Subatomic 101 series · Post 2 · Karl Simon, CTO
In the first post I drew the line between LLMs, agents, and AI Co-Workers. This one is about the thing underneath all three — the part nobody puts on a slide because it isn’t exciting, and the part that decides whether any of it works.
Your data.
I’ll say the thing I say on every call: AI on bad data doesn’t fail quietly. It fails confidently. It gives you a wrong answer faster than you could have gotten one yourself, in a complete sentence, with no hesitation. AI is doing exactly what it’s supposed to do. It’s reasoning over what you gave it. If what you gave it is a mess, the reasoning is a mess — just quicker.
So before we talk about hiring AI, I want to teach the part that comes first. There’s a well-understood way to get data right. Subatomic performs true standardization, unification, and cleansing through a medallion architecture, and the concept is simpler than the name.
The problem it solves
Walk into almost any firm and the data looks like this: a CRM, core software, email and a note taker or a pile of documents that somebody swears are the real source of truth. Each system is internally fine. The problem is they don’t agree.
“Karl,” “Ka,” and “Carl” are the same client in three systems and three different people to your software. A balance in one tool is a day stale in another. A field means one thing to the CRM and something subtly different to the planning tool. None of this is anybody’s fault — it’s what happens when a firm grows one tool at a time over fifteen years.
Here’s the part most people get wrong: they think the fix is connecting the systems. Plug the APIs together and you’re done. Connecting APIs is the easy part. Making the data from all those systems mean one thing — that’s the actual work. That’s orchestration, and a medallion architecture is how you do it in stages instead of all at once.
The three layers, plus the one most people forget
The idea is to move data through levels of refinement, each one cleaner than the last. The shorthand is bronze, silver, gold.
Bronze is raw. Every system lands here exactly as it is — CRM, custodian, planning, documents, all of it — copied in without being changed. You’re not fixing anything yet. You’re just getting everything into one place so it stops living in fifteen separate islands. Bronze is the loading dock. Nothing here is trustworthy yet, and that’s fine. It isn’t supposed to be.
Silver is cleaned and resolved. This is where the work happens. Duplicates get removed. Records get matched — “Sam,” “Sammy,” and “Samuel” become one person, with one identity, across every system. Formats get reconciled. Conflicts get rules: when two systems disagree about a number, which one wins, and why. By the end of silver, the data is consistent. It agrees with itself.
Gold is business-ready. This is the single source of truth — the version of the data that’s been cleaned, resolved, and shaped into the things the firm actually reasons about: a household, a portfolio, a financial plan, a relationship. Gold is what you’d want a sharp new analyst to read on day one so they don’t make a beginner’s mistake. It’s the truth, in business terms, ready to be used.
Bronze to gold is a path well-worn enough that the industry has a name for the end state — an enterprise-grade “data lakehouse.” For most mid-market firms that’s been out of reach, priced for institutions with a data team. Getting a firm from bronze to gold affordably is a big part of what earns us the right to do anything else.
Then there’s the layer the textbook diagrams leave off, and it’s the one that matters most for where this is going.
The Digital Workforce runs on gold. Once the firm has a single source of truth, the AI built on top of it — the Chief of Staff, the AI Co-Workers — acts on truth instead of guesswork. This is the whole point. The medallion layers aren’t a data-engineering vanity project. They’re the difference between an AI Co-Worker that reasons over the firm’s real, reconciled picture and one that confidently quotes a stale balance from a system nobody trusts.
Why the order is non-negotiable
This is the part I won’t move on, so I’ll be blunt about it.
You cannot skip to the AI. Every firm wants to. The AI is the exciting part, the data foundation is the boring part, and there’s enormous pressure — internal and from vendors — to put something impressive in front of people fast. I understand the pull. I’m telling you it’s a trap.
If you put AI on bronze-level data — raw, unresolved, three versions of every client — you don’t get a smaller version of the benefit. You get a confident engine producing wrong answers at scale, and you get them fast enough that nobody catches them until they’ve compounded. The firms that get burned by AI aren’t usually the ones who picked the wrong model. They’re the ones who skipped the foundation and let a capable system reason over data that didn’t mean one thing yet.
Data foundation first. AI second. Not because it’s tidy, but because the reverse actively makes the firm worse — faster mistakes are still mistakes, and now they’re at machine speed.
How we think about it
I’ll be careful here, because this is a 101 and not a sales pitch. The medallion architecture isn’t ours. It’s a well-established pattern, and any competent data team can describe it. What I can tell you is how we approach it, because the approach is where firms actually get stuck.
We treat the foundation as something to get right quickly and affordably, not as a multi-year consulting engagement that bills forever and ends in a binder. The reason matters: data warehousing has been done for thirty years. It is not the moat, and we don’t pretend it is. We build it faster and cheaper than the traditional consulting market so that a mid-market firm can actually afford the foundation — because until the foundation exists, nothing built on top of it is real. Getting bronze-to-gold done earns the right to do the interesting work. It doesn’t replace it.
The interesting work — codifying how your firm reasons, on top of gold — is the next post. But it only counts if the floor underneath it is solid. That’s this post. Get the data right, and everything after it gets easier. Skip it, and everything after it is built on sand, no matter how good the AI is.
The one line to remember
Connecting your systems is easy. Making them mean one thing is the work — and it’s the work that has to come first. Everyone wants to start with the AI. The firms that get it right start with the data the AI runs on.
Next in the series: what actually goes on top of gold — how a firm’s own reasoning gets codified into something an AI Co-Worker can use, and why that’s the part competitors can’t shortcut.