70% of companies in Colombia aren't ready for AI (and it's not an AI problem)

A recent study says 7 out of 10 companies in Colombia aren't ready for AI. The real reason, and a checklist before hiring an AI project.

A recent study from Noticias RCN found that 7 out of 10 companies in Colombia aren’t ready to use artificial intelligence. That’s a scary number, especially if your company already feels the pressure to “get on the AI train” before falling behind. But before rushing to buy a model, a chatbot, or a generative AI plan, it’s worth asking a more uncomfortable question: ready for what, exactly?

Because when you look closely at what these studies mean by “not ready,” they rarely mean a lack of technology. They mean something much more mundane and far more urgent: scattered data, reports that don’t match across departments, processes that live in one person’s head. AI doesn’t fix that. It exposes it.

The problem isn’t a lack of AI — it’s a lack of a foundation

Think about how most mid-sized companies in LATAM actually operate today. Sales keeps its own spreadsheet. Finance keeps a different one. Operations runs a third, separate one — and sometimes they don’t even agree on how many active customers exist this month. Every department “knows” its own number, but nobody has the real number: the one that would come out the same if three different people checked it on the same day.

That’s the actual starting point for most mid-sized companies. And it’s exactly the ground an AI project can’t stand on. Not because the technology is bad, but because a trained model — or even a simple generative AI automation — needs a source of information it can trust. If that source doesn’t exist, the model inherits the mess. It processes it faster, presents it with more confidence, but it’s still the same old error.

This is a pattern we see over and over: a mid-sized company hires an AI project with high expectations — automating reports, predicting demand, answering customer questions. The technical team delivers something functional in the demo. Three months later, almost nobody is using it, because the sales team doesn’t trust the numbers the model is working from. The project didn’t fail because of the model. It failed because there was never a single source of truth to automate on top of.

Why do companies skip this step?

It’s understandable. AI sells. “Let’s clean and centralize the database before automating anything” doesn’t generate the same excitement in a leadership meeting as “let’s roll out an AI assistant for the sales team.” One sounds like boring infrastructure spend. The other sounds like innovation.

The problem is that order matters. Building on unreliable data isn’t actually faster — it only looks faster at first, until the project reaches production and nobody trusts what it says.

The sequence that actually works

1. One single source of truth per critical process. You don’t need to centralize the whole company at once — pick the process that hurts the most (sales, inventory, collections) and make sure everyone looks at the same number.
2. Clean and current data, not perfect data. You don’t need a corporate-grade data warehouse to get started. You need information that’s complete, free of obvious duplicates, and updated on a reasonable schedule.
3. A process, not just a tool. Centralizing data does nothing if every department keeps updating its parallel spreadsheet “just in case.” It has to be clear who updates what, and when.
4. Only then, automation or AI. Once a reliable foundation exists, that’s when a model, an automated dashboard, or an AI assistant generate real value — because they’re standing on something solid.

A concrete example: how many active customers does a company really have

Here’s a case that illustrates this, names changed: a mid-sized services company, around 120 employees, wanted an AI dashboard to predict how many customers would renew next quarter. Before touching the model, one simple question had to be answered: how many active customers do they have today?

Sales said 340, counting everything in the CRM without filtering by status. Finance said 287, counting only customers billed in the last 90 days. Operations, which kept its own spreadsheet because “the CRM is never up to date,” said 301. Three answers, three different numbers, for the same question on the same day.

That 53-customer gap between the highest and lowest figure isn’t a minor detail — it’s the difference between an optimistic revenue forecast and a realistic one. No prediction model, however sophisticated, can resolve that ambiguity on its own: feed it sales’ number and it predicts on an inflated base; feed it finance’s number and it ignores customers operations knows are still active through another channel.

The fix wasn’t a better model. It was getting all three departments in the same room to agree on a single rule for what counts as an “active customer” (billed in the last 90 days or an active contract without billing during a seasonal lull), and making sure all three teams pulled that definition from the same system. Only once that was solved — which took two weeks of work, not months — did the AI prediction project have real ground to stand on. The model itself was built afterward, and it worked, because it finally had a reliable starting number.

When does this apply, and when doesn’t it?

Not every company needs to pause everything before touching AI. If your company already has a central system (an ERP, a well-maintained CRM, even a basic data warehouse) where departments genuinely trust the numbers, the bottleneck probably isn’t data — you might already be ready for a more ambitious AI project.

The real warning sign is different: if asking “where does this number come from?” gets a different answer depending on who you ask, that’s the problem to solve first. There’s no technological shortcut for that — not even the most expensive model on the market fixes a data source nobody maintains.

This also doesn’t apply to small, well-scoped projects. If the goal is automating one specific, narrow task — classifying incoming emails, or generating a weekly summary from a single report that’s already reliable — you don’t need a full data governance project before starting. The “foundation first” diagnosis mainly applies to projects that cross information from multiple departments or make business decisions based on that data.

What you can do this week

You don’t need a months-long project to take the first step. This week you can:

• Pick one number that’s critical to the business (active customers, this month’s sales, available inventory).
• Ask two people from different departments what that number is today.
• If the answers don’t match, you’ve just identified the real first problem to solve — before any conversation about AI.

That five-minute exercise says more about how ready your company is for AI than any general study ever will.

FAQ

Do I need a full data project before using any AI tool?

Not always. For narrow, well-scoped tasks with a single source of data that’s already reliable, you can automate directly. The “foundation first” diagnosis applies when the project crosses information from several departments or makes business decisions based on it.

How long does it take to have a reliable data foundation?

It depends on company size and how many parallel systems exist today, but for a single critical process (not the whole company) it’s usually a matter of weeks, not months, if the scope stays narrow.

Does this mean it’s not worth starting with AI yet?

No. It means it’s worth starting with what actually moves the needle: knowing with certainty what’s happening in the business today. That already creates value on its own, with or without AI on top.

---

If you’re interested in the angle of why buying a big vendor’s full package doesn’t always solve the underlying problem, Raifen wrote about vendor lock-in vs. open stack last week — same principle, different angle.