AI doesn't replace people, it replaces bad decisions

Why the fear of falling behind on AI leads mid-sized companies to buy tools with no real business problem behind them

A few weeks ago, a company laid off 21,000 people and replaced them with artificial intelligence. The headline traveled fast. Around the same time, local media in Colombia reported that a new company adopts AI every five minutes in the country.

Two stories, one effect on any manager reading them: the sense that something has to be done with AI, right now, before falling behind. And that is exactly the moment when mid-sized companies — the 50-to-500-employee kind, without an in-house data team or a multinational’s budget — make the worst technology decisions of the year.

Not because AI doesn’t work. Because fear is a bad starting point for any investment decision.

The problem isn’t the technology, it’s the order of the questions

When an AI decision starts from the fear of falling behind, the order of the questions flips. The tool gets picked first — the one from a LinkedIn post, the one a vendor sold best, the one “everyone is using” — and only afterward does anyone look for where to plug it in.

That order almost guarantees one of two outcomes: either the tool doesn’t match any real process in the company and ends up unused, or it half-matches and creates extra work instead of saving it — someone has to review what it produces, fix it, explain to a client why the chatbot said something strange.

The right order runs the other way: problem first, tool second. Which process is costing the company more hours, more money, or more errors than it should? Only then does it make sense to ask whether AI — or any other technology — is part of the answer. Sometimes it is. Sometimes the real problem is that two systems don’t talk to each other, and no language model fixes that.

This confusion isn’t exclusive to small companies. Studies from international consulting firms on generative AI adoption in large corporations show something similar: most initiatives still haven’t demonstrated tangible value, largely because they launched as a response to competitive pressure rather than a concrete use case with a business owner behind it. If that happens to companies with dedicated teams and large budgets, the risk is even bigger for a company with no room for experiments that don’t pan out.

Why the fear of replacement doesn’t translate the same way to mid-sized companies

The 21,000-layoff headline gets misread when it’s applied to a different scale of company. In an organization of 5,000 employees, replacing entire processes with AI is an organizational architecture decision: multi-year budget, a dedicated team, risk governance, transition plans. It’s a decision made with scale as an ally.

In a company of 100 or 200 employees, there are almost never “spare positions” to replace. What there usually is, in most cases we see, is overloaded people: the person who builds the sales report by hand every Monday, the one answering the same type of email twenty times a week, the one reconciling two systems that were never properly integrated.

In that context, AI doesn’t take anyone’s job. It gives hours back to someone who was already drowning. That’s a difference of substance, not nuance: optimizing an operation that works but is saturated is one conversation, and redesigning an entire organization around automation is a completely different one. They carry different risk levels, different budgets, and different time horizons — and confusing the two is the first way to spend badly.

What a real use case looks like (and what a fear-driven one looks like)

Signal	Real use case	Fear-driven reaction
Where it comes from	A specific process costing measurable time, money, or errors	A headline, a demo, the pressure of “everyone’s doing it”
Project owner	Someone from the operational area who will actually use the result	An innovation or tech team disconnected from day-to-day work
How success is measured	Hours saved, errors reduced, money recovered — defined before starting	”We now have AI” as an achievement in itself
Size of the first step	A narrow process, with a visible result in weeks	A large platform, with results promised “in the medium term”
What happens if it doesn’t work	Adjusted or dropped with little loss	Stays installed anyway, because it’s already paid for and needs justifying

None of these rows depends on how sophisticated the technology is. They depend on whether the project started from a real problem or from a reaction.

When does AI make sense, and when doesn’t it?

It makes sense when there’s a repetitive process, with enough volume to be worth it, and an owner who can clearly say “this costs us X hours a week” or “this generates X errors a month.” It also makes sense when the company already has its data organized in one reliable place — because without that, any AI inherits the same disorder the operation had before.

It doesn’t make sense when the main motivation is competitive (“our competitor already has it”) without an identified problem of its own behind it. It also doesn’t make sense when company data is still scattered across loose spreadsheets, systems that don’t talk to each other, or information that only lives in one person’s head — in that scenario, the step that’s actually needed first is organizing the data, not automating on top of the disorder.

And there’s a third, more subtle case: when the process in question depends on human judgment that’s hard to standardize — a negotiation, a decision with legal implications, a relationship with a key client. There, AI can be useful support (summarizing information, drafting something), but replacing the entire decision usually costs more than it saves, in money and in trust.

How to build a first project that isn’t a reaction

If reading this far made you recognize that your company is in reaction mode — evaluating or buying tools with no concrete process behind them — there’s a simple way to reset the conversation before spending budget.

First, list the processes that repeat most and eat the most time. No formal assessment needed: just ask each team “what do you do every week that feels like it takes longer than it should?” In most 50-to-500-employee companies, that list comes together in under an hour of conversation, and it tends to repeat: reports built by hand, reconciliation between systems, answering the same customer questions, tracking orders or inventory.

Second, put a number on each process on that list. How much time per week, how much money, how many errors. Without that number, there’s no way to know later whether the project worked — and without that measurement, any tool that gets bought stays installed by inertia, never actually evaluated.

Third, pick the process with the best impact-to-effort ratio, not the most ambitious one. The point of the first project isn’t to impress anyone — it’s to produce a measurable win that gives real grounding to the second project. A company that automates a small process well and can show the numbers has more internal credibility for the next step than one that launched a big platform and still can’t say whether it worked.

This order — list, measure, pick by impact/effort — doesn’t require a data team or a big budget. It requires resisting the urge to start with the tool, and starting with the problem instead.

This week’s takeaway

Before the next meeting where someone proposes “adding AI” to something, ask that proposal one question: which specific process does it solve, and how will you measure whether it worked? If the answer is clear and narrow, there’s a real project there. If the answer is “so we don’t fall behind,” there isn’t a project yet — there’s a reaction, and reactions rarely make good investment decisions.

FAQ

Is AI actually replacing jobs in Colombia and Latin America?

In some industries and very specific tasks, yes — but mostly in repetitive, high-volume functions inside large companies. In mid-sized companies across the region, the dominant pattern looks different: AI gets used to absorb work that previously wasn’t getting done well, not to eliminate existing positions.

How do I know if my company has a real AI use case?

If you can name a specific process, say how much time or money it costs today, and say who would own the result, you probably do. If the answer starts with “we should have some AI because…”, you don’t yet.

Do I need a data team to get started?

Not for the first project. You do need the data for that specific process organized and accessible — you don’t need a full data infrastructure before taking the first step, but that first step does need to rest on reliable information.

What size should a first project be?

One that can be measured in weeks, not quarters, with a visible result for whoever asked for it. Starting big without having tested something small is the most common way to waste a first AI budget.

---

Want to dig deeper into how to spot a real AI use case for your operation? You can also check our article on vendor lock-in vs. open stack to learn what questions to ask any vendor before committing to a tool.