AI automates the mess too

AI tools now promise to answer your emails, chase your invoices and sort your documents. Maltese businesses are trying them: in 2025, 21.5% of Maltese enterprises with 10 or more staff used at least one AI technology, slightly above the EU average of 20.0% (Eurostat). But 65.9% of Maltese SMEs say neither they nor their staff have had any AI training (Malta Chamber of SMEs and MISCO, SME Barometer, Q1 2025). The tools are arriving faster than the know-how.

An old lesson with a new tool

In 1990, Michael Hammer wrote in the Harvard Business Review that companies were spending heavily on computers and getting little back, because they used them to speed up old ways of working. His advice was to stop asking “how do we do this faster?” and start asking “should this step exist at all?”

AI is the same story with a new tool. The economists Ajay Agrawal, Joshua Gans and Avi Goldfarb make the point in Power and Prediction: dropping AI into one task, inside a workflow that stays the same, gives small gains. The big gains come from rethinking how the work fits together. AI doesn’t fix a messy process. It runs it faster, mistakes included.

It’s a conversation we have often. A business owner tells us they tried AI and gave up, because it was taking more time than it saved. When we ask what they used it for, the same pattern comes up: they had a problem and asked the AI for a solution. The question is usually something like “How do I get clients to pay on time?”, with no context and no details. The AI answers with a generic list: send reminders, offer a discount for paying early, ask for deposits. None of it fits their business, so they ask again, get another list, and give up. A question asked once the process is clear sounds different: “We invoice by email at the end of each month, about a third of clients pay after 60 days, and reminders only go out when someone remembers. Draft a polite reminder for days 7, 14 and 30.” That one gets a useful answer in a minute. The difference isn’t the tool. It’s knowing your process well enough to describe it.

Five things a process needs before you automate it

1. A clear start. Everyone knows what triggers it: an enquiry, an order, a new client.
2. The same inputs every time. A defined list of what’s needed, not “whatever the client sends”.
3. Written steps and rules. As Atul Gawande shows in The Checklist Manifesto, most failures come from skipping a step everyone knows, not from lack of knowledge. If you can’t write the checklist, you can’t automate it.
4. Known exceptions. What happens when something is missing or unusual, and who decides.
5. One owner and one measure. One person responsible, and one number you check each month to know it works.

Score your process out of five. Five: it’s ready to automate. Three or four: fix the gaps first, it’s usually quick. Two or fewer: the process comes first, and the tool can wait.

Then start small, and safely

Ethan Mollick, who studies how people use AI at work, describes its abilities as a “jagged frontier”: surprisingly good at some tasks, surprisingly weak at others, and the only way to know is to try it on your own work. So start with low-risk tasks such as drafting replies, summarising or sorting documents. Keep personal data out of public AI tools, and have a person approve anything that goes to a client. AI drafts; people decide.

A simple test for this week

Pick one task you’d love AI to take off your hands. Run it through the five points above. If it scores five, try it. If not, you’ve just found what to fix first.

Ideas from: Michael Hammer, “Reengineering Work: Don’t Automate, Obliterate”, Harvard Business Review (1990); Ajay Agrawal, Joshua Gans and Avi Goldfarb, Power and Prediction (2022); Atul Gawande, The Checklist Manifesto (2009); Ethan Mollick, Co-Intelligence (2024). Figures: Eurostat (2025); Malta Chamber of SMEs and MISCO, SME Barometer Q1 2025 (427 businesses).