MCP stands for Model Context Protocol. It is an open standard that lets an AI assistant connect directly to the systems where your data actually lives, rather than being handed screenshots and copied-out spreadsheets. In an accounting context it is the difference between asking a general question about VAT and asking a question about your VAT.
The problem it solves
An AI model on its own knows a great deal about accounting and nothing about your business. To get useful answers you end up exporting data, pasting it in, and checking that what comes back relates to the figures you sent. That works once. It does not work as a routine.
An MCP server sits between the model and a system, such as a ledger, a document store or a practice management tool, and exposes a defined set of actions. The model can then look things up and, where permitted, make changes, without anyone copying data back and forth.
What it looks like in practice
The useful cases in a small business are unglamorous and that is rather the point. Asking which customers are over their terms and drafting the chasing emails. Pulling the figures behind a variance before a board meeting rather than during it. Checking a supplier invoice against what was ordered. Working out what a VAT return would look like on current figures, three weeks before quarter end.
None of that is new work. It is work that currently waits for someone to have an hour free.
Where the limits are
Three things are worth being clear about.
First, access is not judgement. A model that can read your ledger can tell you what is in it. Whether a cost belongs in this period, whether a claim is defensible, whether an arrangement carries a risk: that is the part a qualified person is for, and it is the part that matters when HMRC asks.
Second, connections need boundaries. Read-only is a reasonable default. Anything that writes to a ledger, sends an email or moves money should be something a person approves, not something a model does because it seemed sensible.
Third, this is client data. If you are connecting an AI tool to systems holding personal or financial information, the same duties apply as to any other processor: know where the data goes, know what is retained, and check that your engagement terms and privacy notice actually cover it.
Why we are interested
We build our own reporting on Xero data and Power BI, so the appeal of a model that can query the same source directly is obvious. The honest position today is that the technology is ahead of the governance. We are using it internally, on our own data, before we go near anyone else's.
If you are experimenting with it yourself, the question worth asking is not whether it can reach your accounts. It is who checks what it did.
