These two terms get used interchangeably in sales material, which is unhelpful, because the difference determines what you should build and what it will cost to run.
This article explains both in plain language, using one running example throughout, and ends with the decision most businesses should actually make.
The running example: handling an enquiry that arrives by phone at 7pm.
What is generative AI, in plain terms?
Generative AI produces new content in response to a request. You provide a prompt, it returns text, an image, code, audio or a summary.
The critical property is that it is reactive. Nothing happens until you ask, and nothing happens after it answers.
In our example: you paste a voicemail transcript into a chat window and ask for a summary and a suggested reply. It produces both, well and quickly.
Then it stops. It has not sent the reply, has not checked your calendar, and has no idea whether you did anything with it.
This is where most business use of AI sits today, and it is genuinely valuable. It removes the blank page and the first draft, which is often the slowest part of a task.
What is agentic AI, in plain terms?
Agentic AI is given an objective rather than a request, and works out the steps itself.
Three things make it different: it can use tools, it runs in a loop, and it decides when it is finished. The loop is the part that matters. It plans a step, takes it, looks at the result, and decides what to do next.
In our example: the objective is “handle out-of-hours enquiries”. A call arrives at 7pm. The system answers it, works out what the caller wants, checks the calendar, offers two appointment slots, books the chosen one, sends a confirmation text, writes the summary into the CRM, and flags the job for the morning.
Nobody specified that sequence for that call. The system chose it based on what the caller said. Had the caller been reporting a burst pipe, it would have taken a different path entirely.
Anthropic’s engineering write-up on building effective agents is the clearest technical description of this distinction, and it makes the same point we make below: most tasks do not need an agent.
The difference in one table
| Generative AI | Agentic AI | |
|---|---|---|
| What you give it | A request | A goal |
| What it returns | Content | A completed outcome, or an attempt at one |
| Can it use tools | No, by itself | Yes, that is the defining feature |
| How many steps | One | As many as it decides it needs |
| Does anything change in your systems | No | Yes, that is the point |
| Predictability | High for a given prompt | Lower, varies by run |
| Cost per task | Low and stable | Higher and variable |
| Review needed | On the output | On the actions, before they happen |
| Typical failure | A wrong answer you can see | A wrong action already taken |
The last row is the one to sit with. A bad draft costs you a minute. A bad action costs you an appointment, a customer or a refund.
How is agentic AI different from the automation you already have?
Most businesses already run automation: a form submission creates a CRM record, a booking triggers a confirmation email, a tag starts a sequence. Tools like these follow a flowchart you drew.
That automation is deterministic. Given the same input it does exactly the same thing, every time, forever. It is completely predictable and completely inflexible: anything you did not anticipate falls through.
An agent is goal-directed. You describe the outcome, and it works out the route. It handles the case you did not anticipate, and it might handle it in a way you would not have chosen.
| Deterministic workflow | Agentic system | |
|---|---|---|
| Design effort | High up front, you map every branch | Lower up front, higher in testing and guardrails |
| Handles the unexpected | No, it fails or does nothing | Usually, in some fashion |
| Same input, same output | Always | Not guaranteed |
| Debugging | Straightforward, follow the path | Harder, you are reviewing decisions |
| Running cost | Negligible per run | Meaningful per run, and variable |
| Right for | Repetitive, well-defined tasks | Tasks where the path genuinely varies |
This is the distinction almost nobody explains properly, and it is the one that should drive your decision. If you can draw the flowchart, you do not need an agent. Build the flowchart, because it will be cheaper, faster and more reliable.
Are chatbots AI agents?
Usually not, and the test is simple: does anything change in your systems when the conversation ends?
A chatbot that answers questions from a knowledge base is not an agent, however capable it is at answering. It produces text and stops.
A system that answers the question, books the appointment into your calendar, sends the confirmation and writes the record is acting. That is agentic behaviour, whatever it is marketed as.
Our own AI receptionist sits deliberately between the two. It converses generatively, and it takes a defined set of actions such as booking, texting and escalating, within limits we set. That middle ground is where most useful business systems actually live.
What is agentic AI useful for in a small business?
Four patterns are working in practice, and they share a feature: the path varies, and mistakes are recoverable.
Enquiry handling end to end. Answering, qualifying, booking and logging, across phone, form and message, with escalation rules for anything sensitive.
Research and monitoring. Watching a set of sources, gathering what changed, and producing a briefing. Errors here are visible and cheap.
Multi-step outreach. Finding prospects that match a profile, personalising the approach, following up, and stopping when someone replies.
Reconciliation and chasing. Finding what is missing, chasing the right person, and escalating what remains. Tedious, variable, and low risk when it goes slightly wrong.
Where they are not working: anything with money leaving the business without review, anything regulated, anything where a wrong action cannot be undone, and anything where the correct answer is genuinely contested.
Where does it go wrong?
Four failure modes, all well documented and all manageable if you plan for them.
Confident errors. The system states something incorrect with the same fluency as something correct. In a generative context you notice. In an agentic context it has already acted on it.
Runaway loops. An agent that cannot achieve its goal may keep trying, consuming time and money. Step limits and budgets are basic hygiene, not optimisation.
Cost variability. A task that costs pennies most days can cost pounds when it goes badly, because the number of steps is not fixed. Cap it.
Unreviewed actions. The genuine risk. If the system can send, book, refund or delete, define which of those need approval first.
The guardrails that address these are unglamorous and effective: a human approval step on anything irreversible, hard limits on steps and spend, a full log of every action taken, a defined stop condition, and a narrow set of tools rather than broad access. The NIST AI Risk Management Framework is a sound structure for thinking about this, and the UK government’s approach to AI regulation sets the direction on accountability.
How do you tell which one a vendor is selling you?
The labels on the website are not reliable, so ask about behaviour instead. Four questions settle it quickly.
“What does it change in my systems?” If the answer is nothing, it is generative. If it books, sends, updates or files something, it is acting, and you need to know exactly which of those it can do.
“What happens when it cannot complete the task?” A good answer describes a stop condition and an escalation. A vague answer means nobody has tested the unhappy path, which is where agents fail.
“How many steps can it take before it stops?” There should be a number. If there is not, the cost of a bad run is unbounded.
“Show me the log of what it did.” Every action an agent takes should be recorded and reviewable after the fact. A system that cannot show you its own decisions cannot be governed, and should not be given access to anything that matters.
None of these are technical questions, and you do not need to understand the underlying model to ask them. They are the same questions you would ask about a new member of staff with access to your systems.
Do you need agentic AI, or just a better workflow?
For most businesses reading this, the honest answer is a better workflow.
Ask three questions about the task in front of you.
- Can you draw it as a flowchart? If yes, build the flowchart. Deterministic automation is cheaper, faster, more reliable and easier to fix.
- Does the right path genuinely vary each time? If every case follows one of four routes, that is four workflows, not an agent.
- What happens if it does the wrong thing? If the answer is expensive or irreversible, you need approval steps, and once you have approval steps on everything you have rebuilt a workflow with extra cost.
Agents earn their place where the variation is real, the volume is high enough to matter, and the cost of an error is low or caught by review. That is a narrower set of problems than the market suggests, and it is growing.
The businesses getting value from this technology today are almost all doing something simpler than they were sold. They automated a repetitive process properly, added generative AI to remove the drafting, and kept a person on the decisions. That is not a fashionable answer, and it is the one that works.