Agentic AI vs generative AI: what changes when you put one in production

Every explanation of agentic AI vs generative AI arrives at the same sentence. Generative models produce content; agentic systems take actions. That is true. It is also on every vendor's first slide, and it will not help you decide anything.

The difference that matters shows up somewhere less philosophical: the invoice and the incident channel.

A generative mistake is a draft, an agentic mistake is an event

A model that invents something produces a bad paragraph. Somebody reads it before anything moves, and the cost of the error is that person's attention.

An agent that invents something has sent the email, issued the refund, or written to the record. Somebody reads it afterwards, if at all. The error rate can be identical. The consequence is not, and everything below follows from that.

What changes once it is running

The first thing to change is monitoring. A generative endpoint is either responding or it is not, which is a question a health check can answer. An agent returning confident nonsense is responding, and every dashboard in the building will agree that it is fine. If your monitoring cannot tell answering from answering correctly, it is not monitoring the agent.

Then permissions, which stop being a setting and become part of the design. A generative model needs a prompt. An agent needs credentials, to the CRM and to whatever moves the money. Every tool you connect widens what it can get wrong at speed, and the access handed out during a pilot is usually still there a year later, because taking it back is nobody's job.

Cost behaves differently too. A generative call has a price you can multiply by volume. An agent that retries and calls three more tools to answer one question has a price you learn at the end of the month.

Retrieval stops being an implementation detail. When a generative model gives you a wrong answer you can argue with it — change the prompt and run it again. An agent working from a stale source keeps working from it, confidently, until somebody finds the source and fixes it, which assumes somebody on your side can open the source at all. That is a separate problem, and the one worth asking about first.

AI agents vs agentic AI: is that a different question?

Mostly a marketing one. "Agent" usually names a system with tools attached. "Agentic" usually names the property of pursuing a goal without being told each step. Anyone drawing a hard line between the two is selling a taxonomy.

The line worth drawing is the one above. Does a mistake produce a draft, or an event?

Agentic vs generative: which one do you actually need?

If a person reads every output before anything happens, you want the generative version, and you should push back on anyone selling you more. It costs less and it fails softly.

The agentic version earns its keep when that person is the bottleneck you are paying to remove. Removing them means building the monitoring and taking the ownership they were providing for free. That part is never in the pitch.

Most of the agent projects that quietly die were generative problems sold an agentic answer.

Before you commission one

We build these and then keep running them, which is a different relationship with the question than building one and handing it over. The products we run ourselves are where every paragraph above comes from. Nobody arrives at a list like this from a spec sheet.

The questions worth answering before you sign anything are in the readiness scorecard. Twelve of them, and no email.