What agentic AI means for someone who runs a function

Agentic AI arrives in your inbox as one word covering three different things. Chat, where you hold every step. Workflow, where you fixed the steps in advance. Agent, where the order of steps is produced while the work runs. Place any pitch into one of those three layers and most of the noise drops away.

What are the three layers of agentic AI?

I sort them into three layers, separated by who holds the judgment at each step.

Layer one is chat. You paste last month's trade-spend lines into ChatGPT, ask for the five largest deviations against plan, read the list, and decide what to do about the South zone. The software produces the list. You hold every step of judgment in the work.

Layer two is workflow. You write the steps once, in advance. Clean the file, compute deviation against plan, draft the note in the format your zonal heads already read. The order is fixed by you before anything runs. The software fills the slots, the same way every month.

Layer three is agent. You hand over a goal and some tools, and the order of steps is assembled at run time rather than written down by you beforehand. "Work out why the South zone missed plan and draft the corrective note" is a goal. It is not a sequence.

That third layer is where the word agentic comes from, and it is the layer engineering teams define most narrowly. Their distinction is plain enough to carry into a board room. A workflow runs on paths set down in advance. An agent's path is assembled at run time.

How do I start with AI?

Almost every function runs at layer one today, whether or not anyone calls it that. Start there, on one document you already own. Not a pilot. Not a tool evaluation. One file that crosses your desk every month and that you could check line by line if somebody made you.

I know the seat this advice lands in. Your week is launches, targets, teams and the numbers, with nothing left over. Layer one is the only layer you can run inside a week like that, because it needs no build, no vendor and no IT ticket.

There is a second reason to start there. Layer one is where you find out what the output gets wrong on your own category. Take eight competitor price points in a market you know cold and ask for a summary. Some of the eight will be stale. Spotting which ones is what you are actually learning, and it carries into layer two and layer three unchanged. Learning AI for your function is mostly that: building the habit of checking, on material where you are the expert in the room.

Which layer is the pitch on my desk selling?

A deck that puts the word agentic on slide one is often selling layer two in layer three language. Three questions sort them, and none of the three needs a technical answer.

Who fixed the order of steps, you or the vendor, and can somebody show you that order on a page? If nobody can show it to you, what you are buying behaves like layer three whatever the deck says.

What happens at the step that fails? In a pharma or FMCG function the failing step is usually a stale item master, a renamed column, a distributor claim that got settled offline. A layer two build stops there and says so. A layer three build carries on, and the error reaches you later wearing a finished format.

Whose name is on the output? Somebody signs the price revision that goes to the field. That signature does not move because the draft arrived faster. When IT owns the AI policy and nobody owns daily use, this is the question that gets skipped in the pilot, and it is the one the board asks you first.

What else do people running a function ask about agentic AI?

Do I need to understand how agents work under the hood?

No. You need to know which layer a thing sits in and where the judgment sits inside it. That is readable from the outside. Ask who fixed the order of steps and what happens when one of them fails, and the architecture stops mattering to your decision.

Can I put my own data into a layer one chat?

Check your own policy first, because that line is usually undrawn rather than drawn against you. Until somebody draws it, work on material you would be comfortable emailing outside: published competitor data, your own drafts, numbers already released. The habit of checking builds fine on those.

My juniors already use this more than I do. Am I behind?

No. Speed at layer three without a check at layer one is the risk you are being asked to manage. Ask one of them to show you where an output was wrong last month. If nobody can point at a wrong output, nobody has been checking.