Enterprises are increasingly discovering that raw model intelligence cannot compensate for a lack of explicit business logic, a gap that AtScale is attempting to bridge with its new specialized software suite. The company announced AtScale for AI, a collection of capabilities designed to link natural-language queries to governed business definitions, ensuring that AI agents do not hallucinate metrics or misinterpret core KPIs. By integrating semantic modeling with automated testing, AtScale is positioning itself as the connective tissue between unpredictable large language models and the rigid, governed data environments required for professional financial and operational decision-making.
AtScale for AI Integrates CoModel, Ava, and Canary
The new suite aims to establish a continuous improvement loop for enterprise AI by addressing the "business logic problem," where models struggle to define terms like "revenue" without explicit instruction. AtScale is deploying this via four interconnected components, two of which are available immediately: AtScale Ava and the AtScale AI Computation Engine (ACE). AtScale Ava functions as a clarification layer, intercepting natural-language questions to verify intent—such as distinguishing between "bookings" and "recognized revenue"—before passing the request to the ACE. This engine then applies governed metric definitions to enterprise data to ensure consistent results across various applications and AI agents.
The company is also preparing the release of AtScale CoModel and AtScale Canary to complete the lifecycle of AI accuracy. CoModel is intended to assist teams in using AI to maintain semantic models, though it maintains a requirement for human review to govern changes to business definitions. AtScale Canary will serve as an evaluation tool, allowing enterprises to test AI-generated answers against validated results using a framework based on the BIRD-Interact benchmark. This allows organizations to identify specific failure points and track accuracy regressions over time. By combining these tools, AtScale intends to move AI from probabilistic guessing toward deterministic, governed computation.
Bridging the Gap Between LLMs and Governed Data
AtScale is positioning this suite as a necessary layer for industries where "almost right" answers constitute a failure, such as banking and large-scale retail. The company highlights the potential for significant accuracy gains, citing a case where HSBC utilized AtScale as a semantic layer on a 10 TB table in Google Cloud to move AI-driven deterministic analytics accuracy from 65% to over 99%. This deployment suggests that the primary hurdle for enterprise AI adoption is not the intelligence of the underlying model, but the availability of a shared, governed foundation that prevents the creation of a "second version of the truth."
Furthermore, AtScale is emphasizing its commitment to an open, independent semantic layer through support for Apache Ossie (formerly Open Semantic Interchange). This move is designed to help organizations preserve their business logic even as they swap out underlying AI models, agents, or data platforms. The company's recent evaluation of seven different large language models (LLMs) using the BIRD-Interact benchmark reportedly showed that adding an AtScale semantic layer improved accuracy across every model tested. This research serves as the technical justification for the Canary evaluation tool, aiming to provide a repeatable method for measuring how well an AI agent understands specific enterprise context.
Key Takeaways
- AtScale for AI introduces a suite of tools including AtScale Ava and ACE (available now) and AtScale CoModel and Canary (coming soon) to connect natural language to governed metrics.
- HSBC reported that using AtScale as a semantic layer on a 10 TB table in Google Cloud improved AI-driven deterministic analytics accuracy from 65% to more than 99%.
- The suite utilizes the BIRD-Interact benchmark framework to help enterprises evaluate AI answers against validated business results and track accuracy over time.
TechInsyte's Take
In our view, AtScale is correctly identifying that the "intelligence" arms race in LLMs is hitting a wall of practical enterprise utility. While providers focus on increasing parameter counts, the real bottleneck for the C-suite is the "semantic gap"—the distance between a user's vague question and a company's audited financial definition. By focusing on the semantic layer rather than the model itself, AtScale is betting that the future of enterprise AI lies in orchestration and governance rather than raw generative power. This approach shifts the responsibility of accuracy from the unpredictable model to a controlled, governed computation engine. If successful, this could move AI from a "chat" interface to a reliable analytical tool that can actually be trusted for deterministic reporting in highly regulated sectors like banking and retail.
Questions & Answers
How does AtScale for AI address the risk of AI agents providing inconsistent metric definitions?
The suite uses AtScale Ava to clarify natural-language questions before computation begins. For example, if a user asks for "revenue," Ava identifies whether the user requires bookings, recognized revenue, or contract value, ensuring the AI agent uses the specific, governed definition required by the business.
What role does human oversight play in the AtScale for AI workflow?
Human review remains a critical component of the semantic modeling process. Through AtScale CoModel, teams can use AI to help create and maintain models, but any proposed changes to the shared definitions of metrics, dimensions, or relationships remain subject to human approval to ensure data integrity.
Can AtScale's solution be used with existing large language models?
Yes. AtScale is positioning its suite as an independent layer that works alongside existing models. The company's research involving seven different LLMs suggests that adding an AtScale semantic layer improves accuracy across various models, and its support for Apache Ossie aims to ensure business logic remains portable across different AI platforms.
How does AtScale measure the success of AI accuracy within the enterprise?
The company is introducing AtScale Canary, which allows enterprises to evaluate AI answers against validated results. This tool uses a framework based on the BIRD-Interact benchmark and the company's own question/answer pairs to identify where answers fail and to track accuracy regressions over time.
Source: AtScale
