Back to Glossary
Workflow Orchestration
Synonyms: process orchestration, pipeline orchestration, workflow automation, durable workflows, agentic workflows
Do not index
Definition
Workflow orchestration is the layer that runs a multi-step process in the right order and decides what happens when a step fails. In AI products, the steps are often model calls and human approvals, with the path set in code ahead of time. In multi-agent coordination, a model picks the next step. For designers, it shapes what users see during long tasks, including the human-in-the-loop approval steps.
Use cases
Bad orchestration shows up as stuck tasks and users who can't tell whether to wait or retry.
- The double charge: A checkout flow times out on step 3. The user hits retry, the payment step runs again, and they're billed twice.
- Back to zero: An AI report builder fails on page 9 of 10. Nothing was saved, so it starts over and the user waits another 6 minutes.
- Approval in the void: An expense request waits on a manager's sign-off. The requester sees "Processing" for 4 days, with no hint that a person is the hold-up.
How it's used in practice
- Show the steps as a checklist: List what's done, running, and waiting.
- Match the indicator to the wait: NN/g recommends a spinner for 2 to 9 seconds and a percent-done bar for 10 seconds or more.
- Name the blocker: Write "Waiting for Mai to approve" and add a nudge button.
- Make retry safe: Ask engineering if a failed step can rerun without repeating side effects. If not, resume from the last saved step.
- Report partial success: "8 of 10 invoices sent. 2 failed: missing email." Users fix 2 and keep the other 8.
Design the paused state first. Long workflows spend a lot of time waiting on an API or a person. If the paused screen says what it's waiting for and when it checks again, users file fewer "is it broken?" tickets.
Challenges & limitations
- The UI can't see the state: Many engines track progress on the backend only, and the API often doesn't expose step-level status.
- Time estimates go wrong fast: Steps that call AI models or people can take seconds or days, so "about 2 minutes left" rarely holds.
- Every branch needs a screen: Retries, timeouts, rejections, cancellations. Each path is a state someone has to design.
Free resources
- AI agent orchestration patterns (Microsoft Learn) — sequential, concurrent, and handoff patterns, with when to use each.
- What is workflow orchestration? (IBM) — a plain explainer on how orchestration differs from automating single tasks
