A small business rarely needs someone to install AI everywhere. It may need help turning approved service information into draft replies, organizing routine inquiries, or producing a first version of an internal instruction sheet. A useful paid service starts with one such task and a way to check whether it improved.
Your experience in a particular trade or office can help you recognize the awkward parts of a workflow. Technical confidence alone is not enough: you need to understand what an incorrect answer would cost the business and what staff can realistically maintain.
Watch the existing process before proposing a tool
Ask the owner and the person doing the work to walk through a recent example. Identify the input, the decisions, the output, and the person responsible. Record how often the task occurs and where delays actually happen.
A shop receiving repeated questions about opening hours might need a corrected website rather than a chatbot. A spreadsheet with consistent categories may solve a reporting problem without generative AI. Include these simpler options in your recommendation.
Write the proposed pilot in one sentence: “Use the business’s approved service sheet to prepare draft answers for staff review.” Keep bookings, refunds, payments, and promises to customers outside the initial scope unless there is a separately designed and authorized process for them.
Agree on a small, paid deliverable
A first pilot could include a workflow review, a configured draft-only process, a small test set, staff practice, and written handover notes. Set a time boundary and a fixed list of deliverables. Explain any ongoing software charges and what support is included after delivery.
Spell out the owner’s responsibilities too: supplying accurate reference material, approving the test data, choosing a reviewer, and deciding whether to continue. Do not promise a particular sales increase or staff reduction. The pilot is an experiment with a decision at the end.
If the main problem is choosing between vendors, our article on comparing AI tools for a client covers that earlier selection stage.
Prepare a reliable reference sheet
For a fictional bicycle repair shop, the approved sheet might contain opening hours, services offered, how to request an estimate, and which repairs require inspection. It should not contain invented turnaround times or a promise that every repair has a fixed price.
Ask the owner to verify the sheet and assign a review date. Show staff how to update it. If the assistant has no reliable source for an answer, the expected response should be to ask a staff member rather than to improvise.
A prompt might say: “Draft an answer using only the attached approved service sheet. Mark any unanswered question for staff review. Do not confirm appointments, quote a repair price, or promise completion dates.” Test whether the tool actually follows those limits.
Test the mistakes that would matter
Build a small set of fictional inquiries covering routine questions, unclear requests, missing information, and requests outside the service. Add at least one question that would tempt the assistant to make a promise it cannot support.
For each result, record whether the answer is accurate, whether a staff member must rewrite it, and how long review takes. Compare this with the existing approach using similar tasks. A tiny test can reveal problems; it cannot establish that every future answer will be correct.
NIST’s AI Risk Management Framework emphasizes defined responsibilities, evaluation, and continued monitoring. For a small pilot, put those ideas into plain language: who checks the answer, what counts as failure, and who can stop the process?
Keep access and customer information limited
Use fictional or suitably minimized examples first. Before handling live customer information, have the business approve the service and data use, review the provider’s current terms and settings, and resolve any applicable privacy obligations. An owner giving you a login is not permission to copy everything in the account.
Use individual access with only the permissions the task needs. Let the owner control subscriptions and billing. Avoid handling payment details, passwords, health records, or employee evaluations as convenient demonstration material.
Keep a manual fallback available. If the reference sheet is outdated, the tool stops working, or drafts become unreliable, staff should know how to return to the previous process without waiting for you.
Hand over something people can use
Have the actual staff member perform a few examples while you observe. Provide short instructions covering permitted inputs, checking the answer, correcting the reference sheet, reporting problems, and pausing the tool. A recording can help if everyone involved agrees, but clear written notes are still useful.
Finish with a decision record: continue, revise, or stop. Include the observed time and correction burden, remaining limitations, recurring costs, and a named owner. Offer further support only for a defined need. A pilot that shows AI is unnecessary can still save the business a costly commitment.
AI disclosure: This article was written and revised with AI assistance. Sources checked September 27, 2026. See our disclaimer.
Illustrative photograph by Pete Walls on Unsplash.
