An Agent Acts. A Chatbot Answers. Executives Need the Difference

An email reaches C.H. Robinson from a shipper with a load to move. It used to wait in a queue for a person to get to it, which could take up to four hours. Now an AI agent processes it, and by April 2025 more than 5,200 customers were getting their loads accepted in under 90 seconds. Another agent, launched that January, reads the emails carriers send offering spare trucks, pulls out the details and feeds them into the company’s real-time capacity centre. Across all its agents the company counted over 3 million shipping tasks. “That’s 3 million manual tasks our people didn’t have to do,” said Arun Rajan, its chief strategy and innovation officer (C.H. Robinson, April 2025). By October it described “a digital workforce of 30+ connected AI agents” (C.H. Robinson, October 2025).

Nobody asked those agents a question, though the releases don’t say how far they plan their own steps. An email arrived, and software changed a record in a live business system. That is the line I would use to sort everything sold to you as an agent: can the system change a record, send a message or commit money without a person pressing the button? Scripted automation passes it too, so a second test, who decides the steps, follows below. Sort every “agent” by that test, because the answer decides who in your organisation has to own it.

What is the difference between a chatbot, a copilot and an agent?

Ask a chatbot whether a shipment can make Friday and it tells you what it thinks. The answer is text, and a person reads it and decides what to do. A copilot drafts the reply to the shipper or fills in the form, then waits for someone to accept. An agent gets the goal, say to book this load by Friday at the agreed rate, and books it.

MIT Sloan notes that there isn’t a universally agreed definition of agentic AI. The one it quotes, from John Horton and his co-authors, describes “autonomous software systems that perceive, reason, and act in digital environments to achieve goals on behalf of human principals, with capabilities for tool use, economic transactions, and strategic interaction.” For an executive, the words that matter are act and on behalf of.

Gartner draws the practical line in its own advice. Organisations “can start by using AI agents when decisions are needed, automation for routine workflows and assistants for simple retrieval,” said Anushree Verma, a senior director analyst (Gartner, June 2025). Automation is on that list because it also acts without a person, on steps someone scripted in advance.

What makes a software system an agent?

An agent works in a loop, shown below. Anthropic, which builds these systems, puts the distinction in one line: workflows are “orchestrated through predefined code paths,” while agents “dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks” (Anthropic, December 2024).

The fourth step is the one an executive should watch. Plenty of systems do the first three and stop at a recommendation. That is a capable assistant, and often the right thing to buy. Once the system can take the fourth step without a person pressing the button, it is acting with your organisation’s authority, and you have bought an agent whether or not the contract uses the word.

In a meeting, ask what the system will be allowed to change on its own once it is live. Nothing means an assistant. A price, a booking, a customer record or a payment means it acts for you, as an agent or as scripted automation, and the next question is who owns what it does.

How do I tell a real agent from a relabelled chatbot?

Vendors have every reason to blur the line. Gartner calls it “agent washing”: “the rebranding of existing products, such as AI assistants, robotic process automation (RPA) and chatbots, without substantial agentic capabilities.” It estimates that only about 130 of the thousands of agentic AI vendors are real (Gartner, June 2025).

The opposite mistake costs as much. Verma also said that “many use cases positioned as agentic today don’t require agentic implementations.” Anthropic gives developers the same advice: find “the simplest solution possible,” which “might mean not building agentic systems at all” (Anthropic, December 2024). A fixed workflow that runs the same five steps every time is cheaper to run and easier to audit. Paying agent prices for that work, in my view, buys cost and risk and little else.

Three questions sort most pitches:

• What can it change without a person approving it? Ask for the list of systems it writes to, and the limits on each.
• Who decides the steps? If an engineer scripted every path in advance, you are looking at automation with a language model inside. That can be the right product, priced as automation.
• What happens when it is wrong? Ask how it stops, who is told, and how an action is reversed. A vendor that has run agents in production will have a specific answer.

Why should an executive care?

Because an agent’s mistakes happen in your systems. A wrong answer from a chatbot still has to pass through a person before anything changes. A wrong action from an agent has already happened by the time anyone looks: the load is booked, the email is sent, the refund is paid.

Gartner predicts over 40 percent of agentic AI projects will be cancelled by the end of 2027, “due to escalating costs, unclear business value or inadequate risk controls” (Gartner, June 2025). Two of those three reasons, as I read them, are decisions an executive makes before anything is built: which work deserves an agent, and what controls it runs under.

Every agent also carries authority that somebody delegated to it. How far that authority can run ahead of your controls and the people who answer for it is the subject of my argument on Premature Autonomy, which names the gap the Delegation Deficit. To score your own organisation, use the A7 readiness assessment.

Three more questions each have their own piece. How much freedom to give a particular agent is a question of autonomy levels; what changes when agents hand work to one another is the subject of the piece on multi-agent systems; and the piece on agent accountability deals with who answers when one gets it wrong.

Where should a first agent go?

Gartner advised in June 2025 that, at this early stage, agentic AI be pursued only where it delivers clear value or return on investment. I would add two conditions. Choose work where a wrong action is cheap and can be undone, and name the person who owns the outcome before the agent goes live. C.H. Robinson began with price quotes, orders and truck postings, which it describes as repetitive manual tasks, done in high volume (C.H. Robinson, April 2025). In my view that is the right shape for a first agent in any business: many small actions, each one easy for a person to check afterwards.

Sources and changes

• C.H. Robinson, "At C.H. Robinson, Artificial Intelligence Has Now Performed Over 3 Million Shipping Tasks", press release, 16 April 2025, and "C.H. Robinson unveils the Agentic Supply Chain, enabling companies in every industry to instantly deploy AI", press release, 20 October 2025.
• Beth Stackpole, "Agentic AI, explained", MIT Sloan Ideas Made to Matter, 18 February 2026.
• Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", press release, 25 June 2025 (read in an archived copy).
• Anthropic, "Building effective agents", 19 December 2024.

Changes on 6 October 2026. Cut from about 3,770 words to an Explainer and retitled. The C.H. Robinson figures now follow its two releases separately. Market forecasts, adoption surveys, case cards, a hypothetical startup and unsourced incidents are removed.