The rise of agentic AI: Trends, momentum and risks

From assistants to agents:
The evolution accelerates

The transition from AI assistants to true
AI agents is a watershed moment. Unlike their predecessors, which primarily responded to explicit human requests, today's agents can:

Take initiative based on observed patterns
and predetermined objectives

Coordinate across multiple systems to complete complex workflows

Learn from outcomes and refine their
decision-making processes

Adapt to changing conditions without
requiring reprogramming

This leap didn't happen overnight. It emerged from the convergence of several technological advances:

Advanced reasoning capabilities in large language models

Tool-using frameworks that allow AI to interact with software systems

Retrieval-augmented generation (RAG) providing access to specialized knowledge

Planning algorithms enabling multi-step reasoning and execution

Feedback mechanisms allowing systems to learn from successes and failures

The result is a new class of AI that doesn't just answer questions or generate content. It’s AI that gets things done.

1,063 COTS applications

The average enterprise uses 1,063 commercial off-the-shelf (COTS) applications. At least 50% of them already include AI, and that number is rising fast. Instead of governing 10–30 generative AI use cases, many organizations are facing 200+ embedded AI agents across the tech stack. Most of them operate without oversight.*

*Source: Salesforce