Connect what you have before you build anything new

A fair share of the enquiries we receive describe a piece of software that, on inspection, does not need to exist. The business already has a CRM, a ticketing system, a finance tool and a shared drive full of documents. What they want is for information to move between those systems, and for someone to be able to ask a question across all of them without opening four tabs.

Until recently that meant an integration project. Now it often means connecting what you already have to an AI assistant, and we would rather tell you that in the first conversation than build you something.

MCP in plain terms

MCP stands for Model Context Protocol. It is an open standard, published by Anthropic in late 2024 and since adopted by the other major AI companies, that gives an AI model a standard way to connect to a system.

Each system gets a small connector, which the standard calls a server. The connector describes what the system can do in terms the AI understands: read these records, search these documents, create this kind of item, update that field. The assistant can then do those things, with whatever permissions the connector allows.

It works like a plug socket. Once a system has one, any assistant that speaks the standard can plug in. Claude, Anthropic's assistant, speaks it. Many of the tools businesses already use have connectors, and where one does not exist, building one is a small job.

What it looks like on a Monday morning

Someone in the finance team asks Claude which of last week's support tickets came from customers whose invoices are overdue. Claude reads the ticketing system and the finance system through their connectors and answers, with the ticket numbers and the amounts.

They ask it to draft a reply to each customer and log a note against each account in the CRM. With permission, it does. Nobody built anything. There is nothing to maintain beyond the connections themselves, and the same setup answers next week's question, which will be a different one.

When this is the whole answer

The information already lives in systems the business runs. The work is looking things up, cross-referencing, drafting, summarising, and moving records between places. The volume is human-scale, with a person asking and a person checking. When all of that is true, connecting the systems to an assistant does the job, and it costs a subscription and a day or two of setup rather than a build.

We check for this in the first week of every project, because we would rather lose a build than deliver one you did not need.

When you still need to build

Some things are a system, and they need building. Work that runs unattended thousands of times a day. Anything your customers interact with directly. Logic that is the product itself. Processes that have to be audited, or that a regulator will ask about. A data model that does not exist anywhere yet.

In those cases the AI sits inside a platform that we build properly, with the tests, the access controls and the monitoring that come with production software. MCP still matters there, in the other direction: the systems we build expose their own connectors, so that assistants can plug into them from day one.

Where to start

List the systems the business runs. Check which have connectors. Set permissions carefully, starting with read-only access for a small group, because whatever the assistant can see, the person using it can ask about. Pick one question that someone asks every week and see whether the assistant can answer it.

If that works, keep going. If it does not, you have learned precisely what needs building, and it is usually much smaller than the original request.