# Can you help us use AI in a real business workflow? Canonical URL: https://newphysician.org/can-you-do-this/design-an-ai-workflow/ Yes, by starting with a bounded job rather than adding a chat box to everything. AI can draft, classify, summarize, or assist a decision, but the workflow must define what evidence it receives, what it may do, and when a person must intervene. What a solution might involve Select representative tasks and difficult exceptions. Define acceptable outputs and forbidden actions. Build an evidence path, human review where needed, and a record of the model output and subsequent decision. Compare the new process with the existing one before expanding its authority. What needs care Fluent answers can hide missing information or unjustified certainty. Confidential inputs also require a deliberate provider and retention decision. A successful demonstration does not establish reliability on uncommon but consequential cases or authorize autonomous action. How we would judge whether it works The workflow handles agreed normal and difficult examples, identifies missing information, and routes uncertainty to a person. Staff can inspect supporting evidence and reverse an incorrect action. Performance and failure criteria are written down before launch. Likely engagement Business Engine: $5,000 to $15,000. This is a likely starting tier, not a quote. Scope, access, dependencies, and responsibility for operation need agreement first. Discuss Business EngineWhat to send first Bring one workflow, examples of good and bad outcomes, and the decisions that must stay with a human. Send me the messy version. Experience and scope Selected builds documents the currently authorized public proof. This page describes a proposed approach, not a claim of a completed client engagement in this exact workflow. Related problems Can you test whether an AI workflow misses important facts? Can we automate a workflow but keep a person in control? [Send me the problem](https://newphysician.org/send-me-the-problem/)