Your business doesn't have an AI problem. It has a meaning problem.

Point an AI at your business and ask it a simple question. Which members are at risk this month. How did Tuesday mornings do last quarter. You'll get an answer. It will sound confident. And some of the time it will be quietly, completely wrong.

The usual response is to blame the model. Wait for a smarter one, write a longer prompt, add another tool. But the model was never the problem.

The meaning problem

A studio of any size runs on five or six systems. Bookings in one, payments in another, accounting in a third, marketing in a fourth. Every one of them has an API. Not one of them composes with the others.

The same person exists in all of them: a member in the booking system, a customer in payments, a contact in marketing. Three records. Three ID schemes. No agreement anywhere about what a customer actually is.

Hand that to an AI agent and it has to guess how the pieces join. It guesses mid-answer, silently, with no idea that it's guessing. The answer arrives fluent and plausible, built on a join that never existed.

That's not an AI problem. It's a meaning problem. Nobody ever told the machine what anything means.

Three moves, done once

The fix is unglamorous, which is probably why it's rare. It's three moves, and the point is that you make them once.

Build the ingestor. Every source comes in, API where one exists and CSV where one doesn't, and lands in a single canonical stream. A booking becomes a booking regardless of which platform it came from. The differences between systems get resolved at the door, not in every conversation afterwards.

Build the ontology map. Every entity, relationship and event gets defined exactly once, in machine-readable form that travels with the data. What an active member is. What at-risk means. How attendance and payment state relate. When an agent asks "who's at risk this week," it isn't pattern-matching column names. It's reasoning against definitions. Two agents asking the same question get the same answer, because the meaning is in the data, not in the prompt.

Add the skills. A clean model is necessary but it isn't the point. The point is the frameworks that run on top of it: the analysis nobody had time to do, the work that used to be a consulting engagement. An instructor bonus model that measures capacity against the true median class size for the time slot instead of how many bodies fit in the room. A first-thirty-days read that tells the team which new members need a person, not another automated message.

A brain that knows its own edges

Do those three things and the business has something new: a brain it can talk to. Ask it anything and one of two things happens. It hands back real data, with the reasoning visible. Or it says plainly that it can't answer, and why. The data isn't connected, the definition doesn't exist, the question is outside what it holds.

I'd argue the second behaviour matters more than the first. Most AI in business today will answer anything you put to it. That isn't a feature. A system that knows the edge of its own knowledge is one you can act on. A system that guesses is a liability with a good interface.

The shared part

Here's the piece we didn't fully appreciate until we watched it in use.

Because Kula Intelligence is built as an MCP server, the open standard Claude, ChatGPT and Gemini use to connect to real systems, it isn't one person's clever AI setup that nobody else can reproduce. It's shared infrastructure. The owner, the manager and the person on the front desk all ask the same system, against the same definitions. Each sees exactly the slice their role, their relationships and the member's consent allow. Every query passes through the same policy engine and lands in the same audit log with its declared purpose.

Same language, different windows onto it. Nobody walks into a meeting with a different number for the same word.

Getting a business to agree on what its own words mean turns out to be worth more than any single answer the system gives back.

Where this is live

The technical detail, including what sits behind the endpoint, the semantic catalogue, the five-layer policy engine, and what's connected today versus on the roadmap, is documented in full on the Kula Intelligence MCP page. Connect it to the AI plan you already have and ask your own numbers a real question.

And while fitness and wellness is where we built this, nothing about the method is fitness-specific. Ingest, define, apply. Any business whose data lives in systems that don't talk is the same shape of problem, and the same three moves solve it.