AI workflow

What it means

An AI workflow is a set of steps that a person lays out ahead of time. One or more of those steps hands a job to an AI model, a program that reads and writes text. The person who built the workflow decides what happens and in what order. The AI does its assigned step and passes the result along.

Here is an everyday example. A receipt arrives by email. The AI reads it and pulls out the amount. A fixed rule checks whether the amount is under the spending limit. If it is, the expense is filed. If it is not, it goes to a manager. The AI never decides to skip the manager, because nobody wrote that step for it.

Anthropic, the company behind the Claude models, describes workflows as "systems where LLMs and tools are orchestrated through predefined code paths."¹ An LLM, or large language model, is the kind of AI behind chatbots. "Predefined code paths" means the route was set before the work began.

Where the word comes from

Businesses talked about workflows long before today's AI. In 1996 the Workflow Management Coalition, an industry standards group, defined a workflow as "The automation of a business process, in whole or part, during which documents, information or tasks are passed from one…" participant to the next.² In that older sense, a workflow was a business process run by software.

The meaning shifted once AI models could do some of the steps. In March 2024 the AI researcher Andrew Ng wrote that "With an agent workflow, however, we can ask the LLM to iterate over a document many times," and contrasted this with having a model write "without revising its work."³ For Ng, a workflow was a way to make the AI work in several passes instead of one.

In December 2024 Anthropic drew a sharper line. It called the difference between workflows and agents an "architectural distinction." Workflows follow a fixed path, while agents are "systems where LLMs dynamically direct their own processes and tool usage."¹ Most of the industry now uses the word this way. We could not find a source showing who first put "AI" and "workflow" together, so we leave that out.

How others define it

• Anthropic: "systems where LLMs and tools are orchestrated through predefined code paths."¹
• OpenAI: "A workflow is a sequence of steps that must be executed to meet the user's goal."⁴
• Workflow Management Coalition (1996, before modern AI): "The automation of a business process, in whole or part, during which documents, information or tasks are passed from one…"²

Where they agree and differ

All three describe a series of steps that leads to a goal. Anthropic is the most explicit about the key point: in a workflow, the person who built it decides the steps. In an agent, the model decides.¹ The 1996 definition has no AI in it at all, which shows that the AI is just one more worker inside a process that is much older.

The strongest objection is that this line is fading. In practice, companies put workflows inside agents and agents inside workflows. Even Anthropic doesn't treat one as better than the other. It says "workflows offer predictability and consistency for well-defined tasks," while "agents are the better option when flexibility and model-driven decision-making are needed at scale."¹ We are fairly confident the distinction still helps, because it tells a leader who chose each step: a person or the model.

In the news

In October 2026 Microsoft's Copilot Studio team described "hooks," a preview feature that runs a workflow at set moments while an agent is working, such as just before the agent uses a tool. The team wrote, "Hooks provide predictable checkpoints where the runtime invokes your logic."⁵ The trade publication Cloud Wars summed up the difference: "In the case of a tool, the agent chooses to start it, whereas with a hook the event automatically fires."⁶

The story matters because it shows companies using workflows to put fixed rules around agents. Microsoft's team also gave a caution: "Because Copilot Studio hooks currently fail open, enforce essential rules independently in the tool or underlying service."⁵ "Fail open" means that if the check breaks, the agent keeps going anyway. We are confident in these details because they come from Microsoft's own team blog. The feature is a preview and may change.

What it means for your organization

A workflow is the most framed form of AI at work, with every step decided by a person. The Knowledge Core argues that leaders frame, then free.⁷ Workflows are where framing is easiest to see and check. They suit work that should run the same way every time. Agents suit work where the steps can't be known in advance. Many organizations will need both, and they should know which one is running each job.

One question to ask your team: For each AI system we run, did a person fix the steps or does the model choose them, and who is accountable when a step fails?