Rule-based, agentic or human: sorting every step of a workflow

Take an invoice exception. An invoice arrives, it is matched to a purchase order, the mismatch is investigated, someone decides on an adjustment, the adjustment is posted, the vendor gets a reply and, eventually, a payment is released. It looks like one workflow. It is really seven steps of three different kinds.

Rule-based steps

Reading and filing the invoice follows fixed rules. Matching it to an order and a receipt is rule-based too. These steps do not need an agent, and giving them one adds cost and a new way to be wrong. Conventional automation runs them.

Agentic steps

Investigating the mismatch is different. The answer sits across the order, the goods receipt and a thread of vendor emails, and what matters depends on the case. That is where an agent earns its place: gathering the evidence, comparing it and proposing what to do, inside limits we can write down. Drafting the vendor reply and posting an approved correction are agentic as well, because they follow the outcome and quote the evidence.

Human steps

Deciding the adjustment changes what the company owes, so the person accountable for it decides, with the agent's evidence in front of them. Releasing the payment stays with people whatever else changes.

Why the boundary matters

Once each step has a kind, the design follows. The rule-based steps get ordinary automation. The agentic steps get an agent with explicit permissions, a policy check before every action and a confidence threshold below which it asks a person. The human steps get the evidence they need, delivered where the decision is made.

Drawing that boundary well is most of the work. It is also what makes the system safe to extend later: new authority can be granted step by step, because every step already says who or what is allowed to run it.