What AI needs before it can do real work in your business

You probably use AI already. It drafts the email, summarizes the meeting, cleans up the proposal. Then you go back to the scheduling system, the job list, or the maintenance system and do the real work by hand.

The gap isn’t the AI model. Chat tools are already capable. What’s missing are two things: access to the systems where your work happens, and structure that tells the AI how your business actually runs.

Access: AI has to reach your systems

A chat window only knows what you paste into it. To do real work, AI has to reach the systems you run on: the calendar, the job list, the inbox, the ERP, the maintenance system.

One way to do that is an MCP server, built on the open Model Context Protocol: a connection the AI uses, with you signed in, that gives it access to one system along with instructions for using it. A business can run its own, so that each person’s AI sees the data their role allows and nothing more.

• In a plant, a maintenance planner’s AI can read open work orders and parts on hand, and draft the week’s schedule.
• In a local service business, the office manager’s AI can see the calendar and the open jobs, and draft tomorrow’s routes.

Two rules make this safe. Permissions are applied before any data reaches the AI, so it can’t show someone what they shouldn’t see. And we connect read-only by default: the AI drafts, and a person approves anything that changes a record or goes to a customer. I’ve built these connections for a plant, letting AI query maintenance, materials and production data, and we deployed them read-only. It was the right call for that application, and it’s the right default for most.

Structure: AI has to know how you work

Access alone isn’t enough. An AI that can see your job list still doesn’t know how you price a job, who handles warranty calls, or what “urgent” means in your shop. That knowledge lives in people’s heads.

Structure is two things:

• A knowledge base. Plain written pages on how the business runs: your services, your policies, who does what, where the data lives. The AI reads from it when it answers. This is also why you rarely need to “train your own model.” Handing the AI your information at the moment it answers is simpler to set up and to keep current than changing the model itself.
• Skills. Saved instructions for a task, written as a goal with context rather than a rigid script. “Every Monday, find quotes more than a week old with no reply, draft a short follow-up in our usual tone, and put them in my drafts folder” is a skill. Once it’s written, anyone on the team can use it, and it gets better as you refine it. (Anthropic’s documentation describes how skills package instructions and context so they don’t need repeating in every conversation.)

The instructions, documents, and tools around an AI now matter more than which AI model you pick. That part is in your hands.

A first step you can take this week

Pick one repetitive task that eats time every week. Then write down three things:

1. Which systems it touches.
2. How it gets done today, step by step, by the person who does it best.
3. What “done right” looks like.

That page is the start of your knowledge base, and the outline of your first skill. It also tells you exactly what access the AI would need.

If you’d like help finding the tasks worth automating first, that’s what an AI Assessment does. See how it works for manufacturers and for local businesses.