Local AI Models Protect Sensitive Data

Artificial intelligence has opened a new period of experimentation across IN2’s engineering teams. “We decided to let the genie out of the bottle, and month after month we are watching what happens in different teams,” Ivan Gabrić, IN2's CEO, tells ICTbusiness Media – ICTbusiness.info. The approach reflects an effort to discover practical applications through development work rather than wait for the technology to stabilise.

Gabrić describes teams using new tools to revisit existing systems and develop upgrades that previously demanded substantially more time. The potential is especially interesting for organisations maintaining complex software built over many years. Yet the results are not uniform: different teams adopt the technology at different speeds and identify different opportunities. “Some teams have embraced it extremely quickly,” Gabrić says.

The rapid development of AI can also alter project assumptions in a matter of months. Gabrić refers to discussions involving the Ministry of Health as an example of how quickly the range of possible solutions can change. A capability that was difficult to envisage at one stage may become more realistic as the technology advances. “Things change from month to month,” Gabrić notes.

Speed, however, does not remove the need for regulatory discipline. IN2 works in healthcare and other sensitive environments where data protection is central to the design of a solution. Gabrić is particularly cautious about sending confidential information to external AI services and argues for the use of local models when the nature of the data demands it. “For sensitive data, we want the models to run locally,” he explains.

He is also unwilling to rely solely on a provider’s general assurances about where information is processed or stored. Claims about European hosting, in his view, need to be checked against the actual technical and contractual arrangements. The objective is not to reject new tools, but to understand their implications before deployment. “We have to be certain where the data goes,” Gabrić stresses.

Such checks can slow individual initiatives, but Gabrić considers them part of responsible engineering. Before adopting a new AI capability, teams need to define its purpose, test its performance, and evaluate its risks. He expects adoption to remain uneven across companies and individuals, even as the overall direction becomes increasingly clear.

For IN2, AI is therefore neither a passing experiment nor a reason to abandon established safeguards. The opportunity lies in modernising products and development practices while maintaining control over sensitive information. Gabrić’s central message is that openness to innovation and caution over data must advance together.