Ask a large firm about its AI strategy and, somewhere in the first five minutes, you'll hear it: "we need to get our data in order first." I've heard that sentence from firms sitting on twenty years of delivered projects. The cleanup never starts (or never finishes), and the AI budget gets spent anyway, on pilots that never touch those twenty years of work.
Here is the arithmetic I would put on the whiteboard instead. The gains everyone is chasing come from two factors multiplied together: the data you hold, and the AI you run on it. Look hard at the second factor. It is rented. The same models, the same agents, the same copilots are available to your fiercest competitor at the same subscription price, and they improve for both of you on the same day. Nobody in this industry is going to out-model anyone else. Training your own foundation model is the wrong first move for almost everyone; the market for multipliers is efficient, and it gets more efficient every quarter.
The arithmetic
ROI = Your data × AI. One factor is rented. One is yours.
So if every firm holds the same multiplier, where does advantage live? In the only factor that differs. Your data.
And your data inheritance is bigger than you think. It isn't one Revit model. It is every project your firm has ever priced, drawn, coordinated, built and corrected: thousands of decisions, sitting in whatever formats the last two decades left behind. Nobody else has that. There are only two states to be in with respect to it: years of it inert on a file server, or never having started accumulating it in usable form. Both feel safe. Both lose to the competitor who started.
The standard objection is that the archive is a mess. We've all accidentally used five wall families instead of one because we need to move fast. All true, and beside the point, because multiplication is forgiving in exactly the way a purity test isn't. A messy data catalog times today's multiplier beats a missing one every single time. A multiplier times zero is zero, and zero is what the "data in order first" firm has been running for years, at full subscription price. Process everything, lightly and quickly, first. Patterns surface across a hundred imperfect projects that no lovingly cleaned model will ever show you: your actual net-to-gross across every residential scheme you've delivered; the details that keep reappearing in your fastest jobs; how much of each "new" building your teams quietly re-model from scratch.
Multiplication has no order of operations. "Data first, then AI" does not make sense. Both terms have to be on the board at the same time, and only one of them is yours to grow.
The Test
💡 So, this issue's test. Pick ten delivered projects, any ten, and run the lightest extraction you can across all of them. Then ask that portfolio one question you couldn't answer last week. If the answer surprises you, you've just measured the one term in the equation that belongs to you, and found out how little of it you've been using. Making that extraction take minutes instead of months, on whatever formats the last decade left you with, is the part we've spent years making boring at Speckle. 💡 For the comments: what's the first portfolio question you'd ask if your archive could answer back? And if you think delivered work is a commodity and I'm overvaluing what firms are sitting on, make that case - I'd rather hear it now.