Ammonix proposes AI architecture that learns business patterns and acknowledges uncertainty

Swiss startup Ammonix, founded by former Cortex CEO Peter Ruppersberg, has developed an artificial intelligence agent architecture that learns the specific patterns of each organization. The system combines a specialized model, trained on a company’s rules and data, with large language models such as ChatGPT or Claude, which are used for explanation and interaction. This approach enables decisions to be made locally, without relying on external data centers, and provides traceability to meet regulatory requirements.

Two systems that separate decision-making and explanation

The architecture follows the division proposed by psychologist Daniel Kahneman, winner of the Nobel Prize in Economics, between fast, intuitive thinking (System One) and slower, analytical thinking (System Two). In practice, the specialized model acts as System One, identifying patterns and making quick decisions within a specific domain. Meanwhile, the language model functions as System Two, providing context, justifications, and communication with users, without making the final decision alone.

The agent that knows how to say “I don’t know”

When the system encounters a situation that does not fit known cases, it can request more information or refer the decision to a human. Each recommendation is accompanied by the examples and rules that supported it, ensuring auditability. The startup argues that, in critical decisions, admitting uncertainty is safer than presenting a wrong answer with confidence.

No investors and an eye on Brazil

Ammonix still operates with a small team and has been funded with private resources, without investment rounds from funds. The company plans to launch AmmonixCode, an open-source platform for creating agents based on its architecture. Use will be free for research, while companies will pay for commercial licenses, implementation, and consulting. Local execution is one of the differentiators, especially for organizations with sensitive data or those that prefer to keep decisions within their own infrastructure.

Ruppersberg sees potential in Latin America, where many organizations need sophisticated AI but do not justify large centralized data centers. Brazil is on the startup’s radar, although there is still no established commercial presence in the country.

Applications and efficiency

Ammonix demonstrated four use cases: interpretation of electrocardiograms, health insurance billing, control of a power plant generating energy from waste, and computer vision. In the health insurance billing test, the company claims to have achieved performance comparable to an advanced language model using 4.7 times fewer tokens, a measure of processing. The results have not been verified by third parties.

The startup emphasizes that the approach is more effective in situations with well-defined options, such as diagnostic categories, industrial process states, or reimbursement decisions.

Regulatory context

With the entry into force of laws such as the European Union’s AI Act, the ability to trace decisions has become a business argument. Ammonix does not guarantee automatic compliance, but offers a framework that facilitates the analysis and auditing of its recommendations.

“For us, auditability is not an extra compliance feature—it’s part of the architecture,” says Ruppersberg.

The company also plans to integrate multiple agents working together, with control points for classifying and approving actions, aiming for greater security and governance.