Before you begin
Start with one workflow, one source of information, and a named reviewer.
Use sanitized sample data while you are testing the structure and review loop.
Build the workflow
You do not need to rebuild your product to add AI.
Choose one task inside the existing experience
Extract fields from a document
Classify an item
Match a user with an option
Summarize a record
Recommend a next action
Draft a response for review
Detect an exception
Answer from approved knowledge
Define the feature as an API-like contract
Input
What the feature receives
Output
The exact fields it returns
Success
How you know the output is correct
Fallback
What happens when information is missing or confidence is low
Starter prompt
“Add an AI feature to [existing product] that accepts a defined input and returns [structured output]. Use [approved context]. Flag [edge cases]. Do not change the rest of the product.”
One bounded feature is easier to integrate, evaluate, price, and improve than a general-purpose assistant.
Common mistakes
Rebuilding the entire product to add one AI task.
Returning unstructured text when the product needs defined fields.
Failing to define the fallback for low-confidence output.
Frequently asked questions
Can I use this template for a different industry?
Yes. Replace the role, input, rule set, output, reviewer, and prohibited actions. Keep the first workflow bounded and inspectable.
When should I use production data?
After the structure works with safe examples and the required permissions, security, retention, and review controls are in place.