A 14-day AI pilot checklist for a small business

Before day one: make the experiment small

Pick one recurring workflow and one responsible owner. State the result you want to investigate without promising it will happen: for example, “Can we prepare an accurate weekly summary with less total handling effort?” Agree who will review outputs and how to return to the existing process.

Set the boundaries now: approved information, allowed tool, who can access it, outputs that need approval and cases outside scope. Set a small time and spending limit that your business can afford. The checklist is general operating guidance; it is not a substitute for security, legal or professional review where your task needs it.

Days 1–3: understand the current work

• Write the task’s trigger, required information, final output and usual exceptions.
• Define an acceptable result in observable terms, such as correct facts and no unsupported promises.
• Record at least five current-process examples, including all active checking and correction effort.
• Keep a note of case difficulty and operator experience so the trial comparison can use a similar mix.

Five examples are a starting point for this worksheet, not statistical proof. If the task only occurs once a week, two calendar weeks will not produce a useful sample. Extend the schedule or choose a more frequent task. Do not manufacture easy examples just to finish on time.

Days 4–7: prepare and test in draft

Prepare the smallest version of the workflow. Use invented or appropriately de-identified material while checking the mechanics. Keep outputs away from customers and live operations until the responsible reviewer is satisfied with the process and the data use has been approved.

• Record setup effort once: source cleanup, tool configuration, prompt preparation and team instruction.
• Test a normal case, a case with missing information and one that must be escalated.
• Write down what the human reviewer checks and what stops an output from being used.
• Keep a simple way to pause the trial and finish the work manually.

If the source information is unreliable or the reviewer cannot check the output, pause here. A pilot that identifies missing preparation has produced useful information. It does not need to become a live deployment to count as learning.

Days 8–12: collect comparable trial cases

For an approved, low-risk trial, work through at least five comparable cases with human review. Record the minutes spent doing the task, checking the output and making corrections. Include failed attempts and the manual completion they require. Keep setup separate so it is counted once.

Check quality on every case using the standard agreed before the trial. Note follow-up work that appears later. If you change the tool, prompt or process halfway through, mark the change and avoid pooling unlike versions as if they were one stable workflow.

Days 13–14: decide what happens next

Discuss the result with the person doing the work, not only the person sponsoring the idea. Check whether effort shifted onto somebody else. A faster draft that adds work to a manager’s day may have moved the bottleneck rather than removed it.

Document the examples, costs, limitations and decision. A small comparison can inform your next step; it cannot establish a reliable annual savings figure or prove AI caused the difference. If you continue, keep the reviewer, schedule another check and revisit the process when the tool or source information changes.