61% of companies now "use AI." How long does yours take to answer a simple question?

A study published this month found that 61% of Spanish companies now use AI, seven points above the European average. Good news, and we won’t qualify it: the more people try these tools, the better. But there’s a long way between “using AI” and something far more concrete: anyone in your company being able to ask a question about the operation and get a reliable answer while they still care about it. The first one gets answered in a survey. The second one gets measured with a stopwatch.

The stopwatch test

The test is simple, and anyone can run it this week. Pick an ordinary business question, the kind that comes up in any meeting: how much did we sell last quarter in this region, by product family? Ask it the way it always gets asked — the colleague in admin, the IT manager, whoever it is that “knows how to pull that” — and start the stopwatch. Stop it when you’re holding an answer you trust enough to make a decision with.

In most companies where we’ve run this exercise, the stopwatch doesn’t measure minutes. It measures days.

It’s worth watching where the time actually goes, because it’s not where it seems. The ERP query itself runs in seconds. What takes days is everything around it: the email to the person who knows how to write the query, the request queue that email lands in, the gap in their calendar, the export to a spreadsheet, the reply, and — often — a second round trip because the question wasn’t quite that one. At an industrial distributor we work with, that full loop took between one and three days per question. Not per hard question: per question.

The bottleneck isn’t technical

The first thing to say about that loop is that nothing in it is a technology failure. The data exists, it’s in the ERP, it’s reasonably clean, and the query is trivial. What you’re looking at is an organizational toll booth: only two or three people know how to get the data out, and everything the company wants to know about itself goes through their inbox.

That toll has a visible cost and an invisible one. The visible cost is the hours of the people acting as translators between the business and the database, answering recurring questions one at a time. The invisible cost is worse: all the questions nobody ever gets around to asking. When getting a number costs three days and a favor, people stop asking. Decisions get made on intuition, on habit, or on last year’s figure — not because anyone is careless, but because the loop turns every question into paperwork. A company can be “using AI” for marketing, support or email drafting and still run like this on the inside.

What we built — and why it didn’t start with AI

At that distributor we built a natural-language interface on top of the ERP: any authorized employee types their question in plain language and gets the answer directly. But the work that made that system possible wasn’t model work, it was operations work, and we think the order matters enough to tell it that way.

Before writing any code, we mapped the questions the organization was already asking: the ones emailed to IT, the ones repeated at every month-end close, the ones each department kept in its own spreadsheet. About twenty business questions accounted for the vast majority of requests — along with a less comfortable discovery: many of them were the same question phrased differently, and not always with the same definition behind it. What exactly counts as an “active customer,” what’s included in “margin,” when does an order count as “delivered.” Settling those definitions with the people who use the data every day was the hard part of the project. The interface came afterwards, with its guardrails — role-based access control, query validation, result verification, because a fast wrong answer is worse than a slow right one. And the rollout went in phases — sales team first, then department by department — instead of a company-wide launch day.

The result is measured, not narrated

The before and after fits in one line: a question to the ERP went from taking one to three days to taking seconds, and stopped depending on anyone’s calendar. That was the goal, and it’s verifiable — which is the only thing we ask of any project: if it can’t be measured, it didn’t improve.

What interests us most about this case, though, is the effect that wasn’t in the goal. Once asking stopped costing three days, people started asking more — and, above all, asking better. Follow-up questions. Comparisons nobody used to bother requesting. Hypotheses checked in the same meeting where they came up. The system didn’t make anyone smarter; it removed the toll booth that was keeping the intelligence already in the room from being used.

Adopting is not accessing

Back to the 61%. If your company is in that group, the interesting question isn’t whether it “uses AI” but where: at the edges — text, summaries, email — or at the center, where the operation’s data lives? Our experience, with five systems in real production and zero failed implementations, is that the big value sits at the center, and that getting there is almost never a model problem. It’s an operations problem: knowing which questions your organization actually asks, what each business term really means, and where the days of waiting leak out today. AI is the piece you put in at the end, once all of that is clear. Operations first; technology second.

So the practical takeaway is the one we started with: run the stopwatch test. One ordinary question, the usual loop, the real time until a trustworthy answer. That number — days, almost certainly — is the distance between the AI your company “uses” and the AI it hasn’t used yet.

What did the stopwatch read at your company?

If you’d like to run this against your own data, here’s how we approach data intelligence: we start by mapping the questions your organization already asks — and the ones it has stopped asking because they cost too much — before proposing a single piece of technology.