Your first AI win should be boring.

Start with the work nobody puts in a pitch deck

The most useful place to begin with AI may be the weekly report someone builds by copying the same numbers into the same document. Or the first draft of a customer follow-up. Or the repetitive classification of incoming requests. These tasks lack glamour. That is part of their appeal: the inputs, outputs, and current frustrations are often easier to describe.

Before shopping for another subscription, keep a five-day friction log. Ask your team to note repeated tasks, how long they take, and where they have to stop to find information. Do not ask them to propose an AI solution yet. You want an honest picture of the work, including awkward exceptions.

Pick a task you can actually judge

Choose a candidate with a clear finish line and an owner who can recognise a good result. “Improve operations” is too broad. “Draft the weekly project update from approved notes, with every open issue and owner preserved” is testable. Avoid making a consequential customer decision your first experiment.

Use these four questions to narrow the list. A task does not need perfect answers, but uncertainty should make the pilot smaller.

• Frequency: does this work happen often enough to test repeatedly?
• Input quality: can the necessary information be supplied consistently?
• Reviewability: can a person check the result without redoing the entire task?
• Recovery: can a mistake be caught and corrected before it causes harm?

Run a small, honest comparison

Collect a baseline from real work. Then try the AI-assisted process on comparable examples. Track preparation, generation, review, and correction together. A fast first draft is not a win if it creates a longer editing queue. Record which cases fail, not just the pleasing examples.

For the pilot, keep a person in charge of approval. Agree on a stop rule in advance: pause if confidential information appears where it should not, if important facts are invented, or if the result repeatedly takes longer to check than to produce manually.

At the end of two weeks, choose one of three outcomes: continue, change the process, or stop. Document what you learned in a page the next person can use. The lasting advantage is a repeatable way to improve work—not an impressive demonstration that nobody trusts.