← All posts

What is an MCP connector, and why would a business want one?

· Bob Ottley

Claude, ChatGPT and Copilot are clever, but on their own they know nothing about your business. A connector changes that, and done properly it is one of the highest-value pieces of software a small business can own.

190hours a year, example: a manager pulling ten reports a day
5rules every connector we build follows
0changes made without a person saying yes

The short version

MCP stands for Model Context Protocol. It is an open standard, now supported by the major AI assistants, that lets an assistant talk to another system in a controlled way. Think of it as a door you build into your own systems, with a lock you control and a list of exactly what may pass through it.

Without a connector, you copy and paste: export a report, paste it into ChatGPT, ask a question, and hope the export was current. With a connector, you ask the question and the assistant fetches what it needs itself, from the live system, every time.

What a connector is, in one minute. Best with the sound on.

What it saves

Take a sales or operations manager who needs ten answers a day from their systems: which orders are late, which customers have gone quiet, what is due this week. If each answer takes five minutes of opening a system, filtering and exporting, that is 50 minutes a day. Over 230 working days it is around 190 hours a year, nearly five working weeks, spent fetching rather than deciding.

Example figures. Your numbers will differ; the method is the point.

With a connector, each of those becomes a sentence: "Which customers have not ordered in sixty days, and what did they usually buy?" The answer comes back in seconds, from the real figures.

What it makes possible

  • Questions across systems. "Which of our overdue invoices belong to customers with an open complaint?" needs the accounts and the help desk together. Nobody builds that report, so nobody asks.
  • Drafts with context. "Draft a reply to this complaint" with the order history, previous emails and delivery tracking already in front of it.
  • Changes, safely. "Move the Acme deal to proposal and set the close date to the end of the month" done in the CRM, after it has been read back and confirmed.

The five rules we build to

We built a connector for our own product, TellMyCRM, so that sales teams can ask Claude, ChatGPT or Copilot about their pipeline and update it. The rules we set for it are the rules for every connector we build:

  • It signs in as the person asking. It sees what that person is allowed to see in the underlying system, and nothing more. A sales person sees their deals; a director sees the team's.
  • Reading is not changing. Reading tools and changing tools are separate. Anything that changes a record is read back in plain words first and only written when the person says yes.
  • The figures come from the system. Totals, counts and sums are worked out in code from the real records, never estimated by the AI.
  • It exposes jobs, not the database. We give the assistant specific, named tasks such as "find late orders" or "update a deal stage", each with limits, rather than open access to everything.
  • Everything is logged. What was asked, what was read, what was changed and by whom.

What to connect first

Start with the questions people already ask most often and answer slowest. Write down a week of them. If more than half need the same two or three systems, those are your first connector. If your team does not use an AI assistant at all yet, start with the work, not the tool: a connector adds nothing to an assistant nobody opens.

Have a job like this? Tell us in a few lines what gets copied, checked or chased by hand. We will come back with questions, not slides.

Book a call

Next: Where a connector pays for itself: eight questions worth asking your systems