Agentic workflow

An agentic workflow is a process in which an AI system knows the goal but determines the route itself. It breaks the task into steps, calls tools, checks intermediate results, and reorders things when a step fails to deliver.

How is this different from automation?

Conventional automation is a fixed path. An invoice arrives, check field A, compare against table B, post to system C. Every branch is thought through in advance. Whatever was not thought through falls out and lands on a pile for humans.

An agentic workflow specifies only the goal and the permitted tools. If a value is missing, the system can go looking for it. If a document does not match the expected schema, it can try another route. That is what makes these workflows useful for processes where the inputs are messy, which is to say most real processes.

The price is non-determinism. Two identical inputs can take different routes. Measuring success is therefore part of the build, not a question for acceptance testing.

What is an agentic workflow made of?

Four components show up almost every time.

A goal stated as an outcome rather than a procedure. Tools with clear boundaries, often connected via MCP. A memory for the case in progress, managed inside the context window. And stopping conditions: a step limit, a cost ceiling, and the cases where it hands off.

That last point decides whether the thing can be operated at all. Without stopping conditions a workflow will loop when in doubt until the budget is gone, and still deliver nothing.

When is the effort justified?

When the process is frequent, touches several systems, and the inputs vary. Sorting incoming requests and enriching them with customer data, checking receipts against orders, reading tender documents for disqualifying criteria.

When the procedure is always the same, ordinary automation is cheaper, faster and easier to audit. An AI agent there is the more expensive tool with no added benefit.

The usual mistake is scope. Design a workflow for an entire department and you get a system whose success nobody can measure. A single process with a clear definition of success is the route into production, as described in Agentic software in the Mittelstand.

What has to be measured in production

Because the route is not fixed, "is it running" is not a sufficient question. Four numbers should be collected from day one, or a workflow cannot be steered.

Completion rate: the share of cases finished correctly with no human intervention. That is the actual benefit figure.

Steps per case. If it climbs over time, something has usually changed in the connected systems and the workflow is compensating with detours.

Cost per case. It varies by nature, and an average with no view of the outliers hides exactly the cases where something is going wrong.

And handover rate with reasons. A workflow that hands off frequently for the same reason is telling you what the next sensible extension is.

Those four are also the acceptance criteria for an AI pilot. Collect them only after rollout and you have no baseline against which an improvement could be demonstrated.