Begin with the work
An AI proposal should start with a specific task: preparing a draft account summary, classifying an internal request, or flagging missing information. “Add AI to sales” is too broad to define success, responsibility, or acceptable error.
Map inputs and decisions
Identify where the information comes from, who may access it, what output is required, and who uses the result. If definitions and ownership are unclear, automation can reproduce that confusion more quickly. Resolve the process questions first.
Design a contained pilot
Choose an approved dataset and a narrow use case. Define a baseline for time and quality, then decide how outputs will be checked. Include cases that are incomplete or ambiguous. A demonstration on ideal examples does not establish reliability in everyday work.
Keep a human decision point
Specify who reviews the output and what happens when it is wrong or uncertain. Do not allow an unreviewed draft to become a customer promise. Define the failure path and a way to return to the existing process if the pilot causes problems.
Expand only with evidence
Compare the pilot with the baseline, including review time, corrections, access controls, and user adoption. A workflow that produces drafts faster may still consume more total effort. Expand when the evidence supports the change, and keep the operating owner accountable afterward.