How to tell if your company needs AI, a simple automation, or just cleaner data

We hear the same sentence a lot: "we want to add AI." It almost never comes with the problem AI would solve. And that's the trap. The question isn't "what technology do we install?" but "what is costing us money or time today, and why?" Flip that order and the project gets chosen by trend, not by need, and ends up expensive.

The numbers back the warning. Gartner projected that at least 30% of generative AI projects would be abandoned after the proof of concept by the end of 2025, due to poor data quality, runaway costs, and unclear business value. An S&P Global survey showed something worse: the share of companies that scrapped most of their AI initiatives before reaching production jumped from 17% to 42% in a single year. It's not that AI doesn't work. It's that it gets bought before the problem is understood.

Almost everything an SME wants to fix fits into one of three layers. The first is data: information that's scattered, duplicated, or sitting in spreadsheets nobody reconciles. The second is automation: a repetitive, clear-rule task a person does by hand every week. The third is AI: decisions or text that require interpreting language, images, or patterns you can't write down as a fixed rule.

The order matters because cost and risk climb at every step. Cleaning data is the cheapest move and almost always the most profitable. Automating a rule-based process is predictable and pays back fast. AI is the most powerful tool, but also the most expensive to maintain and the one that depends most on the two layers beneath it being healthy. Skipping the first two to jump straight to AI is like buying a race car for an unpaved road.

Classic sign: two departments report the same number differently. Sales says one thing, finance says another, and nobody knows which to believe. Or the same customer shows up three times spelled three ways. Or month-end closes by copy-pasting across five spreadsheets. That's not an AI problem. It's a data problem, and layering AI on top just amplifies the mess faster.

This isn't theory. Gartner estimated that poor-quality data costs organizations around US$12.9 million a year on average; that figure comes from large enterprises, but the mechanism is the same in an SME at a different scale: late decisions, rework, and missed opportunities. The fix is usually tidy and tech-free: a single source of truth, consistent fields, and a couple of validations when data is entered. Boring, yes. But it fixes the root.

If the task can be described as "when X happens, do Y," it's almost certainly automation, not AI. Copying orders from email into a system. Generating the same invoice every month. Pinging WhatsApp when stock drops below the minimum. Moving form data into a spreadsheet. These are fixed-rule tasks with no ambiguity, and they're solved with tools that already exist and cost little.

It's worth it because the wasted time is real: a Smartsheet survey found that over 40% of workers spend at least a quarter of their workweek on manual, repetitive tasks. In a ten-person SME, that's roughly a salary and a half going into copy-paste. Automating the repetitive stuff isn't a luxury; it's often the fastest-paying investment, and it sets the ground in case AI does make sense later.

AI wins when you need to interpret something that won't fit in a rule: reading hundreds of customer emails and grouping them by intent, summarizing long contracts, answering questions about your own documents, classifying product photos. There the problem is language and patterns, and no rule-based automation solves that on its own.

But there's a condition almost nobody mentions in the sales pitch: generic AI, without your data and your context, gets it right rarely. According to data.world, language models answer real business questions correctly only around 16.7% of the time when they aren't grounded in a well-structured database. Translation: if your data is dirty, AI inherits the mess and delivers it with a confident tone. So even when the answer is AI, the first step is almost always the data layer. AI is the top floor of the building, not the foundation.

Before you quote anything, ask three questions in this order. One: do I trust my data, and does it come from a single source? If not, start there. Two: can the task be written as an "if this happens, do that" rule? If yes, it's automation, not AI. Three: is what's missing the interpretation of language, images, or shifting patterns? Only then does AI enter the conversation, and always grounded on clean data.

Sometimes the honest answer is "don't buy anything yet": clean up, measure for a couple of months, and look again. That saves more than any tool. If after these questions you're still unsure which layer you're in, that's exactly the conversation worth having: a short, no-strings diagnostic chat to put a name to the real problem before you spend. If it helps, let's talk it through.