Your AI doesn't know your business

It's 2:30 in the afternoon, and the controller at a 30-store retailer has just been asked a simple question: "How much are we paying in wages at Queen West, and what's the breakdown by employee? I think we're paying too much, and I need to know where we're bleeding."

It isn't simple. She has to count the store's employees, pull months of wage history from QuickBooks, match it against payroll to see who was paid what and when, and then check each person's hours to make sure no one is being overpaid. All of it gets lined up by hand.

The hardest part is that none of these systems agree on what the store is called. The POS calls it Queen West. Payroll calls it Location 014. QuickBooks tracks it as a class. She knows they're all the same place. No software does.

Twenty years of software, and the work is still manual

The average growing business now runs accounting, payroll, scheduling, a POS, e-commerce, banking and a CRM, plus a few tools somebody forgot they're still paying for. Each is very good at its own slice, but none of them knows the business as a whole. That job falls to a person, usually the operator, and it happens in spreadsheets and in their head.

AI doesn't fix this on its own. McKinsey's 2025 survey found that 88% of organizations already use AI in at least one function, so access to a good model is no longer an edge. What matters is what the model knows about your company, and out of the box, that's very little. Point an agent at QuickBooks and it can query QuickBooks, but your business isn't QuickBooks. It's how QuickBooks relates to your POS, your people, your locations and years of decisions nobody wrote down.

Picture hiring the smartest CFO alive, starting Monday. They don't know your stores or your chart of accounts. They don't know that one system's numbers can't be trusted after midnight, or that "Accessories" means something different in two databases. They'd be useless for weeks, and not because they aren't smart.

Models will keep changing. Context is what compounds.

GPT improves, then Claude, then Gemini, and six weeks later someone ships something better. Intelligence is getting cheap and interchangeable.

What isn't interchangeable is the knowledge that lives inside your company: how your locations map across systems, which source wins when two disagree, what counts as a strange transaction, what a good labour number looks like for Store 7 on a Saturday in December. Most of that lives in people's heads.

That's the layer we're building with Sourcebook. It sits between your systems of record and whatever intelligence you plug in on top. It holds the entities, mappings, definitions and rules, so a new model or agent inherits all of it on day one instead of starting from scratch. [One sentence on how it actually works or what you've built, in plain terms.]

From "what happened" to "why"

A dashboard tells you Store 12 revenue fell 11%. That's useful, but it's also where the real work starts, because now someone has to find out why.

With the context in place, the answer can look more like this: revenue fell 11%, almost all of it between 5 and 9 p.m. Transactions were flat, but basket size dropped. Three high-volume products were out of stock for two days, and scheduled labour didn't change, so labour as a share of revenue went up.

The first version gives you a number. The second gives you something to act on. And once a system can explain a problem, it can start handling parts of it: flagging a reorder, suggesting a schedule change, opening a reconciliation, or escalating the unusual cases to a person.

None of this replaces the operator. A competitor opens across the street, your best manager is thinking about leaving, the city tears up the road outside your store, and for some reason everyone wants pistachio this year. No model sees any of that. The operator's job shifts from doing the reconstruction to supervising it, and spending judgment where it's actually needed.

The system of record stays. The operator stays.

QuickBooks should keep being great accounting software. Shopify should keep being great commerce software. They aren't going anywhere. They just become inputs.

Systems record the business. Sourcebook understands it. AI reasons about it. Operators run it.

[Closing line in your own voice. Something about the controller from the opening getting her evening back, if you want to bookend it.]