An AI-assisted month-end close uncovered roughly $10,000 in payment discrepancies. The bigger surprise was what it suggested about running a business.
I thought our books were in reasonably good shape.
Someone on our team handled the bookkeeping, and I reviewed the financials each month. At a high level, the bank balance generally looked consistent with what I was seeing. Nothing had given me much reason to worry.
Then I asked AI to help close the books.
It began with ordinary questions about how to treat particular transactions. Then it started following discrepancies into older records, comparing accounting entries with bank activity.
Eventually, it found customer invoices marked paid without matching bank receipts. The discrepancies totaled approximately $10,000.
I asked our bookkeeping person to check. They confirmed the discrepancies.
That got my attention. But what stayed with me was not just the amount.
I had given AI a version of the closing instructions we already used in the business. Those procedures needed clarification, and I remained involved in decisions and approvals. But they had given it a starting point for doing useful work.
It made me wonder whether the processes we had built to help people run the business could also help us expand what those people—and I—could accomplish.
The starting point was already there
I run a professional services business. Our accounting mainly involves service revenue, payroll, software vendors, and a smaller amount of product sales.
I was considering changing accounting software and wanted cleaner records before a possible move. That gave me a practical reason to look more closely at the books.
I already had a month-end closing checklist written for a person to follow. It covered transaction categories, matching receipts to invoices, bank reconciliation, older unresolved items, unusual vendor charges, and other checks.
Rather than start with a broad request to “do the bookkeeping,” I reviewed that checklist, updated parts of it, and asked AI to work through it.
As the work progressed, it asked questions, investigated exceptions, and brought proposed corrections to me for approval. It identified duplicate or incorrectly recorded items and problems with amortization.
I had defined what I wanted checked, but I had not prescribed every investigative step. Watching it pursue particular discrepancies was part of what surprised me.
The payment findings were the most consequential example. But finding discrepancies is not the same as recovering money. Correcting the records and collecting any amounts still due are separate outcomes. Some of those amounts may prove difficult to collect.
Even with that distinction, the review had surfaced something important that our existing routine had not.
The work also changed the instructions
This was not a case of handing over a perfect checklist and watching everything happen without further involvement.
The work exposed places where my instructions needed to be more precise.
For example, investigating an unresolved item might require looking at earlier records. But I did not want every monthly close to become an open-ended investigation of the company’s entire accounting history.
I clarified that the normal scope should stay within the closing month and the prior month, with earlier research reserved for specific unresolved items.
I also asked for unreconciled items to be traced to bank receipts or withdrawals, and clarified that a bank transaction might match a journal entry rather than an invoice or bill.
Another refinement was to keep a separate record of issues and resolutions, and refer back to unresolved items in future periods.
These were practical details, but they changed how I thought about directing the work.
A checklist says what needs to be done. The work also needs boundaries: what evidence is sufficient, when to investigate further, when to ask a question, and what should carry forward.
AI did not remove that management responsibility. It gave me another reason to make it explicit.
Having a process is not the same as knowing it worked
My immediate reaction to the discrepancies was not to look for someone to blame.
What concerned me was the gap between having a written process and knowing that it was producing the intended result.
It would be easy to say that someone had failed to follow the instructions. But the findings alone do not establish whether the problem was a missed step, an unclear instruction, an inadequate review, or some combination of those things.
What I could see was narrower: we had documented procedures, the bookkeeping was being done, and I was reviewing the financials. Yet these discrepancies had not surfaced.
That raised a more useful question for me:
What evidence tells us that a process is working—not just that the work has been marked complete?
In this case, an invoice marked paid was not enough. The review needed to connect the accounting record to the underlying bank activity.
Would a thorough human review have found the same issues? Quite possibly. This was not a controlled comparison between AI and a skilled bookkeeper.
The useful result was that this AI-assisted work, with my direction and our team’s checking, helped us examine records more closely and identify problems.
That is meaningful without turning it into a claim that AI is better than people.
Why this feels bigger than bookkeeping
I have long thought of a business as a system of repeatable processes.
That belief is part of why we document how work should be done. The aim is to make the business more consistent and less dependent on someone remembering every detail.
Until this experience, I had mainly thought about those procedures as instructions for people.
Now I see another possible use.
The same work of explaining a process—its purpose, expected outcome, rules, and exceptions—can provide a foundation for directing AI-assisted work.
That does not mean every procedure is ready to automate. Mine needed refinement. Nor does it mean AI will follow every instruction correctly or recognize every mistake.
One useful round of work is not yet proof of a dependable monthly process. I was still answering questions, approving changes, and involving our team in checking findings. Some issues required further work rather than a clean, immediate resolution.
But those limits do not erase the opportunity.
What excites me is that the effort already invested in documenting the business may have more value than I had realized. We were not starting from nothing. We had a description of how important work should happen, and we could use it, test it, and improve it.
The documentation was no longer just something to give someone during training. It was becoming part of how I could direct work in a different way.
More capacity, not less responsibility
It is tempting to turn every AI experience into a prediction about which jobs will disappear.
I do not think this case settles that question.
What it changed for me was my sense of how much a person or a small team might be able to accomplish.
There is a meaningful possibility here: people with experience and judgment may be able to apply them across more work than their individual time previously allowed.
But the judgment still matters. Someone needs to decide what a good result looks like, recognize when an answer is insufficient, and determine which changes should be made.
In this experiment, that work did not disappear. It became a visible part of managing the process.
For another owner considering a similar test, I would start with one recurring task—not with a goal of “putting AI into the business.”
Choose work you understand well enough to evaluate. Look at the instructions you already have. Ask what evidence would show that the task was completed correctly, which exceptions require judgment, and where approval is necessary.
Then test whether AI actually helps. A simpler check may be enough for some tasks; AI does not need to be the answer to every operating problem.
For me, the month-end close offered a concrete place to begin. It also showed me that the quality of the instructions and the quality of the review deserve as much attention as the tool.
I started this exercise because I wanted cleaner books.
I came away looking differently at the business systems we had already built.
The idea of “duplicating yourself” has obvious appeal when there is always more work than time. This was not that. But it made the possibility of extending what one person—and one small team—can accomplish feel much more real.
The $10,000 got my attention. The possibility of doing more with the knowledge already inside our business is what has kept it.
