A useful AI workflow needs more than a capable model. It needs a clear agreement about what it may do, which evidence it must collect and when a person has to decide.
Begin with the action, not the agent
List the changes the system could make: drafting a record, sending a message, updating an account or releasing a payment. These actions carry different consequences. An approval rule should describe the action and its context, rather than relying on a vague instruction to be careful.
For an invoice workflow, extracting fields can be an automated preparation step. Matching a supplier and purchase order can be a deterministic check. Writing an approved record is a separate step with its own permission. Payment execution should not appear as an accidental consequence of approving a draft.
Make review a useful interface
A reviewer needs the proposed change, the source evidence and the reason it was suggested. Present missing fields and conflicting records directly. Offer an edit or rejection path, and make it clear whether the system is waiting or has already acted.
Define what happens when nobody responds, an integration times out or the underlying record changes. A retry must not create a second record. These are ordinary software design questions, even when a language model prepares the proposal.
Evaluate the whole workflow
Test incomplete documents, ambiguous instructions, duplicate submissions and permission failures alongside the happy path. Record whether the system asked for help at the right moment, not only whether its generated text sounded plausible.
Anthropic’s distinction between prescribed workflows and model-directed agents is a useful starting point. Our design recommendation is to introduce flexibility only where it helps a specific task, and retain explicit software boundaries around important actions.
Choose the smallest useful unit of autonomy. Give people enough evidence to make the decisions that remain theirs.
