Your back office cannot answer questions

Someone in your company is already copying numbers out of your back office into ChatGPT. You may not know about it. They open the overdue invoice list, select, copy, paste it into a chat window, and ask for a summary or a draft chase email.

It works. It is also the most expensive thing they do all week, and they do it again every Monday.

What you will be offered instead#

When this surfaces, the proposal is nearly always the same. Replace the software. Something newer, with AI built in.

That is the most expensive possible answer to a problem that sits elsewhere. Your back office holds your customers, your prices, your history, your billing habits. Replacing it means another migration, retraining everyone, and spending 6 months rediscovering the small rules nobody ever wrote down.

The software is not the problem. The problem is that it cannot answer when you talk to it.

A standard socket#

Since November 2024 there has been a standard for this. It is called MCP. In practice it is a socket: one end plugs into your back office, and ChatGPT or Claude connect to the other.

This is not a bet on a fashion. On 9 December 2025 the standard was handed to a foundation under the Linux Foundation. Anthropic, OpenAI, Google, Microsoft and AWS are founding members, and Anthropic counted more than 10,000 such servers running at that point.

So it is no longer one vendor's idea. The question "what if this disappears in two years" has an answer.

What changes for you: your sales person stops retyping and starts asking. "Which customers have ordered nothing in 90 days?" The answer comes from your own database rather than from an invention.

A Monday morning#

The real value shows up when nobody asks anything at all.

A scheduled task runs at 8am on Monday. It queries your back office, pulls every invoice more than 30 days overdue, sorts by amount, and sends you the list with a draft chase for each one. You open your phone and it is there. Nobody prepared it.

Two limits worth knowing before you believe in it. The minimum interval is one hour, on ChatGPT and on Claude alike, so this is not real time and does not need to be for collections. And the output stays a draft. You still decide who gets chased and in what tone.

The questions your accountant will ask#

These are the right questions, and they have precise answers.

What data leaves the company? Only what the socket exposes, and we decide what that is. We do not open "the software". We open a list of permitted requests: overdue invoices, yes; payroll records, no.

Read-only or read-write? Read-only by default. That is not decorative caution. Claude's own documentation warns that a scheduled task can use every tool from a connector you gave it, writes included, without asking again mid-run. A socket that can write into your accounts therefore has to be opened deliberately, never by default.

What if the socket lies? The standard lets a server declare itself read-only. Official documentation has said since March 2026 that this declaration is a hint rather than a guarantee: a server you do not control can announce one thing and do another. That is why we install your socket instead of plugging in the first one we find online.

Has this ever gone wrong? Yes. In June 2025 Asana took its own socket offline after finding a tenant isolation defect. Data could be visible to other organisations for over a month, with roughly 1,000 customers potentially affected.

Not an intrusion: an internal defect. We cite it because that is precisely the risk to cover, and because a provider who swears it never happens has not read about it.

One technical constraint while we are here: the server has to be reachable from the public internet. If your back office never leaves your network, that conversation happens before the work, not after.

What it costs, and what it does not replace#

No new licence. No migration. No training either, because your staff already know how to type into a chat window, and that is the only new skill.

What it does not do: it will not tidy your data. If your invoice statuses mean three different things depending on who typed them, the AI will answer with the same confusion, faster.

A socket on a well-kept system saves hours. On a badly kept one it makes the mess more visible. Useful, but not what you were sold.

Here, this kind of work gets a price and a date before the first line of code. What we open, what we refuse to open, and who may use it, all of that is written into the quote.

If you want to know where to start, it fits in one question. What does someone in your company retype by hand, every week, and has done for long enough to stop noticing?