Agentic AI: when the model stops answering and starts working

There is a fundamental difference between a model that answers questions and a model that executes tasks. That difference has a name: agentic AI. And in 2026, it moved off paper.

What changed

During the first years of LLMs, the pattern was simple: you ask, the model answers. It could be a text, some code, an analysis — but always an answer. You still did the work.

Agentic AI inverts that logic. The model receives a goal, not a question. It plans the necessary steps, executes each of them using external tools — APIs, databases, graphical interfaces — evaluates the intermediate results, and delivers the final product. You define the destination. The model charts and travels the path.

Numbers that confirm the shift

The market for autonomous AI agents was valued at US$ 8.5 billion in early 2026 and is expected to reach US$ 35 billion by 2030. More revealing than the size is the speed: by the end of 2026, around 40% of enterprise applications are expected to incorporate some kind of specialized agent, compared with just 5% in 2025.

This is not a roadmap. It is adoption in progress.

Where it is happening in Brazil

Agibank began using autonomous agents in its WhatsApp service to answer queries about credit and financial services — with no human intervention across much of the flow. Banco do Brasil is moving in the same direction, focused on modernizing customer service and operational gains. Globo is among the companies advancing in implementation.

The average ROI reported by companies that have already deployed agents ranges between 200% and 400% in the first year. The areas with the highest returns are customer service, data analysis, compliance, and onboarding.

What it takes to work

Autonomous agents work well when three conditions are present: enough long context to maintain the state of a complex task, the ability to use external tools via APIs, and self-verification logic that detects errors before moving forward.

Not by chance, the most recent models — GPT-5.4, Claude Opus 4.7, Gemini 2.5 Pro — were all updated with specific improvements in those three areas. The model market is being shaped, in large part, by what agents need.

The risk few people discuss

Delegating autonomous execution to an AI system is not trivial. An agent that misinterprets a goal can carry out dozens of wrong actions before any human notices. Governance, audit logs, and well-defined scope limits are not optional — they are part of the project.

Agentic AI is a real paradigm shift. But the cost of a poorly planned implementation is proportional to the autonomy you grant.

What is coming

Conservative estimates suggest that by 2028, most large companies will have at least one process entirely managed by AI agents. The most common entry point today is service and triage. The natural next step is analysis and decision.

Whoever is learning to work with agents now will be ahead when the next step arrives.