What is an AI workflow?
The AI Glossary
In our classes, we often pause to explain the ideas behind AI and the language used to describe it. These posts grew out of those conversations. They’re for anyone who would like to understand AI better, with no technical background assumed.
An AI workflow is a process that uses one or more AI capabilities alongside other software steps and human work. It might receive a document, extract specified information, check required fields and route an uncertain case to a person. The workflow is the complete sequence, not just the model call.
In one common engineering distinction, a workflow follows paths defined in advance, while an agent has more discretion to decide its next action. Usage varies, and real applications often combine the two. The important question is which decisions are fixed, which are model-selected and which belong to people.
Several patterns are useful. Chaining passes one step’s output to the next. Routing sends different cases to different processes. Parallel work handles independent pieces at the same time. A checking step can compare an output with defined requirements before the process continues. These are design options, not a ladder in which more steps always means a better system.
A workflow needs to preserve state: information about what has happened and what remains to be done. It should distinguish “request sent” from “request completed,” and a failed connection from a negative result. It also needs a response to exceptions, such as missing information or conflicting records.
Adding a second model as a checker may help, but it does not create independent truth. The checker can share the first model’s weaknesses. Some checks are better performed with ordinary software, such as confirming that a date is valid or a required field is present.
For a clinic, a referral workflow might identify the service requested, check for required documents, prepare a follow-up and record the responsible coordinator. The evaluation should include time spent correcting errors, duplicate contacts and cases that remain unresolved. A well-written message is not enough if nobody owns the next step. Start with the work people need to complete, then decide where AI adds useful interpretation or flexibility.
Sources: Anthropic — Building Effective Agents (2024); Anthropic — Demystifying Evals for AI Agents (2026)
To learn more, take a look at the AI in Health Care Certificate of Specialization from the Advanced Learning Academy at Harvard T.H. Chan School of Public Health. You can take an individual course or work toward the full certificate.
AI is a fast-moving field. This explanation reflects our class discussions and related material at the time of writing. Terminology, capabilities and tools may change.
For suggested additions or corrections, please email me at tpanch@hsph.harvard.edu.
Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy