An Idiot’s Guide to Running Your Business with AI

You have a small business. You have heard it is now possible to “run it with AI”. You have also heard the words agents, automations, plugins, and MCP, and perhaps you have reached the completely reasonable conclusion that the people saying these words are trying to make you feel stupid.

They probably are not. But people often explain this stuff badly.

Here is the useful version: you can give ChatGPT or Claude the ability to do specific things in your business, instead of only advising you about them. That does not mean firing your team or letting a robot loose in your bank account. It means taking repetitive admin off people’s plates.

The problem is not that there are no apps

For years, software promised to make running a business easier. At first, it did. Then you collected an invoicing app, expense app, social-media scheduler, email-marketing platform, customer database, booking system, payroll system, and about seventeen other logins.

Then using the apps became a job in itself.

Take expenses. Yes, you may use Xero, QuickBooks, or whatever your bookkeeper prefers. But someone still has to find the receipts, upload them, match them to transactions, and put them in the right category. There are good reasons accountants know how to do this: finance software is not exactly made for people who would rather be doing literally anything else.

Or marketing. “Just post on Instagram” is a sentence said by people who have never had to consistently post on Instagram. You need photographs, editing, captions, timing, replies, perhaps a content calendar, and enough enthusiasm to do it again next week. This is why businesses hire social-media managers.

The apps are not bad. But you are still the person manually passing information between them. You are the integration.

What AI changes

Anything boring and repetitive that can be defined as a process can, in principle, be done by software. The new bit is that AI makes it much easier to create that software, and it can understand a reasonably normal instruction instead of requiring you to click through twelve screens.

“Show me all unpaid invoices more than seven days overdue. Draft polite follow-ups, but do not send them yet.”

Or: “We have a cancellation in Thursday’s beginner class. Tell me which people on the waitlist are suitable and draft a message to the first five.”

ChatGPT can already write a nice message. That is the easy part. The interesting part is allowing it to look at your actual invoices, schedule, customers, and waitlist — and then to take a carefully controlled action.

An AI needs the right access and tools

There is a lot of chatter about AI agents replacing people. In practice, an “agent” is an AI with tools and permission to use them. ChatGPT or Claude already provides the conversational part: it is the person at the desk waiting for instructions.

What it lacks is access to your business.

Giving it access does not mean handing over every password and hoping for the best. Give it a narrow set of permitted actions: read sales numbers, look up a customer, create a draft invoice, or propose a roster change. It can publish a class schedule only after you approve it. Each permitted action is a tool.

You may also hear these called plugins, integrations, or MCP servers. The terminology is a bit of a mess, but the practical idea is the same.

What is an MCP server, in normal English?

MCP stands for Model Context Protocol. Ignore the name. An MCP server is basically a small back-end service that tells an AI what it can safely ask your business systems to do.

Imagine a very literal, slightly overprotective operations assistant sitting between ChatGPT and your business. ChatGPT says, “Dana wants the class schedule for next week.” The MCP server says, “Fine, I have a tool called get_schedule. I will return the schedule.”

Or ChatGPT says, “Dana wants to publish a revised schedule.” The server can say, “I have a tool called preview_schedule_update. I will show the proposed change first.” It might require a second explicit instruction before it does the actual publishing.

That separation is important. ChatGPT is the interface you talk to. Your back end is where the real business rules live. The MCP layer is the controlled doorway between them.

Build around the work your business actually does

ChatGPT plugins that connect to Gmail, Sheets, and Calendar are useful. They connect AI to generic apps, but they do not automatically understand how your business works.

Eventually, you want a simple back end that represents the work your business does. It does not need a glossy customer-facing app. It may be a database, a small internal dashboard, a few connected services, and clear rules.

For a dance studio, the back end might know about:

• classes, teachers, venues, and the published timetable;
• students, memberships, bookings, and waitlists;
• teacher availability and substitute-teacher contacts;
• events, ticket sales, and marketing campaigns.

Once those things are represented properly, you can offer plainly named tools such as find_available_teacher, list_waitlist, create_event_draft, and publish_timetable. Plain is good: it means you have described a real process rather than hoping an AI invents one.

For a coffee roaster, the same idea might cover inventory, wholesale orders, customer support, fulfilment, and the barista roster. The details differ, but the method is the same: define the core information and repeatable processes, then give the AI tools to work with them.

A good first version is much smaller than you think

Do not start with “build me an AI business manager”. That is like asking someone to build you a “business app”. It sounds grand and means nothing.

Start with one weekly annoyance: a task with a clear input, clear output, and a way to check that it was done correctly.

1. Write down the process as it happens now. Include the annoying exceptions.
2. Decide what the AI should be allowed to read.
3. Decide what it should be able to draft, but not send or publish.
4. Only then decide which actions it may take automatically.
5. Test it with real but low-stakes examples until it is reliably boring.

For example, a good first tool might be: “Every Monday, collect the inquiries from the previous week, group them by topic, draft replies, and put them in a review queue.” You still approve them. But you have removed the tedious reading, sorting, and first-draft work.

This is not about pretending AI is infallible

AI can misunderstand things and be confidently wrong. It can also turn a simple question into six paragraphs of nonsense. Do not give it a giant “do whatever you think is best” button.

Good business AI has guardrails. It can read more than it can write, and draft more than it can send. It should ask before spending money, changing a schedule, deleting a record, or contacting a customer. It should leave an audit trail. Use the same rules you would want for a new human assistant.

Your bookkeeper, accountant, social-media person, and operations manager may still be essential. AI should remove repetitive coordination and admin, so specialists can spend their time on work that needs their judgement.

What I do with this myself

I run Disco Media, which includes a few content websites and various app projects. I have been building a back end that understands the things I actually do: manage sites, write and update articles, organise media, build and test apps, and keep track of the work around all of that.

This is a natural extension of the slightly grand claims I made when going all in on AI, and the idea of building your own lightsaber: do not just use AI as a chatty search box. Give it the ability to make useful things, carefully.

I talk to the back end through ChatGPT. In fact, I told it to write this article. It did not simply press a magic “write article” button and wander off for a long lunch. It had access to the relevant site, read comparable posts to match the voice, created the artwork, handled the media, and published the result through defined tools.

That is the part I find genuinely exciting: a business owner can finally use an interface that keeps up with what they mean.

The short version

You do not need to become an “AI agent” person or learn every acronym. Identify the repetitive processes that make your business feel unnecessarily complicated, put the relevant information and rules in a sensible back end, and give ChatGPT or Claude a few safe ways to interact with it.

Then you can spend less time moving information from one stupid app to another, and more time on work that requires you.

If you run a business and want help working out what this could look like for you, contact me here. I like turning a vague “there must be a better way” feeling into a practical system.