Agentic AI: Open-Source Ensures Data Sovereignty and Security

Murat Atıcı, the CEO of Bimser, says the next frontier is multi-agent collaboration—specialised agents working together like human teams

How do you define Agentic AI, and how does it fundamentally differ from traditional chatbots or rule-based automation?
At Bimser, we define Agentic AI as proactive systems capable of autonomously planning, deciding, and acting toward a final goal—going beyond reactive systems that wait for specific instructions. Its key difference from traditional systems lies in bridging the gap between “intention” and “action” independently. With our Agentic AI infrastructure, our Synergy product can autonomously carry out complex, multi-step tasks with a single objective—such as “analyse an email and its invoice attachment, extract the data, save it to the ERP system, and initiate an approval flow with the relevant department.” This capability goes far beyond simple rule engines or chatbots.

Beyond task automation, what complex, multi-step problems are agentic AIs uniquely positioned to solve that current AI models cannot?
Current AI models excel at singular tasks (e.g., text generation, translation). Agentic AI orchestrates these capabilities to solve complex business workflows end to end. We call this orchestration ability our “maestro” vision. For example, an agent can analyse a sensitive document locally (on-premise), mask personal data, and then send anonymised information to a cloud-based model for deeper analysis—something current siloed models can’t achieve.

Likewise, when an invoice arrives, an agent not only extracts structured data but also compares it against quality documents in our QDMS system, initiates corrective workflows if needed, and keeps a full audit trail. This requires the combination of retrieval, function calling, and process automation.

How much autonomy should Agentic AI have in decision-making, and where should humans remain in the loop?
We adhere to Responsible AI principles when implementing autonomy. The level of autonomy is determined by the task’s risk profile. Full autonomy can be granted for tasks like data extraction or translation. For identifying risky clauses in contracts, the agent analyses but seeks human approval before acting. In cases like procurement or financial commitments, the agent never makes autonomous decisions—it acts only as a co-pilot offering insights and analysis. We embed this risk-based decision logic into our AI infrastructure.

What infrastructure challenges (such as power, cooling, latency) arise when deploying Agentic AI at scale?
Agentic AI consumes significantly more infrastructure resources than a simple API call. We’ve proactively addressed this in our strategic planning. We provide customers with data sovereignty and model flexibility, enabling on-premise LLM installations. This allows them to plan power, cooling, and latency considerations internally.

We also support OpenAI, Gemini, and Claude—routing requests intelligently based on their cost and latency profiles. With our multi-tenant cloud architecture, we cryptographically isolate each client’s data and Retrieval-Augmented Generation (RAG) infrastructure while maintaining scalability.

What’s the next frontier for Agentic AI—will we see AI “agents” collaborating like human teams?
Absolutely. The next frontier is multi-agent collaboration—specialised agents working together like human teams. Our vision at Bimser is to enable communication between agents not only within our own product family but also across third-party B2B ecosystems.

For instance, a “QDMS Quality Agent” detecting non-compliance could trigger a corrective action process by notifying the “eBA Process Agent.” Concepts like the Remote Model Context Protocol (RMCP) Server will orchestrate such inter-agent collaboration. This architecture mirrors the division of labor and specialisation found in effective human teams.

Will open-source models keep pace with proprietary Agentic AI systems, or will there be a widening gap?
Our strategy is to combine the best of both worlds. Rather than a widening gap, we foresee increasing specialisation and collaboration. We aim to standardise on-premise deployments of open-source models like Llama 3 and Mistral—essential for clients who don’t want their sensitive data sent to the cloud. Meanwhile, proprietary models like OpenAI, Gemini, and Claude offer superior general reasoning.

Our hybrid approach anonymses data on-premise, then leverages proprietary cloud models for further processing. Open-source ensures data sovereignty and security, while proprietary systems deliver peak performance. The winning platforms will be those that integrate both flexibly—exactly what we’re building toward.