Last week we talked about the loop between the creative side and the structured side. This week is about feeding it. Most small businesses have their data in three or four places that do not talk, and an MCP server is the least painful way we have found to pull it together.
What it is, without the acronym
An MCP server is an API for AI.
That is genuinely the whole thing. An API is just data access, a defined way for one piece of software to ask another for information. What the industry has been building over the last couple of years is a version of that access designed specifically for AI systems to use. Same idea, new consumer.
You do not need to understand how it works any more than you need to understand how your card reader talks to your bank. What matters is what it lets you stop doing.
The problem it solves
Picture the typical setup. Customer names and contact details in a CRM. Jobs, scheduling and delivery in an ops tool. Opens, clicks and newsletter numbers in a marketing platform. All three contain something you need. None of them contain enough on their own.
Large companies solve this by buying an ERP where everything lives in one box. Most small businesses cannot justify that, and honestly should not. So instead, managers spend their week reconciling customer names across systems, chasing which deal is which, and exporting everything into spreadsheets to try to line it up by hand.
And the owner still cannot answer a straightforward question, because they are looking at one system at a time and the real answer lives across all three.
People do fix this, in bursts. Somebody sets up a clean export routine and keeps it going for a month or three. Then it gets busy, and it stops. Every time.
What it looks like when it works
With an MCP connection in place, you can sit in one chat window and ask across systems. Pull the customer record, check what is still open in ops, and tell me whether this account is actually profitable.
And it goes and does it. Sally has done four deals with you, one is still in the workflow, it is due next Tuesday. No clicking into the CRM, no second tab for ops, no export.
That is a real improvement and it is available to businesses that could never afford the enterprise version of it.
The asterisk, which is large
It is still 70%.
When it goes out and filters across those systems, it is doing the same probabilistic thing it always does. It will grab a near match. It will quietly return four deals when there were six.
Which brings up the thing that separates people who get value here from people who give up: you have to be willing to tell it no. Go back and get the right data. That is not a failure of the setup, that is operating it correctly. People who never push back are the ones who conclude it does not work.
So you still have to know your own numbers well enough to notice when they are wrong. There is no version of this where you hand over the expertise.
Have it write to markdown, then check it
Get the answer out as a markdown file rather than as chat. Same reasoning as last week: low overhead, readable both ways.
Now you can actually audit it. Sally, four deals, next due Thursday. And you can say: no, six. That correction is only possible because the answer is sitting still in a file instead of scrolling away in a conversation.
Then get it into real structure
This is the part we would push hardest on, and it is the same shape as the spreadsheet episode.
A markdown file is fine for a one off question. If it is a lot of data, or a question you will ask every month, it needs to go into a database.
What we do on systems we have not built ourselves, where it is cheaper to keep the tool that exists than to replace it: on a schedule, pull the data out and land it in a database we control. From there you can point AI at it, or skip AI entirely and run ordinary structured reporting.
For anything a decision rests on, we would take a structured report over a chat answer every time. Not because AI cannot do it, but because we do not want the 70% anywhere near it.
The short version
- An MCP server is an API for AI. It fetches data on the model's behalf. That is all you need to know about it.
- It removes the export and reconcile treadmill, which is the part that always lapses.
- Call the 70% bluff. Tell it no. Send it back for the right data.
- Land the answer in markdown so you can check it against what you know.
- Move anything recurring into a database, and report off that.
That is the series. Five episodes, one argument: AI is a genuinely powerful creative tool sitting at about seventy percent, and it becomes a business system only when you put structure underneath it and keep checking the output. Everything else is detail.
If you are working through this in your own business and want a second opinion on where to start, get in touch. That part is free.
