There's a reason this question matters more than it looks: the speed at which a company can answer itself is the ceiling on everything else.

If a question takes three days to answer, the decision that depends on it takes a week, and the change that depends on that decision takes a quarter. And one thing simply never happens: nobody automates what they can't yet query.

Why it takes three days

Not because the information is missing. Because it's scattered, and because the path to it runs through people.

The typical route: you ask. Someone works out which systems hold the data. Someone exports. Someone else cross-references two files because the codes don't match. Someone interprets what to do with the odd cases. Someone assembles the deck. Between every step there's a wait, because each person has their own job and this joins the queue.

Add up the waiting times — not the working times — and there are your three days. The actual work was forty minutes.

And there's a bigger cost that stays invisible: with slow information, supervision becomes a job. Someone has to watch, consolidate, compare and flag, because the system doesn't do it on its own. That person isn't supervising the business: they're supervising the gap between the systems.

What happened to us, and we didn't plan it

Virfon has served more than 150,000 Colombian and Venezuelan immigrants, connecting them with their families. In 2025 it faced a serious EBITDA problem, and decisions had to be made.

Several of its managers moved to related companies with better professional prospects. Nobody was laid off: they were helped into the next step, without a single day out of work. But their supervision duties had to be covered with what remained.

The decision was to build a replica of all the operation's raw data on a Mac mini, and work it with AI and vibe-coding. No corporate data platform. A desktop machine holding all the information unsummarized, un-aggregated, with no advance decision about which questions would be answered.

The expectation was modest: cover most of what those supervisors did.

We didn't replace them. We went far past them.

What showed up once AI looked at raw data

The AI began studying information we had held for years and had never truly seen, because we had always looked at it through already-processed reports. It found things nobody had asked for:

Abusive usage patterns in trial periods. No report was designed to show them, because nobody had thought to ask. Weeks later, once validated, they became an agent that analyzes every new trial and suspends it two hours in if the behavior matches the pattern.

Which customers are genuinely loyal. Not the highest billing ones: the ones who behave as if leaving never crossed their minds. A distinction no revenue report shows.

Churn risk, read in behavior before it shows up as a cancellation.

Business-type consumption patterns inside the base — customers using the service like a company while contracted as individuals. That's up-sell that had existed for years and nobody could see.

And a capability that was previously unthinkable: when a customer calls for service, an agent now hands over their 360-degree picture in seconds. Before, some data sat in billing, some in the help desk, some in accounting, and assembling that 360 for a single customer cost hours. Nobody asked for it, because nobody was going to pay those hours to handle one call.

Why we hadn't done it before

Not for lack of tools or capable people. For a structural reason: we were merging reports instead of analyzing data.

A report is data already processed according to what someone decided mattered. Cross-referencing three reports to answer a new question is thankless, fragile work, and the result inherits every decision made while building each one. That's why the interesting questions went unasked: the cost of answering them was out of proportion to the curiosity behind them.

Raw data doesn't have that problem. It doesn't arrive with the questions already decided.

The lesson worth taking

Everyone wants to start with the decision-making agent. It's the visible part, it demos well, and it's in fashion.

But the agent is the last step. The real order is: query → discover → validate → automate. Without the first there's nothing to discover; without validation there's nothing you can safely leave deciding on its own. An agent built on fragmented information decides on the same three contradictory answers you get today — only faster, which is the worst possible outcome.

And the least impressive detail of the whole story is worth repeating: it ran on a Mac mini. This isn't about how much you invest in infrastructure. It's about no longer querying your business through reports someone imagined years ago.

What it looks like when it works

It doesn't look like a new dashboard. It looks like this: you have a question in a meeting, you ask it, and the answer appears before the meeting ends. And when that answer opens a second question — which is what always happens with good ones — that one gets answered right there too.

That second question is what gets lost today. In the three-day model nobody asks it: it isn't worth starting another cycle over a curiosity. That's where most of the value stays — and that's where the abuse patterns, the loyal customers and the unseen up-sell were.

A question to close on: in your last leadership meeting, how many questions went unasked because everyone knew the answer would take a while?

---