How to Choose Your First AI Project (A 5-Question Test)

I have watched more first AI projects die from a bad choice of problem than from bad technology. The pattern is consistent: someone picks the task that annoys them most, rather than the task that has the most repetition and the clearest right answer, and three weeks later they are correcting the output more often than they were doing the job themselves.

Choosing a first AI project for business owners with no technical staff is a scoring exercise rather than a gut call, and the gut is usually wrong. Annoyance and automatability are unrelated. The invoice chase that makes you grind your teeth might happen six times a month and require a judgement call each time. The enquiry you answer twelve times a week in nearly identical words does not annoy you at all, and it is the one worth automating.

Here are five questions to score a candidate task with. Do it on paper for two or three tasks before you spend anything, and use your real numbers rather than an impression.

Question one: how often does this actually happen?#

Count it, do not estimate it. Owners routinely double the frequency of tasks they dislike and halve the frequency of tasks they have stopped noticing.

Volume matters because every AI project has a fixed cost in setup and correction time that has to be earned back per repetition. Under about ten occurrences a week, that cost rarely comes back. Between ten and fifty, the case is usually strong. Above fifty, the task is probably already costing you a person's time and the arithmetic makes itself.

Do the counting on a fortnight of real data. Your phone log, your inbox, your booking system. If you cannot find the number anywhere, that is worth knowing too, because a task you cannot count is a task you will not be able to prove improved.

Question two: would two of your staff answer it the same way?#

This is the repeatability test, and it is the one that quietly kills projects.

If you asked two experienced people in your business to handle the same case, would they produce roughly the same result? For "what does a standard callout cost", yes. For "should we take this job", almost certainly not, because that involves reading the customer, the calendar and your appetite for the work.

AI learns from consistency. Where your business has a settled answer, an agent will reproduce it faithfully. Where your business genuinely improvises, the agent will improvise too, and you will not like the results. Feed it the settled half and keep the improvised half.

A useful side effect of asking this question is discovering how much of your business has no settled answer at all. That is worth fixing regardless of whether you automate anything.

Question three: what does one wrong answer cost?#

Score this in dollars and in trust, and be pessimistic.

A wrongly categorised email costs almost nothing. A wrong price quoted to a customer costs either the margin or the awkward retraction. A wrong instruction about a medication or an electrical hazard is not in the same category at all and should not be automated in any form.

The rule I use is straightforward: if a single bad output could cost more than a month of the system's fees, the task needs a human check before anything reaches the customer, and that check has to be somebody's actual job. Tasks where mistakes are cheap and visible are the best first projects because you learn quickly and pay little for the learning. The recurring ways this goes wrong are collected in seven AI mistakes that cost small businesses money.

Question four: does the information already exist in writing?#

An AI system can only answer from what it has been given. If the knowledge lives entirely in your head, or in the head of the person who has been there eleven years, someone has to extract it before anything works.

That is not a reason to abandon the project. It is a cost to put in the estimate. Writing down your service list with real prices, your service area, your hours and the twenty questions you answer every week takes a focused afternoon, and it is the same afternoon whichever vendor you eventually choose. What to write and what to leave out is covered in training an AI agent on your business.

Watch for the opposite problem too. Businesses with a large document library often assume they are ready, then discover half of it is out of date and the AI is quoting a 2023 price sheet. Volume of documentation is not the same as usable knowledge.

Question five: who checks it in week three?#

Every AI project has an enthusiastic week one. The question is who is looking at the output in week three, when the novelty has gone and there is a real job on.

Name the person. Give them fifteen minutes a week and a specific thing to look at: a sample of conversations, an escalation log, a count of enquiries that turned into quoted work. Without an owner, systems drift quietly, and you find out six months later that it has been telling people you do not service Hobsonville.

If nobody in the business can take that fifteen minutes, the honest answer is that you are not ready for this project yet, and a smaller one would serve you better.

The best first AI project for business owners is usually the dull one#

Take a small building company weighing up three options. Score each question from one to five, where five is favourable.

Enquiry response wins, and it usually does. Invoice chasing feels more urgent because unpaid invoices hurt, but the volume is too low and the relationship risk is too high for a first project. Site reports fail on the data question before you get anywhere near the technology.

The second point worth noticing: the winner is not the task the owner complained about. It rarely is.

Run the same table for your own three candidates and resist the urge to adjust a score once you have seen the total. The scoring only helps if you fill it in before you know which answer you want. Where two candidates come out close, break the tie on the data question rather than on volume, because a task with the information already written down can be running inside a month, while a task that needs knowledge extracted from someone's head will take a quarter and stall the first time that person gets busy.

Turn the winner into a project you can finish#

Once you have a winner, keep it small enough to complete inside a month. One task, one owner, one number you are trying to move.

Before you switch anything on, record where that number sits now. Two weeks of baseline data is enough, and without it you will spend month four arguing about a subscription with no evidence. Then work out whether the money makes sense using the ROI method rather than a vendor's projections, and run the thing quietly alongside your existing process before it touches a customer. A staged rollout is laid out week by week in the 90-day implementation roadmap.

If you are still deciding which of the broader options fits your business at all, the overview of AI solutions for small business covers what the technology reliably does today and what each version costs.

Score your three candidates this week. The exercise takes twenty minutes and it will save you a quarter.