Subatomic 101 series · Post 3 · Karl Simon, CTO
In the first post I gave you the three-layer vocabulary — LLM, agent, AI Co-Worker. In the second I made the case that none of it matters until your data means one thing, and we walked bronze to gold. This post is about what goes on top of gold: cognition. Whose reasoning the system runs on.
I have a version of this conversation every week. A prospect asks how an AI Co-Worker is different from the agents the big labs are shipping, and it’s a fair question — the market is blurring the two categories as fast as it can, because the market hasn’t named the second one yet. But the blur is not harmless. Treating agents and AI Co-Workers as the same purchase is the most expensive mistake a firm can make in this space right now. A firm that needs the operational weight lifted off its team ends up with a faster keyboard. A firm that needs a faster keyboard ends up paying for a full hire.
So let me draw the line cleanly, in language that holds up after the demo is over.
What an agent actually is
An agent is a generalist. It arrives with a set of pre-built skills — summarize this, retrieve that, draft this email, fill this form. The vendor decided in advance what the agent is good at, and it’s good at those things for everyone the same way. That’s the design philosophy: build something that works passably well for the widest possible audience.
Agents are usually tied to whichever underlying model their vendor sells, and they’re built to make existing workflows faster. You were already searching, summarizing, drafting — the agent does those things at higher speed and lower marginal cost.
None of that is wrong. Agents have a legitimate place, and the big labs and the agent-builder platforms make reasonable ones. If your team needs to type faster, search faster, or move information faster, an agent will help.
But that’s an answer to that problem. Not the next one.
What an AI Co-Worker actually is
An AI Co-Worker is the opposite design philosophy. It doesn’t come out of the box. It gets hired, and then it gets onboarded — onto the firm’s data, the firm’s reasoning, the firm’s SOPs, the firm’s decisions. The unit of design is one firm. Not the mass market.
That changes what’s underneath. An AI Co-Worker is not a wrapper around someone else’s model. The cognitive layer — what we call Subatomic IQ — is the product. The LLM beneath it is interchangeable; we pick the model that works best for a given job and swap it when something better ships. The firm’s intelligence doesn’t live in the LLM. It lives in the cognitive engine sitting on top of it — and that engine only works because it reasons over the gold data layer we built in the last post. Cognition on bronze is guesswork with confidence. Cognition on gold is the firm’s actual judgment, running.
And the work is categorically different. An AI Co-Worker doesn’t speed up a task. It carries a role — meeting prep, client follow-up, compliance review, data operations — end to end, trained on this firm’s way of doing it. The point is not to replace the people. It’s to take the operational weight off them so they spend their time where humans actually create value.
The shorthand I use with buyers: AI Agents replace tools. AI Co-Workers replace the work — not the people. Hire an agent if you want your team to type faster. Hire a Co-Worker if you want your team focused on the work only people can do.
The $50,000 question
Here’s where the distinction stops being philosophy and starts being money.
A wealth-management client calls her advisor and says she’d like to pull $50,000 from her IRA next month.
An agent retrieves the rule. There’s a 10% penalty if she’s under 59½. There’s ordinary income tax. There may be state tax depending on jurisdiction. The agent gives that answer correctly, fluently, and quickly.
An AI Co-Worker trained on this firm’s reasoning answers a different question entirely. It already knows this client’s tax bracket, the unrealized gains in her taxable account, the Roth conversion ladder the advisor put in place two years ago, and the cash position she could draw from instead. It knows how this firm thinks about a $50K withdrawal — which scenarios get flagged before any number gets quoted, which compliance checks happen first, which next-best-action this firm would surface for this household.
Same question. Completely different answer. One is a faster lookup the advisor still has to act on. The other is the firm’s own reasoning, executed before the advisor even picks up the phone — so the conversation with the client is about counsel, not arithmetic.
When the stakes are low, the difference looks like a nice-to-have. In wealth management — and in most knowledge work where a decision actually matters — the stakes are not low.
The cognitive engine is the part nobody can shortcut
This is the part most buyers underestimate, and it’s why I spent the last post on data without apology.
The data warehouse isn’t the moat. Data warehousing has been done for thirty years; we build the foundation faster and cheaper than the traditional consulting market, and that earns us the right to play. It doesn’t earn us the right to keep playing.
The moat is what sits on top. Subatomic IQ codifies the firm’s reasoning, decision-making, and SOPs into a tree of thoughts — the way this firm thinks about this problem, with an interrelated understanding of how the firm’s products, services, and processes connect. Industry best-practice baselines come included. The firm’s secret sauce goes on top of that.
That work is consulting work. Someone has to sit with the firm, extract the reasoning, structure it, encode it, and tune it as the firm evolves. It cannot be shipped in a box, because the whole point is that it isn’t the same for everyone. A vendor can shortcut the model — everyone rents the same handful of LLMs. Nobody can shortcut the part where your firm’s judgment gets encoded, because that part requires your firm.
The five questions to ask any vendor
If you’re evaluating any AI investment this year, bring these to every vendor on the table.
Is this onboarded to my firm, or is it a set of generic skills designed for everyone? If the demo looks the same as the demo your peer firm saw last week, you’re looking at an agent, not a Co-Worker.
Is the intelligence in the LLM, or above it? If the answer is “we use [X model]” and the conversation stops there, the cognitive layer isn’t theirs — and it isn’t yours either.
Does it speed up a task or carry a role? If the success metric is “faster at X,” you’re buying a tool. If the success metric is “X gets done end-to-end so the team can focus elsewhere,” you’re hiring a Co-Worker.
Whose reasoning does it encode? If the answer is “industry best practices,” it’s a starting point. If the answer is “yours,” it’s a finished product.
What happens when we part ways? Co-Workers leave the way people leave — the data, outputs, and decisions stay with the firm. If the answer is anything else, the vendor is selling a subscription, not a hire.
Five questions. Any vendor worth hiring can answer all five without flinching.
Where this goes next
You now have the vocabulary, the foundation, and the cognition layer. What’s left is how it all fits together — because a firm doesn’t run on one Co-Worker any more than it runs on one employee.
Next in the series: the architecture — how the pieces assemble into a stack that actually runs a firm, and why the answer to “too many tools” was never another tool.