How to Choose Your First AI Project: A Practical Guide for Small Businesses

Start with the work, not the tool

Before comparing products, describe one recurring task in a single sentence: “Our team copies enquiry details from an inbox into a tracking sheet.” Record what starts the task, who does it, what information they need, and what a finished result looks like. Watch the work happen. The frustrating step is sometimes an unclear policy or a missing field, rather than a problem that needs AI.

Make a shortlist of three candidates. Good starting points can include drafting responses from approved FAQs, organizing incoming enquiries for staff review, or finding answers in a maintained internal guide. Keep decisions about hiring, lending, medical care, and other high-consequence outcomes outside a first general-purpose pilot.

Use a simple decision scorecard

Rate each candidate as low, medium, or high on five questions. These ratings are a discussion tool, not a mathematical promise of return:

• Frequency: does this happen often enough to learn from a small pilot?
• Clarity: can two team members agree on what a correct result looks like?
• Information: is there a reliable source the system is allowed to use?
• Reversibility: can a person spot and correct a mistake before it causes harm?
• Ownership: who will maintain the workflow after the demonstration?

Choose a frequent, well-defined, reversible task with usable information and a named owner. If none qualifies, improve the process or documentation first. Paying to automate confusion usually makes the confusion arrive faster.

Measure the existing process

Collect a small, representative sample of ordinary tasks, difficult cases, and failed handoffs. Record time spent, corrections, wait time, and whether the work was completed. Remove personal or confidential information from test examples unless you have established a suitable basis and approved environment for using it.

Compare the pilot against the same task, including the time people spend reviewing AI output. As an illustrative calculation, saving four minutes on 100 monthly tasks is 400 minutes, or about 6.7 hours before setup and maintenance. That is an estimate of time, not guaranteed cash savings. An employee may use the time for other work rather than reduce payroll cost.

Write the pilot agreement

A useful one-page brief states the audience, permitted task, approved information, prohibited actions, success measures, and stop conditions. Decide what the system does when information is missing. “Ask a person” is a product feature when a wrong answer would be expensive.

For example, an enquiry assistant could prepare a summary and suggest the next team to contact. It should not promise a price, availability, or appointment unless that information comes from an authorized current system. Test duplicate enquiries, contradictory instructions, and requests outside its remit.

NIST: AI Risk Management Framework provides a voluntary framework for managing AI risk across the life of a system. Use it as a reference for your review, not as a certification or substitute for sector-specific obligations.

Budget for operating the result

Ask for an itemized estimate covering discovery, data preparation, integration, hosting, model usage, monitoring, staff training, and maintenance. Clarify what happens when a supplier changes a model, an API fails, or your process changes. A low demonstration price says little about the effort needed to support a dependable workflow.

Before expanding, compare measured results with the baseline. Keep the pilot if quality and usefulness justify its cost; revise it if a specific failure can be fixed; stop if there is no meaningful improvement. Keep a record of the decision so a new tool does not restart the same experiment from scratch.

Bring a useful brief to a first conversation

• The repeated task and the person responsible for it
• A few anonymized examples, including one difficult case
• The tools and approved information involved
• The outcome you want to measure and a practical budget range

AI Empower’s AI consultancy can help assess the opportunity, map the workflow, and define a practical first version. Arrive with a business problem; the right answer may combine process changes, a conventional integration, and a small AI component.