industry keeps telling enterprises to buy a smarter model, but a model can't reason its way to your definition of revenue," said Chris Lynch, CEO of AtScale. "Enterprises need to define what an answer means, compute it correctly, and prove it holds up. AtScale for AI brings those capabilities together so accuracy becomes something businesses can measure, improve, and demand."

To achieve this, AtScale for AI introduces a continuous improvement loop across four connected capabilities. The foundation is AtScale ACE, a computation engine that applies governed metric definitions and calculation logic to enterprise data, optimizing query performance without requiring data movement.

Sitting on top of this engine is AtScale Ava (Ask, Verify, Answer), a natural-language clarification tool available today. Before a computation even begins, Ava intercepts the user's prompt. If an executive asks an AI agent about "revenue," Ava identifies the ambiguity and clarifies whether the user means bookings, recognized revenue, or total contract value. Only after resolving this intent does Ava pass the request to the computation engine.

Rounding out the suite are two forthcoming tools: AtScale CoModel, which uses AI to assist data teams in creating and maintaining semantic models while keeping human review in the loop, and AtScale Canary, an evaluation framework designed to track regressions and identify where AI answers go wrong over time.

Real-World ROI: From Guesswork to Precision

The theoretical benefits of a semantic layer are well-documented, but the practical application for AI accuracy has yielded striking results in early enterprise deployments. Because AtScale acts as a universal translator—sitting between the raw data in cloud data warehouses and the consumption tools—it ensures that an AI agent is reading from the exact same sheet of music as a human data analyst.

At HSBC UK Commercial Banking, the implementation of AtScale's semantic layer over a massive 10-terabyte table in Google Cloud resulted in a dramatic performance leap.

"In banking, an AI answer that's almost right is still wrong. The challenge was never model intelligence; it was making sure every agent calculated metrics with precision and accuracy - no exceptions," said Dr. Kevin McLafferty, Head of Data & Analytics for UK Commercial Banking at HSBC. "By building on the governed definitions we already trusted, AtScale gave our agents that context and precision without reinventing the logic behind it... accuracy on AI driven deterministic analytics took a quantum leap in performance from 65% to better than 99%."

Similarly, global retail giant Carrefour has utilized AtScale to govern thousands of key performance indicators across its sprawling business operations.

"AI shouldn't create a second version of the truth," said Nicolas Treanton, Data & Analytics Director at Carrefour. "Extending that foundation to AI means our agents can use the same definitions as our analysts and BI tools, rather than producing different answers depending on how someone asks the question."

To quantify these anecdotal successes, AtScale simultaneously published findings from a companion evaluation benchmarking seven leading large language models—including GPT-4, Claude 2, and Llama 2—using the BIRD-Interact framework. BIRD-Interact, a rigorous open-source dataset for evaluating text-to-SQL models, focuses heavily on realistic business intelligence scenarios. The research demonstrated that routing queries through a semantic layer improved accuracy and reduced hallucination rates across every single tested model, proving that context, rather than raw neural network size, is the key to deterministic accuracy.

The Semantic Layer War: Anchoring Autonomous Agents

AtScale's aggressive pivot toward AI governance arrives at a pivotal moment in the broader data infrastructure market. The semantic layer space is becoming fiercely competitive, evolving rapidly from a niche data modeling requirement into the central nervous system for enterprise AI.

Competitors are approaching the same problem from different architectural vantage points. For instance, dbt Labs has heavily popularized the dbt Semantic Layer, which allows data engineers to define metrics directly within their transformation workflows. While highly integrated into the modern data stack, dbt's focus has historically leaned toward traditional reporting rather than dynamic, out-of-the-box AI agent governance.

Meanwhile, open-source solutions like Cube offer highly performant, developer-centric API gateways for data analytics. Cube excels at building data APIs and customer-facing dashboards, but it lacks the explicit, packaged AI hallucination prevention and natural-language clarification features that AtScale is introducing with Ava and Canary.

On the other end of the spectrum, massive platform vendors are building their own walled gardens. Databricks Unity Catalog provides robust, centralized metadata management and access control, functioning brilliantly as a governance catalog but stopping short of being a universal semantic translator for cross-platform natural language. Snowflake Cortex enables machine learning and AI building blocks directly within the Snowflake Data Cloud, yet it does not serve as a universal, platform-agnostic semantic layer for heterogeneous enterprise data environments.

AtScale is positioning itself as the independent, universal alternative. By connecting to various data platforms without requiring data to be moved into a proprietary ecosystem, AtScale for AI targets the enterprise reality: data is scattered, models will become commoditized, and the only true differentiator will be the proprietary business logic governing them.

Open Standards and the Path Forward

To mitigate the risk of vendor lock-in—a common fear among enterprise architects adopting new data infrastructure—AtScale is heavily backing Apache Ossie, formerly known as Open Semantic Interchange.

Currently an active project within the Apache Incubator, Apache Ossie aims to establish a common, open interchange format for semantic models. Operating under the Apache Software Foundation's strict meritocracy and open-development principles, the project is designed to standardize how metrics, dimensions, and hierarchies are defined and shared across the industry. By advancing these shared semantic standards, AtScale is ensuring that organizations can preserve their core business logic even as they inevitably swap out underlying AI models, autonomous agents, and cloud data platforms in the years to come.

As the enterprise AI narrative shifts from the initial hype of generative text toward the demanding reality of deterministic analytics, the infrastructure supporting these models must mature. AtScale will detail its BIRD-Interact benchmark methodology and further explore this shift in an upcoming webinar, "Make Your AI Show Its Homework," scheduled for October 28.