The artificial agent does not replace the human but works with him

How does agentic AI differ from the Robotic process automation or intelligent automation that we knew until recently?

The main difference is that agentic AI integrates large language models with analytics, process automation and business rules. Traditional Rpa did not uselarge language models. Today, however, we put these components together - chatbots, automation, analytics and Llm - to build real intelligent agents capable of making complex decisions. It is a clear evolution.

Talking about multi-agent systems, are you moving in that direction as well?

Absolutely. With Sas Intelligent Decisioning, companies can build multi-agent agent flows. Imagine an evaluation process for a mortgage application: a 'router' agent identifies the type of activity required, and activates the most suitable agent. For example, for credit risk, for fraud detection or for suggesting the next action. Each agent has a specific role, and they can cooperate in a hierarchy.

What if an agent generates problematic or partial content? Do you have controls in place?

Certainly. In a multi-agent flow we can insert an agent dedicated to checking the bias in the texts generated by an Llm. If it detects an anomaly, it can activate a second agent for a double check or involve a human being. Escalation is an integral part of the system. It is one of the key features of our tools.

What models do you use to detect content bias?

We use a combination of Nlp models and business rules. The approach is holistic: we combine linguistic intelligence with the specific constraints of each sector or customer.

About governance: what kind of security do you integrate into your systems?

Security barriers may concern the masking or detection of sensitive information, compliance with company policies, model performance. These are all rules that the organisation can define, and which are applied within the decision-making flow.

But how do you identify sensitive data, technically? .

We use Ocr and classification and detection techniques. It is an integrated process in our analysis and governance tools.

You often associate agentic Ai with the presence of multiple agents. Is this correct?

In part. Agentic AI implies that a system has 'agency', i.e. the ability to act. A single agent can already be an agentic system. Multi-agent systems are an advanced, but not exclusive version of it.

And what about human intervention? Do multi-agent systems risk reducing human control too much?

This is a real risk. But in our experience, and based on what customers tell us, fully autonomous systems often fail in critical contexts. In regulated industries - such as finance, healthcare or insurance - you cannot afford margins for error without human control. That is why humans must remain in the decision-making cycle, at least at certain stages.

Does Sas intend to develop its own Llm?

No, we are not investing in creating our own Llm. We believe the value lies in the ability to integrate any model chosen by the customer - OpenAI, Anthropic, Mistral or others - into a secure, traceable and governed environment.

But in the end, will artificial intelligence replace human workers?

It is not so much a question of jobs, but of skills. Some roles will disappear, sure, but many others will emerge, as happened with the Internet in the 1990s. The difference is that this time we are not talking about 30 years, but 5 or 10. In the future, they will not ask you if you know how to use AI: they will take it for granted.