What an agentic workflow is, without the hype

An agentic workflow is a piece of software that carries a business process from start to finish. It takes in work, makes decisions about it, acts on those decisions through tools, and hands the result on. The difference between it and a chatbot is that it does the work rather than talking about the work. The difference between it and a script is that it can cope with variation.

A script follows a fixed path. If the invoice arrives as a tidy PDF with the fields where they should be, the script wins every time, and it wins cheaply. The moment the invoice arrives as a photograph of a crumpled page, or the supplier changes their layout, or the amount does not match the purchase order, the script stops and a person gets involved. An agentic workflow reads the messy version, works out what it is looking at, checks it against the records it can reach, and either finishes the job or flags it with a note explaining what went wrong.

The parts that make one up

Strip the marketing away and a serious agentic workflow has four components. First, workers: individual AI agents, each with a written role, a set of tools it is allowed to use, and limits on what it may do without sign-off. Second, shared state: a place where work in progress lives, so one worker can pick up where another left off without the two of them chatting. Third, review: gates where output is checked before it takes effect, by a person or by another agent with a narrow brief. Fourth, logs: a record of every action and every decision, so the answer to "why did it do that" is never a shrug.

If a vendor cannot show you all four, you are being sold a demo, not a workflow.

What it is not

It is not artificial general intelligence, and it does not need to be. The processes worth automating are specific: read this inbox, reconcile these payments, draft this report, chase this document. Narrow and reliable beats broad and impressive.

It is not an overnight replacement for a team. Sensible installations start with one process, run alongside the people who currently do it, and earn their way into production with measured results.

It is also not a chatbot wearing a lanyard. If the system's only output is text for a human to act on, the human is still the workflow. That can be useful, but it is a different, smaller thing.

Where it fits, and where it does not

The good candidates share a shape: the work repeats, the volume is real, the inputs vary enough to defeat a script, and success is checkable. Invoice processing, client onboarding, support triage, reporting, research and content production all qualify in most firms.

The poor candidates are one-off projects, decisions that hinge on taste or politics, and anything where being wrong once is catastrophic and there is no way to check the output before it lands. Those stay with people, and the honest vendors will tell you so on the first call.

That is the whole definition. Software that does defined work, copes with variation, and can prove what it did. Everything else is packaging.