You’ve probably had this moment. You paste a question into Claude or ChatGPT, something specific to your business and the answer comes back... fine. Plausible. Polished, even. But useless. Because the AI doesn’t know your numbers, your customers, your context. It’s guessing in a well-educated way.
This is the default mode. AI as a brain in a jar. Incredibly smart. Completely disconnected from everything that matters to you.
Most people stop here. They conclude that AI is good for writing emails and summarising articles but not for real work. They’re wrong, but they’re wrong for an interesting reason.
The problem isn’t the brain. The problem is that the brain has no eyes and no hands.
What eyes and hands look like
Imagine you could wire your AI directly into your business data. Not copy-paste a spreadsheet into a chat window. Actually connect it, so when you ask “how’s our cash position?”, the AI pulls your real bank balance, your outstanding invoices, your upcoming bills, and gives you an answer grounded in your numbers from this morning.
That’s what a custom connector does. It gives AI eyes to see your data and hands to work with your tools.
The technology behind this is called MCP or Model Context Protocol. You don’t need to remember that. What you need to know is that it’s become surprisingly easy to build a bridge between your AI and your business systems. And when you do, something fundamental shifts.
The AI stops being a clever stranger. It becomes a working partner who’s read every invoice, every contract, every transaction you’ve ever processed.
Three layers to get there
Building this bridge isn’t one step. It’s three, and each one compounds the one before it.
Layer 1: The data layer
Your business already generates data. Accounting software, CRM, project tools, they’re all quietly recording everything. The first job is to pull that data into a clean, queryable form. Not a data warehouse project that takes six months and costs a fortune. Something simpler: connect to your existing systems, extract what’s there, clean it up so it makes sense.
This is plumbing. Unsexy but essential. If the data is wrong, everything built on top of it is wrong. Get this right and you have a foundation. Get it wrong and you have an expensive hallucination machine.
Layer 2: The analysis layer
Raw data is overwhelming. Thousands of transactions, hundreds of contacts, years of history. Nobody wants to stare at that. The second job is to build views pre-computed answers to the questions that actually matter.
What’s our revenue trend? Which customers are growing? Where are we losing money? How does this quarter compare to last year?
These views turn a database into a dashboard. They encode business logic: this is how we calculate margin, this is how we define churn, these intercompany transactions should be excluded. The views carry institutional knowledge that would otherwise live in one person’s head.
Layer 3: The connector
Now the interesting part. You take those views and expose them through a connector that AI can talk to. The AI can browse your dashboard pages. It can run queries against your metrics. It can search for specific topics across your reports.
Here’s where it gets powerful: the AI doesn’t just retrieve data. It reasons over it. You ask “why did revenue drop in November?” and it pulls the monthly numbers, checks which clients churned, looks at whether project revenue fell while retainers held, and gives you a narrative that a human analyst would take half a day to produce.
Not a generic answer. Your answer, from your data, in seconds.
What changes
Before the connector, your conversations with AI are theoretical. “What should I think about regarding cash flow management?” You get textbook advice. After the connector, the conversations are operational. “Our AR is R370K and climbing. Which clients should we chase first and what should we say to each one?”
The shift is from AI as advisor to AI as analyst. An advisor gives you frameworks. An analyst gives you answers.
A few things I’ve watched happen once this bridge is in place:
Leadership meeting prep drops from days to minutes. The AI has already read the dashboard. It knows the numbers. You ask it to draft talking points for the revenue discussion and it writes them with the actual figures, the actual trends, the actual comparisons. No more assembling slides from six different spreadsheets.
Questions get cheaper. When every question costs 30 minutes of an analyst’s time, you ask fewer questions. When every question costs 30 seconds and a prompt, you ask better ones. “What would happen to our margin if we lost our three biggest clients?” becomes something you can explore at 9pm on a Tuesday, not something you commission and wait a week for.
Context compounds. The more the AI sees, the more useful it gets. After a quarter of data, it knows your seasonal patterns. After a year, it can spot anomalies you’d miss. It remembers that November is always slow, that Client X pays late every time, that your fastest-growing service line has the thinnest margins.
The surprisingly low bar
The mental model most people have is that this kind of integration requires a massive IT project. Enterprise middleware. Months of consulting. Six-figure budgets.
It doesn’t. The data layer is your existing accounting software. You’re already paying for it. The analysis layer is SQL views that a competent developer can build in days, not months. The connector itself is a few hundred lines of code that slots into a web server you’re already running.
The entire bridge from “AI can’t see my business” to “AI can query every metric I track” can be built in days or weeks. The hard part is deciding which questions matter. That’s a leadership problem, not an engineering problem.
The real unlock
We spend a lot of time debating which AI model is best, which prompting technique is optimal. None of that matters as much as whether the AI can see your data.
A mediocre model with access to your real numbers beats a frontier model guessing from generic training data. The unlock isn’t better AI. It’s connected AI. Give it eyes and hands, and the brain you already have access to (for free or nearly free) becomes something genuinely useful.
Your data is the moat. The connector is the drawbridge.
— Graham
"The views carry institutional knowledge that would otherwise live in one person's head" — this is the core vulnerability in every knowledge-intensive function, and your three-layer approach addresses it elegantly for business data.
The same pattern plays out in software engineering. How we calculate margin becomes how we structure error handling. Churn definitions become architectural standards. Intercompany exclusion rules become security boundaries. The knowledge exists, but it lives in senior engineers' heads and scattered Slack threads — not embedded in the tools making decisions.
Your distinction between AI as advisor vs analyst applies here too. An AI coding tool operating as advisor gives generic best practices. One operating as analyst — connected to the organization's actual conventions, post-mortems, and architectural decisions — produces output that reflects how the team actually builds software.
We've been building that connector layer for engineering teams: https://encephalon.net/whitepaper. Different domain, same three-layer architecture.