What is an MCP server, in plain English (and when a small company needs one)

An MCP server is a translator that sits between an AI model and one of your actual business systems, so the AI can read real data and take real actions instead of guessing from whatever you happened to paste into the chat box.

That's it. That's the whole concept. Everything else in this post is just working out when you personally need one.

The short answer, if you're in a hurry

You need an MCP server when you keep copy-pasting the same information into ChatGPT or Claude every day: order numbers, customer emails, spreadsheet rows, calendar availability. If you're doing that more than a couple times a week, an MCP connection turns "paste this in, then paste the answer back out" into "the AI just looks."

You don't need one if you're mostly asking an AI to write things, draft things, or brainstorm things that don't depend on your live business data. Plain chat, no wiring required.

Okay but actually, what is it

Think of an electrician running a new circuit to your kitchen. The power company (the AI model) already generates plenty of electricity. Your house (your business systems: QuickBooks, your CRM, your inbox, a spreadsheet) already has outlets everywhere. The problem is the wire between them doesn't exist yet, or it exists but only the electrician knows how to hook it up safely.

MCP is the wiring standard. An MCP server is the specific wire run to one specific outlet: "here's how to safely ask my invoicing system for last month's totals," or "here's how to safely create a calendar event on my behalf." Once that wire exists, any AI tool that speaks MCP can use it, the same way any lamp can plug into any outlet built to code.

Before MCP existed, every AI tool that wanted to talk to your invoicing system needed its own custom, one-off integration, built by whoever made that AI tool, if they bothered at all. MCP means one integration works with any MCP-compatible AI assistant, not just one vendor's app.

Why does the AI need a "translator" at all? Isn't it already smart?

The AI model itself is smart at language and reasoning. It is not, on its own, connected to anything. It doesn't have your login. It doesn't know your database schema. It has no idea what "job #4471" refers to unless something tells it.

An MCP server is that "something." It's a small program that knows two things: how to talk to your actual system (your database, your accounting software, your file storage) and how to describe that system to an AI model in a shape the model understands. The model sends a request like "get me all invoices from June that are unpaid." The MCP server translates that into the actual database query, runs it, and hands back the real answer.

Without it, you're stuck being the translator yourself, by hand, every single time.

When does a small company actually need one?

You probably need an MCP connection if any of these describe your week:

• You or an employee spends real time each day copying data out of one system (a spreadsheet, a CRM, a ticketing tool) and pasting it into an AI chat to get a summary, a draft, or an analysis.
• You want an AI assistant to actually do something, like send a follow-up email, update a record, or create a calendar hold, not just describe what you should do.
• You have more than one system that needs to talk to the AI (say, your inbox and your invoicing tool), and you're tired of switching contexts between them.
• Your team keeps asking the same "what's the status of X" question and someone has to go look it up manually every time.

When do you not need one?

• You're using AI mainly to write: emails, social posts, product descriptions, job listings. No live data involved, no wiring needed.
• You're a team of one or two and the "system" you'd connect is just... your own memory of what's going on. Sometimes a spreadsheet plus a good prompt is genuinely enough.
• The system you want to connect changes its structure constantly and nobody maintains documentation for it. Wiring something unstable creates more debugging than it saves.
• You haven't actually tried the manual, copy-paste version yet. If you haven't felt the pain, you don't know what to automate. Do it by hand for two weeks first.

What does it cost, roughly?

For a small business, connecting an AI assistant to one or two real systems (say, email plus a scheduling tool, or a CRM plus a shared inbox) is usually a project measured in low thousands of dollars, not tens of thousands, and the ongoing cost is closer to a modest monthly hosting bill than a payroll line item. It scales with how many systems you connect and how unusual those systems are. A standard tool like Gmail or Google Calendar is cheap because the wiring is mostly already built. A homegrown internal tool with no documentation costs more, because someone has to reverse-engineer it first.

What can actually go wrong

MCP servers are still fairly new, which means the tooling around them is uneven. A few honest caveats:

• Not every AI tool supports MCP yet. Check before you build anything.
• Giving an AI assistant the ability to take real actions (send emails, edit records) means you need to think about what happens if it gets something wrong. Start with read-only connections, add write access once you trust the setup.
• "It worked in the demo" and "it works reliably every day for three months" are different bars. Budget for a shakeout period.

Is this worth it for a five-person company?

Usually yes, but scoped tight. The mistake I see is companies trying to wire up everything at once. The move that actually pays off is picking the one manual, repetitive, copy-paste task that eats the most time each week and wiring up just that. Get it solid, then decide if a second connection is worth it.

If this sounds like your Tuesday, I set these up for a living.

Work with me