Last week, Barbados’s Senior Minister Kerrie Symmonds put some numbers to something anyone who works with small businesses in this region already suspected. Across the Caribbean, only 4 to 5 per cent of business transactions are conducted digitally, against 18 to 19 per cent in the United States. Citing research commissioned by that country’s Small Business Association, he noted that among Barbadian micro, small and medium enterprises, 78 per cent use social media and customer-facing platforms. Only 13 per cent use payroll software. Only 14 per cent use digitised inventory management. A quarter use no digital management tools at all.

Read those last four numbers together and a pattern appears. We have automated the shop window. The stock room, the payroll and the ledger are still being run by hand.

Jamaica’s picture rhymes. A Mastercard study reported in May found that only 8 per cent of small merchants use point-of-sale systems, and that cash still accounts for roughly 72 per cent of personal consumption spending. Set that against the roughly 422,000 registered micro, small and medium enterprises the Ministry of Finance and the Small Business Association of Jamaica count in this country, and the scale of the gap becomes clear.

What gets automated first versus what actually pays: 78 per cent on social media, only eight per
cent of small merchants on point-of-sale, against 13 per cent on payroll software and 14 per cent on
digitised inventory. (Branded graphic by PGH Consulting, LLC)

The order most businesses choose, and the order that pays

This is not only a Caribbean habit. In a survey commissioned by Intuit last year, covering more than 2,200 American businesses with up to 100 staff, the most common uses of AI were marketing at 43 per cent and customer service at 36 per cent. Bookkeeping came last, at 29 per cent.

Now set that against where the returns actually turn up. The Hackett Group’s 2026 study of finance functions found accounts payable to be the most mature area of AI adoption, with a third of organisations already scaling solutions there, followed by travel and expense management. Financial planning, forecasting and reporting sit well behind.

The reason is not mysterious. Paying suppliers and processing expenses are high-volume, repetitive and rule-bound: the same shape of task, many times a week, with a right answer that can be checked. Forecasting is none of those things, and neither is marketing copy, which is why it feels rewarding to automate and rarely changes the cost base.

So the first useful question is not which tool to buy. It is which of your processes has that shape.

Before automating anything: how often, are the rules stable, is the input consistent, and would
you notice if it went wrong? (Branded graphic by PGH Consulting, LLC)

A four-question test

Before automating anything, ask four things about the process.

How often does it happen? Something you do forty times a week is worth an hour of setup. Something you do twice a year is not.

Are the rules stable? If the way you code an expense changed twice this year and will change again, you are automating a moving target.

Is the input consistent? An agent can read a supplier invoice reliably when the supplier’s name is spelled the same way every time. It cannot when the same company appears in your records four different ways.

Would you notice if it went wrong? This is the one people skip, and it matters most. An automation that fails loudly is a nuisance. One that fails silently, producing plausible numbers nobody questions, is a genuine problem.

Score your five most repetitive tasks against those four questions and the answer usually picks itself. For most small businesses it is supplier invoices, expense handling, or chasing payment.

88 per cent of data leaders are confident their data is ready for AI. 43 per cent say data
readiness is their biggest obstacle. A further estimate: 60 per cent of AI projects without AI-ready data
may be abandoned through 2026. (Branded graphic by PGH Consulting, LLC)

The confidence gap

Here is the part that gets skipped, and the evidence on it is unusually clear.

In January, a study by Precisely with Drexel University’s LeBow College of Business surveyed more than 500 senior data leaders at large enterprises. Some 88 per cent expressed confidence in their organisation’s data readiness for AI. In the same survey, 43 per cent named data readiness as the single biggest obstacle to AI delivering results.

Both figures come from the same people. They are confident, and they are blocked by exactly the thing they are confident about.

Gartner points the same way, predicting in February last year that through 2026 organisations would abandon 60 per cent of AI projects that were not supported by what it calls AI-ready data. The failures it describes are not clever failures of the technology. They are ordinary failures of records.

What speed does to a mistake

The clearest illustration of that remains a trading firm rather than a small business, but the mechanism is the same at any size.

On 1 August 2012, Knight Capital deployed new code with a defect in it. In the first 45 minutes of trading, its system sent more than four million orders while attempting to fill 212. The firm lost more than US$460 million. The detail worth remembering is that 97 automated system emails flagging an error condition had gone out before the market opened. According to the regulator’s account, nobody acted on them, because the firm had never designed those messages to function as alerts and staff did not generally read them. The information existed. Nothing had been built to treat it as a warning.

The United Kingdom’s Post Office Horizon accounting system ran a second version of the same lesson over two decades. More than 700 subpostmasters were prosecuted over shortfalls the system reported, and the courts and a public inquiry have since found the software carried bugs and defects capable of producing exactly those shortfalls. The organisation trusted its system over the people doing the work.

Neither is an argument against automation. Both argue for knowing what a process actually does before you make it faster, and for keeping a person close enough to the output to notice when it stops making sense.

One hour, no software: take your last 100 records, check ten fields, count the errors. That
number is your starting point. (Branded graphic by PGH Consulting, LLC)

The measurement you can run on a Friday afternoon

There is a way to find out how good your data is that costs nothing and takes about an hour, developed by researchers including Thomas Redman and published in the Harvard Business Review.

Take the last 100 records your business produced of one type, invoices say, or new customer accounts. Pick ten to fifteen fields that matter, the ones somebody downstream relies on. Then go through record by record and mark each as correct or not.

When the researchers ran this with 75 executives, on average 47 per cent of newly created records contained at least one critical error, and only 3 per cent of scores reached an acceptable standard. That study is from 2017, and a later replication across nearly 200 measurements found no meaningful improvement.

Your own number is the one that matters. If a third of your last hundred invoices carry an error in a field somebody depends on, you have found your first project, and it is not an automation project.

What the numbers here do not tell you

Two caveats, since this column is partly about not trusting figures uncritically.

Most published statistics about automation failure are shakier than they look. The widely repeated claim that 95 per cent of AI pilots fail comes from a preliminary working paper, misreads that paper’s own data, and has been publicly disputed by academics in the field. Several other famous numbers here trace back to a conference remark or an uncited magazine line. Treat any dramatic failure rate you see quoted, including in this newspaper, as something to check.

The figures above are better sourced, but most survey large organisations. The direction they point is useful. The absolute numbers are not your benchmark.

What to try this week
1. Run the measurement described above on your last 100 invoices or customer records. Ten fields, one hour, count the errors. Whatever number you get is the most useful thing you will learn about your business this month.
2. List your five most repetitive tasks and score each on the four questions: frequency, stable rules, consistent input, and whether you would notice a failure.
3. Before automating anything, clean the master data it depends on. One spelling per supplier, one code per expense type, one version of each customer record.
4. Pick the highest-scoring process, which is usually supplier invoices or expenses rather than anything customer-facing, and automate one step of it, not the whole thing.
5. Run that one step alongside your existing process for a full month and compare the two before you switch over.
None of this requires new software, and the first three steps cost nothing but attention. Automation applied to a process you have not examined will reproduce whatever is wrong with it, at speed and at scale, which is the opposite of what you were trying to buy.

Peta-Gaye Hardy is the founder of PGH Consulting, LLC, where she helps finance and operations teams adopt AI in practical, low-risk ways. She writes the weekly AI in Finance & Business column and is based between Jamaica and the United States. Learn more at www.pghconsultinggroup.com. Follow on Instagram and YouTube @pghconsultinggroup, and connect on LinkedIn at linkedin.com/in/peta-gaye-hardy.

Disclosures: This article is informational and does not constitute investment, tax, legal, or accounting advice. Readers should consult a qualified professional before acting. Regional digitisation figures are as stated publicly by Senior Minister Kerrie Symmonds in August 2026, citing research commissioned by the Small Business Association of Barbados. Jamaican payment figures are from a Mastercard study reported in May 2026; the count of registered micro, small and medium enterprises is attributed to the Ministry of Finance and the Small Business Association of Jamaica. Research findings are attributed to Precisely and Drexel University’s LeBow College of Business (January 2026), Gartner (February 2025), The Hackett Group (March 2026), Intuit (April 2025), and Nagle, Redman and Sammon in the Harvard Business Review (2017); all are as published at the time of writing and most survey large organisations rather than small businesses. The Knight Capital account follows the United States Securities and Exchange Commission’s 2013 administrative order. AI tools can produce errors, and every figure, clause, or claim they produce should be verified against a source before being shared or acted upon. The author has no commercial relationship with any company or product mentioned and was not compensated by them. The examples described are illustrative and do not depict any real business.