The challenge of validating and governing AI agents for successful implementation in industry

Rosa Porta,

The Barcelona-based deep tech startup SOMIA Solutions has developed a platform to validate, monitor and govern AI agents, with applications in areas such as regulatory compliance and industrial production.

Artificial intelligence opens up new opportunities to transform industrial processes. However, there is a fundamental challenge between developing a solution that works in a demonstration and deploying it in a real-world environment: how can we ensure that its responses are reliable, that it operates correctly, and that errors can be detected and corrected?

This is precisely the challenge addressed by SOMIA Solutions, a Barcelona-based deep tech startup specialising in the validation and quality control of artificial intelligence agents.

During one of the sessions of the DONAIndústria programme, Tetiana Klymchuk, co-founder of SOMIA, shared with the 20 participating students not only her professional journey and the lessons that have shaped the company's evolution, but also the opportunities offered by the technology they are developing for industry.

 

From democratising AI to ensuring its reliability

SOMIA was founded in 2024 with an initial idea: to democratise access to artificial intelligence and make it accessible to companies without specialised teams. However, with the arrival of ChatGPT, access to generative AI expanded much faster than anticipated.

Faced with this new landscape, SOMIA identified a growing need: although companies could already develop AI agents, they often lacked the necessary mechanisms to demonstrate their reliability before deploying them in real-world environments, or to ensure that they continued to operate correctly over time.

Their response to this challenge was to develop their own platform to manage the entire lifecycle of an AI agent, providing a system for continuous monitoring and more reliable decision-making. To achieve this, the solution is structured around four stages:

  • Validate: define quality criteria and test agents using real-world scenarios before deployment, combining rules, AI evaluators and human review.

  • Monitor: track real-world executions and detect potential issues, with information on quality, cost and latency.

  • Improve: turn identified errors into new test cases and regression tests to assess subsequent versions.

  • Govern: maintain traceability of tests, reviews and approvals, with evidence to support and justify decisions.

 

Two use cases with direct applications in industry

SOMIA's technology can be applied to different industrial processes. Notable use cases include certification and regulatory compliance assistants, which facilitate the consultation of standards such as TISAX or ISO, and production data agents, which allow users to retrieve production information using natural language, without requiring specific technical knowledge.

However, one of the key lessons Tetiana Klymchuk shared during the session was that technology alone cannot solve everything. Developing useful and reliable AI solutions requires starting with real problems, having access to quality data and incorporating the knowledge of the people who understand an organisation's processes. In this respect, human oversight remains essential to validate results and ensure proper control of AI agents.

Ultimately, SOMIA's case highlights that the challenge of artificial intelligence in industry goes far beyond developing new technological tools: it is about integrating them reliably into business processes, with clear quality standards, human oversight and continuous improvement. A transformation in which technology is essential, but where people's knowledge and talent remain decisive.

Modify cookies