There is an uncomfortable fact behind most businesses that feel stuck with AI: the models are not the problem, and neither is the budget. The problem is that the single thing that would make AI genuinely valuable to them, their own data, was never really collected. It streamed past every day for years and went straight down the drain.
This is the practical companion to a point I have made before, that the most valuable thing in AI is the data nobody has yet. That piece argued why your proprietary data is the real edge now that everyone rents the same models. This one is about the unglamorous next question: if that is true, how does a normal business actually start collecting the stuff? Because most are not, and they are leaving their best asset on the floor.
The asset you are already making and losing
Your business generates its most valuable data as a byproduct of simply operating. Every quote you win or lose, every service call, every exception a staff member quietly handles, every correction someone makes to a record, that is a record of how your specific operation works, and it exists nowhere else on earth. A competitor can buy the same software and the same market reports. They cannot buy your history.
And yet most of it evaporates. By common estimates, more than half of the data organisations hold is “dark,” collected and stored but never used, and for some kinds of operational data the unused share climbs far higher, with some analyses suggesting most companies actively use only a sliver of what passes through them. Read that not as a scolding but as an opportunity sitting in plain sight. The raw material is already flowing through your business. Almost nobody is catching it. The ones who start will have something in three years that the ones who did not simply cannot manufacture after the fact.
That last part is the quiet urgency. Data has a vintage. You cannot go back and record the reasoning behind a decision your team made last spring if nobody wrote it down at the time. The history you want to have in 2029 is being created, and discarded, right now.
Five practical ways to start collecting it
None of this requires a data team or a big platform. It requires a few deliberate habits, applied to the work you already do.
Capture the decision, not just the outcome. Most systems record what happened, the invoice, the booking, the closed ticket. Far more valuable is why. Why did this quote get discounted? Why was this order flagged? A single extra field asking “reason” turns a flat record into something you can learn from, because the reasons are where the patterns hide.
Structure it at the moment of entry. A dropdown with five clear options is worth ten times a free-text box, because it produces data that is consistent enough to count. Free text is where meaning goes to hide. The small friction of choosing a category at entry is what saves you from an unusable pile later, and it is exactly the structure an AI needs to read your records well, as the piece on embeddings explains.
Keep the exceptions and the corrections. The gold is not in the routine cases, it is in the edge cases: the order that did not fit the normal flow, the record someone had to fix, the judgment call that overrode the default. Those are the moments that encode your team’s actual expertise, and they are precisely what a future system needs to learn to handle the hard 10% instead of only the easy 90%.
Put it in one place you own. Data scattered across inboxes, spreadsheets, and three peoples’ memories is not really collected, it is just stored, badly. A single system of record, timestamped and owned by you, is what turns scattered activity into an asset that compounds. Owned by you is not a throwaway phrase: data you control is an advantage that stays with the business, not something rented back to you.
Collect it cleanly and with consent. More data is not automatically better, and in Canada it carries real obligations. Collect what you have a clear use for, be straight with customers about it, and hold it under Canadian privacy expectations. First-party data you gathered properly is both the most valuable and the safest kind to build on. Volume you cannot account for is a liability, not an asset.
Why this is the one advantage a competitor cannot buy
It is worth being clear about why this is worth the bother, because collecting data is a cost today for a payoff later, and that is a hard trade to make without seeing the logic.
The logic is that everything else in AI has become a commodity. The models are rented and identical to your competitor’s, which is the whole argument of why a chat subscription is not AI engineering. Compute is a utility. What cannot be commoditised is the specific, hard-won record of how your business actually operates, because you are the only one making it. Data is also the rare asset that compounds: the longer you collect well, the further ahead you get, and the harder you are to catch. A rival can copy your website in a week. They cannot copy four years of your operation’s memory.
That memory is also what makes any future AI genuinely yours rather than generic. An AI reading your clean, structured history can answer from your actual policies, notice your particular patterns, and support your real decisions. The same AI with no data of yours to stand on is just a smarter version of the tool everyone else already has. The data is what turns a rented model into a system that runs your business.
Plant it now
The mistake is treating data collection as something to sort out later, once the AI project starts. By then it is too late to have the history that would have made the project worth doing. The businesses that get remarkable results from AI a few years from now will, almost without exception, be the ones that started keeping good records before they had a use for them, the way a good vineyard is planted years before the first bottle.
You do not need a strategy department to begin. Pick the one workflow that matters most, add the fields that capture the decision and the exception, put the record somewhere you own, and be consistent about it. Capture it, structure it, own it. Do that steadily and you are quietly building the one thing in AI that money cannot buy and time cannot be rushed: a record of your business that is entirely, defensibly yours.