Most AI conversations in small businesses start in the same place: a demo. Someone has seen a tool that writes, summarises, classifies, or answers questions, and the question becomes “where could we use this?” It is an understandable way to start, and usually the wrong one.
The tool is rarely the scarce resource. The scarce resource is a clear view of the work: where time is lost, where quality is inconsistent, where a customer waits, and where a person is doing something repetitive enough to justify a new system.
Start with a workflow, not a tool
Choose a piece of work that happens often, has a recognisable input and output, and can be measured. Examples in a small professional firm might include drafting a first-pass proposal, triaging inbound enquiries, turning a call transcript into a structured summary, checking documents against a checklist, or preparing a weekly client update.
These are not glamorous. That is a feature. The first useful AI project is usually the one nobody would put on a conference stage, because it removes friction from work that already has a clear owner and a clear standard.
The best first AI project is usually the one nobody would put on a conference stage.
Three filters
Once you have a workflow in mind, test it against three filters. They are deliberately simple, and they usually eliminate most candidates.
1. Frequency and volume
A task that happens twice a month may be annoying, but it is rarely worth a new system. A task that happens twenty times a day has a compounding cost. Estimate the hours, the wait time, and the rework. If the numbers are small, the project is probably a process fix, not a technology project.
2. Judgment and consequence
Some work benefits from a first draft; some work requires a decision that should not be delegated. Sort tasks by the cost of being wrong. A low-stakes internal summary is a good early candidate. A client-facing recommendation, a legal conclusion, or a price quote may need a human decision at the centre—even if AI prepares the material. The goal is to redesign the workflow around human review, not to pretend the review is unnecessary.
3. Data readiness
Does the work have accessible inputs? Are they clean enough to use? Where do they live, who owns them, and what is sensitive? A useful pilot often needs less data than expected, but it needs the right data. If the inputs are scattered across inboxes and personal drives, fix that before buying a tool.
The first pilot should be boring
Pick one workflow, one team, and one measure. Set a baseline before you start: cycle time, error rate, rework, response time, or the number of handoffs. Run the pilot for four to six weeks—long enough to see a pattern, short enough to stop if the evidence is poor.
Keep the scope narrow. A pilot that tries to change three teams, four tools, and a client-facing process at once will produce a story but not a learning. The point is to find out whether the new workflow is better, for whom, and under what conditions.
Measure what changed, not what was deployed
“We launched an AI tool” is not an outcome. The useful questions are: Did the work get faster? Did quality improve or decline? Did the team spend less time on low-value tasks? Did clients notice? Did the cost per unit change? Did the exception rate rise in a way that shifted work downstream?
Include the cost of review, integration, training, and maintenance. A tool that saves an hour of drafting but creates forty minutes of checking may still be worthwhile—but only if the quality is better or the person’s time is freed for something more valuable.
Keep the human in the loop where it matters
AI is good at producing plausible material. That is exactly why review matters. Build the workflow so a named person is accountable for the output, with a clear checklist and an easy way to correct the system. Log the corrections; they are the best evidence of where the tool is weak and where the process needs to change.
The aim is not to remove people from the work. It is to move their attention to the parts that need judgment, relationship, and accountability.
Then, and only then, talk about scale
Scaling means more than turning on more licences. It means the workflow is documented, the data access is managed, the review is consistent, and the economics still work at higher volume. If the pilot succeeded because one enthusiastic person held it together, you do not yet have something that scales.
The most valuable output of a first AI project is not the tool. It is a clearer understanding of how your business works, where the bottlenecks really are, and what you would need to change to remove them. That understanding is useful whether or not the tool survives.