AI Workflow — AI Glossary

Reference details

Related terms

• Agent Workflow
• AI Agent
• Prompt Chaining

How it works

1. The task is decomposed into steps with declared inputs and outputs, so each step is independently testable and the shape of the whole is reviewable before it runs.
2. Control flow is explicit — sequence, branch, loop, parallel fan-out — which is what distinguishes a workflow from an agent deciding its own next action.
3. Steps mix model calls, deterministic transformations, tool invocations and human checkpoints, and the deterministic steps are where most of the reliability lives.
4. Each step needs a defined failure behaviour: retry, fall back to a cheaper or simpler path, or halt and escalate. A workflow without these has one failure mode, which is stopping.
5. Reliability compounds multiplicatively across steps, so the end-to-end rate is the product and not the average — which is the arithmetic that governs how many steps a workflow can afford.
6. Observability is per step, because an end-to-end success rate tells you something failed and a per-step rate tells you what.

Example: The per-step reliability a multi-step workflow demands

Reliability multiplies. A 10-step workflow where each step succeeds 99% of the time completes 0.99^10 = 0.904 of the time, so roughly one run in ten fails despite every step being right 99 times in 100. Inverting the question is more useful for design: to reach 99% end-to-end across 10 steps, each step must succeed with probability 0.99^(1/10) = 0.9990 — three nines, not two. At 20 steps it is 0.99^(1/20) = 0.9995. That is the constraint behind every long workflow: the per-step bar rises as steps are added, and model-based steps do not reach three nines on open-ended work. The design responses follow from the arithmetic rather than from taste — reduce the step count, convert model steps to deterministic ones wherever the logic permits, and place human checkpoints where the compounding has already eaten the margin. A 10-step workflow with 7 deterministic steps at essentially 1.0 and 3 model steps at 0.97 completes 0.97^3 = 0.913 — marginally better than the all-99% version at 0.904, while exposing 3 steps to model error rather than 10.

Commonly confused with

A workflow and an agent differ in who decides the next step. A workflow's control flow is authored, so its paths are enumerable, its cost is predictable and it cannot solve a problem its author did not anticipate. An agent decides at run time, so it can handle unanticipated cases and its cost and path are not known in advance. Most production systems described as agents are workflows with a model call at one or two decision points, which is usually the right design — and calling it an agent obscures the fact that its reliability is the product of its steps rather than an emergent property.

When to use it

Reach for it when:

• Tasks with a known decomposition, where authoring the control flow buys predictability and testability.
• Where per-step observability matters, since end-to-end rates localise nothing.
• With deterministic steps substituted for model steps wherever the logic allows, which is the main lever on compounding.
• Where cost must be predictable, which authored control flow provides and run-time decisions do not.

Reach for something else when:

• Genuinely open-ended tasks whose decomposition is not known, where an authored flow forecloses the needed path.
• Long chains of model-based steps without checkpoints, where the multiplicative reliability makes end-to-end success unlikely.
• Without per-step failure behaviour, which leaves stopping as the only failure mode.
• Described as an agent when the control flow is authored, which hides the compounding that actually governs its reliability.

Referenced by

This term is used by the following ModelRefs references:

• tool chain

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to AI Workflow — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path