Where a connector pays for itself: eight questions worth asking your systems
A connector lets Claude, ChatGPT or Copilot answer from your own systems. That is easy to say and hard to picture. So here are eight ordinary questions, by type of business, and a simple test for whether a connector is worth building for you.
The test
A connector earns its keep when three things are true.
- The answer sits in more than one system. The accounts know who is late. The help desk knows who has a reason to be. Nobody builds the report that joins them.
- The question changes. A dashboard answers the question you thought of last year. People ask a different one every week.
- There is something to do next. A chaser, a purchase order, a refund, a booking. The connector prepares it and waits for your yes.
If you only need the same five numbers every morning, a scheduled report is cheaper, and we will say so.
If you send invoices
"Who owes us more than £2,000 past 45 days, and do any of them have an open complaint?"
The customer with a complaint should not get the same chaser as everyone else. A connector reads the accounts and the help desk together, then drafts one email for the clean ones and a different one for the rest. This is usually the easiest case to make, because it is tied to cash.
If you hold stock
"What will we run out of before the next delivery arrives?"
Stock is in one system, sales in another and the delivery date in a supplier's email. The sum is simple. Finding the three numbers is the work. The connector does the sum in code and has the purchase order ready to approve.
If you send people to jobs
"Which of tomorrow's jobs are not ready, and what have we finished but not invoiced?"
A job with no parts allocated. An outdoor job with rain forecast. Five finished jobs nobody has billed. Uninvoiced work is found money, which makes this one easy to put a price on.
If customers pay you monthly
"Which payments failed this week, and who is worth a phone call?"
A failed card is a cancellation nobody chose. The connector lists them, works out what each customer is worth and separates the ones who deserve a call from the ones who need an email.
If customers email you
"Which emails to our accounts and support inboxes have had no reply in two days?"
The reason an invoice is unpaid is often sitting in a shared inbox, in the customer's own words. Read alongside the accounts, an unanswered email stops being an oversight and becomes the explanation.
If you sign contracts
"Which contracts renew in the next 90 days, and which allow a price increase?"
The answer is in documents nobody has opened since they were signed. A connector reads them where they already live, in SharePoint or OneDrive, and only the ones that person may already open.
If you are inspected
"Which certificates, checks or training expire in the next 30 days, and who is not booked?"
For property, care and construction, a missed date is a fine or a lost contract. The dates are recorded. They are just not read together.
If you sell software
"Can our customers ask their own AI assistant about their data in our product?"
They will expect to, soon. We built exactly this for our own product, TellMyCRM, and it is live in Claude, ChatGPT and Microsoft 365 Copilot. For a software company a connector is not a saving. It is a feature customers stay for.
A real one
This is our own connector answering one question about a deal, recorded in Claude on demo data. The CRM said the deal was 80 per cent likely and closing that day. The emails showed the customer had not replied for 94 days. The accounts showed an overdue invoice. Three systems, one answer.
Recorded in Claude on demo data. Best with the sound on.
When it is the wrong answer
- Nobody uses an AI assistant yet. A connector adds nothing to an assistant nobody opens. Start with the work.
- The records are wrong. A connector reads what is there. If the stock count is fiction, the answer will be too. Fix the process first.
- One fixed report would do. Then build the report.
The questions above are examples. The figures in a real answer come from your own systems and are worked out in code, not estimated by the AI.
Have a question like these? Write down the ones your team asks most often and answers slowest. We will tell you honestly whether a connector is worth building.
Book a callNext: What is an MCP connector, and why would a business want one?