HSBC says that putting a semantic layer under its AI agents lifted accuracy on deterministic analytics from 65% to better than 99%, measured on a 10 TB table in Google Cloud. The layer in question is AtScale's, and the company has now packaged the approach as AtScale for AI.
AtScale's argument is that enterprise AI has a business logic problem. A question such as “What was our revenue last quarter?” needs an agreed definition of revenue, the right reporting period and a governed calculation, and a more capable model cannot supply definitions the enterprise never made explicit. The suite, announced on 15 October 2026 and powered by the company's AI Computation Engine (ACE), has four parts.
AtScale CoModel uses AI to create and maintain semantic models, with proposed changes subject to human review. AtScale Ava (Ask, Verify, Answer) connects a natural-language question to the intended governed metric before computation begins, asking whether “revenue” means bookings, recognised revenue or contract value. ACE applies the governed definitions and calculation logic to enterprise data and returns consistent results across agents, applications and analytics tools. AtScale Canary evaluates AI answers against validated results, using a framework based on the BIRD-Interact benchmark and the customer's own question and answer pairs, and tracks regressions over time.
Chief executive Chris Lynch said: “The AI industry keeps telling enterprises to buy a smarter model, but a model can’t reason its way to your definition of revenue.” Dr. Kevin McLafferty, Head of Data & Analytics at HSBC, said: “In banking, an AI answer that’s almost right is still wrong.”
Carrefour is also quoted. Nicolas Treanton, its Data & Analytics Director, said the retailer already has a semantic layer governing thousands of KPIs and wants its agents to use the same definitions as its analysts and BI tools.
Ava and ACE are available now; CoModel and Canary are listed as coming soon. AtScale published a separate benchmark in which adding its semantic layer improved accuracy across all seven LLMs tested on BIRD-Interact, and will host a webinar on the results on 28 October 2026 at 2pm ET. The launch comes with a relaunched AtScale website.
To stay across the latest in cloud, AI and enterprise tech analysis from Compare the Cloud, subscribe to our weekly newsletter at https://www.comparethecloud.net/newsletter
