AI orchestration

When an AI workflow breaks in production, the cause is often one of these steps, for example a key that expired over the weekend.

How it works

The orchestrator, whether a workflow tool or your own code, runs the sequence. It fires on a trigger, calls each system in order, branches on the results, retries steps that failed, logs every run and pauses for approval before anything is sent outside the company. Credentials live here too.

There are two broad styles. In a workflow tool such as n8n, the steps are fixed and you can list them in advance, with one bounded AI step doing a defined job. In an agent setup, the model decides the next step itself. A practical test is to count the steps that make judgment calls against the steps that only move data. When judgment dominates, an AI agent may fit better than a long fixed workflow.

Example

An operations lead automates supplier invoices. When an invoice email arrives, an AI step reads the PDF and returns the supplier, amount and order number in a fixed format. The flow looks up the matching order in the accounting system. If the figures match, it creates a draft bill. If they don't, it sends the invoice to a person in Slack. Every run is logged with its outcome.

Common mistake

Turning on retries for every step. Retrying the lookup is harmless, but retrying "create bill" after a timeout can create the same bill twice. Mark which steps are safe to repeat before enabling retries, and give the others a duplicate check instead.