The short version: 40Two built a small read only server that lets an AI assistant query our QuickBooks Online file directly, every morning, without being able to change a single transaction. In its first couple of weeks it surfaced about a dozen hours of unbilled contractor time and a client we had been hosting free since 2023. Below is what we built, what it caught, and how to do the same thing safely.
Why do small businesses find out about billing problems so late?
Because nobody looks at the books daily. Most owners see their real financial position once a month, usually about a week after it would have been useful. The bookkeeper catches up, a report gets emailed around, and by then the invoice that never got sent has been sitting there for sixty days.
Our books at 40Two were fine. QuickBooks was doing its job. The problem was the lag, and the things that quietly slipped through it:
- Contractor hours that were worked, logged in a spreadsheet, and never turned into an invoice.
- A client we had been hosting since 2023 who had never been put on a paid hosting plan.
- Invoices sent, never opened by the customer, and never followed up on.
None of those are accounting errors. They are attention errors. The fix for an attention error is not a better report. It is something that looks every day without being asked.
What is an MCP server, and what did we build?
MCP stands for the Model Context Protocol. It is an open standard that lets an AI assistant connect to a real system and read real data instead of guessing. Think of it as a plug that a tool like Claude can reach into.
We wrote our own MCP server for QuickBooks Online and we run it ourselves, on our own infrastructure. Four decisions mattered more than the code:
- It is read only. The assistant can run queries and pull reports. It cannot create, edit, or delete a single transaction. Nothing it does can change our accounting file.
- It signs in through our own Microsoft 365 accounts. Access is controlled by the same identity system that controls our email, so it can be revoked in one click like any other staff account.
- It talks to Intuit through our own registered application. No third party sits in the middle holding a copy of our financial data, and no other company’s outage becomes our problem.
- It runs on infrastructure we control. We know exactly where the data goes and where it stops.
That last set of points is the whole reason we built it rather than clicking install on something. When the data is your general ledger, “trust us” is not an answer.
What does the AI assistant actually do every day?
Three routines now run on their own, with no one starting them.
A morning financial pulse, every weekday at 7:00. The assistant pulls receivables, payables and anything that looks off, and it lands in the same brief that carries my calendar and my project list. I know the shape of the money before I have finished my coffee.
An aged receivables review, every Monday. Every open invoice with its age, how many days past due it is, when it was last emailed, and whether the customer has actually opened it. Then it drafts the follow up emails, one per invoice worth chasing, ready for me to read and send.
A bookkeeping activity check, every day. It watches the audit trail to see when the books were last touched by a human. That one sounds trivial. It is not. Knowing that nothing has been posted in eleven days is exactly the kind of thing you want to find out on day eleven rather than at year end.
What did it find in the first couple of weeks?
Working through contractor timesheets against what had actually been billed, it surfaced roughly a dozen hours of developer time that had been worked and never invoiced. That turned into a real invoice, sent the same day.
It also flagged a hosting client we had been carrying for three years for free. Not out of generosity, just out of nobody noticing. That is now a conversation about a proper plan.
And it turned invoice follow up from a task I felt guilty about into a Monday morning email I skim and approve.
None of this is glamorous. It is the unglamorous stuff that quietly costs small businesses real money every year.
Should a small business connect AI to its accounting system?
Yes, if the connection is read only and the access is controlled by your own identity system. You do not need to care about protocols or servers to get the point.
The useful version of AI in a small business is not a chatbot that writes your marketing copy. It is an assistant with safe, read only access to the systems where your business actually lives, and a standing instruction to look at them every day and tell you what changed.
Accounting is the obvious first one because the payoff is measurable. The same pattern works on your hosting bills, your project time, your support queue and your sales pipeline.
The parts that need doing properly are access, permissions and where the data lives. That is the work. The rest is plumbing.
Common questions
Can the AI change anything in our books?
No. The server we built exposes queries and reports only. There is no create, update or delete path, so there is no way for the assistant to alter a transaction even if it were asked to.
Where does the financial data go?
Between Intuit and infrastructure we control. We registered our own application with Intuit rather than using a reseller or a middleman service, so no third party holds a copy of the ledger.
Who can use it?
Only people who can sign in with a 40Two Microsoft 365 account. Access is governed by the same identity system as email, so removing someone removes their access to the books at the same time.
How long did it take to build?
A couple of weeks of part time work, and most of that was waiting on Intuit to approve production access rather than writing code.
Does this only work with QuickBooks?
No. QuickBooks Online is where we started because the payoff was easiest to measure. The same read only pattern applies to Xero, to hosting and infrastructure billing, to project time tracking, to a support queue and to a sales pipeline.
If you want one of these
We build these for clients now, on the same principles we used on ourselves: read only by default, your identity system, your infrastructure, your data staying where it should.
If you are running QuickBooks, or Xero, or a system you have been meaning to get a proper handle on, book a call and let us talk about what looking at it every day would be worth to you. It is a relaxed 30 minutes, even if it turns out not to be a fit.