The word changed some time in the last year. Everything is an agent now. The same software that was a chatbot in the spring is an agentic platform by autumn, at a higher price, with the same underlying capability.
Underneath the marketing there is a real distinction, and it is worth understanding, because it is the difference between something that helps a person work and something that does work without one.
The actual difference
A chatbot answers. You ask, it responds, the exchange ends, and nothing in the world has changed.
An agent acts. It is given an objective rather than a question, and it can take steps toward that objective: look things up, call other systems, make a decision about what to do next, and repeat until it decides it is finished.
The distinction is not intelligence. It is permission and consequence. A chatbot produces text a person then chooses to act on. An agent changes something, and by the time you read the summary, it has already happened.
Where it genuinely earns its place
Agents suit work with a specific shape: many small decisions, each one easy, none of them worth a person’s attention individually, but collectively eating real hours.
Sorting inbound enquiries by what they are actually about and routing them to the right person. Reading a supplier document and pulling the same six fields out of it every time. Checking a submission against a checklist and flagging only the ones that fail.
What those share is that the decision is genuinely a judgement, so a fixed rule handles it badly, but the consequence of getting one wrong is small and recoverable.
When you do not need one
This is the more useful half of the answer, and it applies to most businesses asking the question.
- The rule is actually a rule. If the logic is “over five hundred euro goes to the finance manager”, that is a condition, not a judgement. Ordinary workflow automation does it more cheaply, more reliably, and it will produce the same answer every single time, which an agent will not.
- It happens twice a week. Automation has a build cost and a maintenance cost. Low-frequency tasks rarely repay either, however annoying they are.
- Getting it wrong is expensive. Anything touching money leaving the business, legal commitments, or a customer relationship should have a person in the loop.
- The process is not written down. If three people do it three different ways and each believes theirs is correct, automating it just makes the disagreement faster.
That last one stops more projects than any technical limitation, and it is always the cheapest thing to discover early.
The part that gets skipped
An agent that can act can act wrongly, at machine speed, without anyone watching.
A person doing the same task badly makes one mistake and usually notices. An agent makes the same mistake four hundred times before anyone opens the report. This is not an argument against agents. It is an argument for the three things that rarely appear in the sales conversation.
- Boundaries. What it may do, what it may never do, and a hard limit on scale. Not “handle refunds” but “handle refunds under this amount, and escalate everything else”.
- A record. What it did and why, readable by a person who was not there. If you cannot reconstruct a decision afterwards, you cannot defend it to a customer.
- A stop. One control that halts it, that a non-technical person can reach without ringing anybody.
A proposal that does not cover all three is describing a demonstration, not a system. This is closely related to why quiet failure is the expensive kind.
How to start, if you are going to
Pick the least important thing that qualifies. Not the biggest opportunity, the least consequential one.
Run it alongside the existing process for a few weeks, with the agent suggesting and a person deciding. You are not testing whether the technology works, you are finding out how often it is wrong and in which direction, which is the only thing that tells you where the boundaries belong.
Then let it act on the narrow slice where it has proven reliable, and leave the rest with people. That sounds unambitious. It is also the version that is still running in a year.
The honest summary
An agent is software given an objective and permission to act on it. Useful where there are many small judgements, low stakes on each, and a written-down process underneath.
Most businesses asking about agents actually have a rules problem, and rules are cheaper, faster and more predictable. Being told that is worth more than being sold an agent.
If there is a task eating hours and you cannot tell whether it is a judgement or a rule, describe what happens step by step. That distinction is usually clear within a paragraph, and it decides everything that follows.
Common questions
What is the difference between a chatbot and an agent?
Permission and consequence, not intelligence. A chatbot answers and the exchange ends. An agent is given an objective and can act on it, so by the time you read the summary something has already happened.
Are agents safe to let loose on customer work?
Only with boundaries, a readable record of what it did and why, and a stop control a non-technical person can reach. A proposal without all three is describing a demonstration, not a system.
How do we know whether we need one?
Ask whether the decision is a judgement or a rule. If the logic is “over five hundred euro goes to finance”, that is a rule, and ordinary automation does it more cheaply and more predictably.