Here's the thing nobody warned me about back then: the software was never the hard part.
When we set out to digitize that company, I assumed the technology would be the mountain. It wasn't. We could build the systems. We had good people and enough room to make mistakes and recover from them. What we couldn't shortcut, no matter how good the engineering got, was the thing sitting underneath all of it: years of methodology, verified data, and a hard-won sense of what a number actually means and whether you can believe it.
You don't vibe-code that. You certainly don't vibe-code a ratings agency.
I think about this constantly right now, because AI has made it look like anyone can build a prequalification tool over a weekend. And honestly, they can build something. A GC's internal team can wire up a form in an afternoon. An AI startup can wrap a model around a questionnaire and put the word "platform" in the pitch deck. The demo looks great. I've seen a few, and some of them are genuinely slick.
But a prequalification isn't a form. The form is maybe twenty percent of the thing, and it's the twenty percent you can see. The part that matters is the eighty percent that never shows up in a demo.
What that eighty percent actually is
Let me be specific, because it's easy to wave your hand at "data" and move on.
It's the data itself, first. Ten years of subcontractor financials, safety records, project history, and qualification decisions across thousands of GC and sub relationships. You can't synthesize history. A tool built last month has nothing real to reason over, and no amount of compute conjures up data that was never collected.
It's verification, which is the unglamorous one nobody wants to talk about. Anyone can collect self-reported numbers. We check the data a sub submits against the source documents behind it: the financial statements, the certificate of insurance, the bonding letter, the OSHA logs. Knowing how to read those documents, pull the right figures, normalize them so a profile compares cleanly against every other one, and catch the number that doesn't match what was claimed, that's a craft. And turning that craft into something that runs at scale isn't the easy part of the software, it's the hardest engineering we do. You build it by doing the work over and over, and being on the hook when you get it wrong. There's no weekend version of that.
It's the network, and the fact that it's standardized. The whole thing compounds because a sub fills it out once and every GC reads the same profile. A newcomer starts with neither side of that and an empty database.
And it's trust, which is the slowest of all. Risk managers put real liability on the line when they lean on this data. They trust it because of track record, not because somebody shipped a clean interface. Trust is the one thing you genuinely cannot fake, and it's usually the first thing people assume they can skip.
AI doesn't level this. It widens it.
Here's the part I find most interesting, and a little counterintuitive. The instinct is that AI lets a newcomer leapfrog all of this. In a data business, it does the opposite.
An AI assistant is only ever as good as the verified data it can reach. Point a very capable model at thin or unverified data and you get answers that are confident, fluent, and wrong. In prequalification, the confident wrong answer is the dangerous one, because someone is about to stake a contract on it. So as AI spreads, the companies sitting on deep, verified data get more valuable, not less. The vibe-coded tool gets exposed faster than it used to, because now it's generating authoritative-sounding output on a foundation of nothing.
Do the hard part first
Which is, I'll admit, why I'm more excited about what comes next than nervous about it.
The verified data, the methodology, the standardized network, the trust that took years to earn, none of that is something we're scrambling to assemble now. It's the foundation we've been building at COMPASS from the start. AI doesn't replace it. It makes it more valuable. We're leaning into AI deliberately, and we can do that with confidence precisely because what sits underneath is already verified, already standardized, already trusted. That's the whole difference between an assistant that guesses and one that actually knows.
And the payoff doesn't land on us, it lands on the people doing the work. The sub who fills it out once and gets read fairly by every GC on the platform. The risk manager who can put their name behind a decision because the number underneath it holds. That's who the hard part is for.
So the work goes on the way it always has. Do the hard, unglamorous part first. Then let the new stuff build on top of something real. The good news for anyone willing to do it that way: the foundation is the one part nobody can skip, including us.
“As AI spreads, the companies sitting on deep, verified data get more valuable, not less.”
— Dave Warford
