Agentic AI in the enterprise: a working explainer

Four tests for the word

Marketing uses "agent" for almost anything that talks. Gartner calls the relabelling of assistants, chatbots and automation "agent washing", and estimated in 2025 that only about 130 of the many vendors selling agentic products offer the real thing.¹ There is no settled industry definition, so you will need a working one. Ours has four tests, and something counts as an agent only when it passes all of them.

1. It pursues a goal across several steps. One reply to one prompt is a chatbot.
2. It uses tools. It can query a system, write to one or call another service.
3. It chooses its own next step. A script that always follows the same path is automation.
4. It acts without a person approving every step. If someone clicks "approve" each time, that person is doing the agent’s job.

Chatbot

Answers in a single reply. No tools, no multi-step goal.

Assistant

Helps a person inside an application, for instance by drafting or summarising. The person drives each step.

Workflow automation

Follows a fixed path designed in advance. It may call a model, and the path never changes at run time.

Multi-agent system

Several agents with different roles passing work between them.

Plenty of useful products fail one of the tests, and they are still worth having. Fund and measure them as what they are.

How the loop works

An agent repeats the same cycle until the goal is met or it hits a limit. It takes a goal ("reconcile these two supplier lists"), plans the first step, calls a tool such as a query or an API, reads what came back and checks it against the goal. Then it decides whether to take another step, hand over to a person or stop.

Two pieces of infrastructure sit around that loop. The first is how the agent reaches tools. The Model Context Protocol has become the common standard, with support in ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot and Visual Studio Code. In December 2025 it moved to the Agentic AI Foundation, a directed fund under the Linux Foundation, reporting more than 97 million monthly SDK downloads and 10,000 active servers at the time.²

The second is how the agent learns what your data means. A query can return a table, and the agent still has to know which column is revenue and whether it includes tax. Gartner recommends a context layer for exactly this reason.³

Where adoption stands

Read two numbers together. McKinsey found in 2026 that 40% of respondents from large organisations are scaling AI agents, up from 27% a year earlier, while smaller organisations stayed at 22%.⁴ Gartner forecasts that more than 40% of agentic AI projects will be cancelled by the end of 2027, because of cost, unclear value or weak risk controls.¹

The sequence is the useful part. Single-task agents come first, agents that cooperate inside one product follow, and agents working across products come later. Your first project belongs in the first stage.

Where first projects go wrong

Choosing a first project

Look for a narrow job that people repeat several times a week, where a mistake is visible and cheap, and where someone can check the output in minutes. Prefer work that gathers and summarises over work that changes records. Make sure you can count something before and after, such as time saved, errors caught or cases closed, and that one business leader sponsors it while one engineer runs it.

Questions people ask