Too many tools, no system

The average small business now runs a median of five AI tools. That is usually the problem, not the progress.

A survey this March put the average small business at a median of five AI tools. Every owner I meet in Trinidad is somewhere near that number: something for social posts, something for writing, a chatbot a nephew set up, a scheduling thing, and a transcription app nobody remembers subscribing to.

None of it is connected. The bookings live in one place, the client's history in another, and the money in a third, so a member of staff spends their day carrying information between them by hand. Adding a sixth tool makes that worse, not better.

The pattern has a name in this year's post-mortems: tool sprawl with no real system of record. It sits alongside automating a process that was never stable, and the brittleness of stacks glued together with no-code connectors, as the reasons small business automation quietly stops being used.

The test

Pick your busiest day last month. Answer these without asking anyone, in under a minute each:

• How many customers came in, and what for?
• What did we take, and what is still owed?
• Which regulars have not been back in six months?
• Who on the team did what?

If any answer takes longer than a minute, that is your project. It is almost never the thing you were about to buy.

What to fix, in the order that works

One. Stop the double entry. Anywhere the same fact is typed twice, that is first. It is the cheapest to fix and the fastest to feel. A clinic I work with had every website enquiry arrive by email and then get written into a diary by hand. That is not a technology problem, it is a plumbing problem, and it went away in a day. The enquiries now land in the patient record before anyone picks up a pen.

Two. Make one place true. One record per customer, holding everything about them. This is the system of record, and without it every clever thing you buy afterwards has nowhere to put what it learns. It is also what ends the twenty minute hunt for a file.

Three. Let money be seen. What was earned, what was collected, what is still owed, on a screen your responsible person can open without preparing anything. Most businesses discover unpaid work at this step. Not because anyone was dishonest, because nobody had a list.

Four. Then the clever things. Assistants, follow-up sequences, analysis. These sit on top of the three above. Buy them first and they fail, which is exactly what the failure statistics are describing.

Automate the process you actually run, not the one you describe

Ask an owner how a booking works and you get the official version. Watch the desk for an hour and you get the real one: the regulars who text a specific staff member directly, the price that changes for the man who has been coming for nine years, the appointment written in pencil because it might move.

Automate the official version and the staff will keep running the real one alongside it, and now there are two systems. That is not staff being difficult. It is a design failure.

So the sequence is: watch the job being done, write down what actually happens including the exceptions, then build for that. If a rule cannot survive contact with the way people really work, the rule is wrong.

Make the software refuse the thing that causes the problem later

The most useful lines of code I write are usually refusals.

A policy cannot be logged without its number and at least one document attached, because a record with no paperwork is exactly the thing that goes missing when someone needs it in six months. An invoice cannot be raised at a made up price, because the price comes from a loaded price list, four hundred and thirty four items in one case, rather than from whoever is at the counter.

Owners often flinch at this. It feels like the system is being awkward. It is the opposite: it is the one moment where being awkward is cheap. Chasing a missing document three months later is not.

What good looks like

You should feel the change in a fortnight. If a change takes three months to show up, it was too big, and you should have done a smaller one. Two hours a week back, permanently, from removing a single piece of double entry beats a platform nobody logs into.

And when someone offers to sell you the sixth tool, the question is not what it does. It is: what does it write to, and what reads from it. If the answer is "it stands alone", you are about to buy another island.

Sources: Kaizen AI Consulting, small business AI in 2026, Automaton Agency on small business automation failure patterns.