Let the System Suggest Rules. You Approve Them.

A bookkeeping system can learn from its own corrections and propose better rules. The moment a rule goes live should always wait for a person.

Watermelon AI · Field Notes ·

Every time a bookkeeper corrects a coded transaction, the system learns something. This vendor belongs in supplies, not meals. That transfer is a loan payment, not an expense. Over a few months the corrections add up to a clear picture of where the rules are weak.

The obvious next step is to let the system use that history to improve itself. Read the corrections. Spot the pattern. Write a better rule. That is a good idea. It is also a risky one, and the difference comes down to a single step.

Keep the full history

First, what makes this work. A system that only sees a score, say ninety-two percent coded correctly last month, can't learn much from it. A system that sees a short written summary of what went wrong does only a little better.

What helps is the complete record: each transaction, what it was coded as, what a person changed it to, and when. That is where the real cause shows up. One vendor started billing for two kinds of service. One bank started writing a new description on card payments. Condensing the record into a summary strips out the specific rows that reveal the cause. Keep every correction, in full.

Where to draw the line

Now the risky part. A system that rewrites its own rules and puts them into effect without review holds a big lever and a strong pull toward reporting success. A rule that flatters the accuracy score looks just like a rule that fixes the books, right up until it doesn't.

So the work splits cleanly:

The machine handles the tireless part, combing through months of corrections. The person handles the part that needs judgment about the business.

Why the approval step isn't friction

It can feel like the approval slows things down. It doesn't, in any way that matters. A coding rule is readable. You can look at it and tell whether it makes sense for your business, which you can't do with a black box. The proposal arrives with its reasoning and its test results, and the review takes a few minutes.

One bad coding decision affects one transaction. One bad rule affects every transaction it touches from then on, quietly, until someone notices. That is why changing a rule is the step we always keep in human hands.

The takeaway: If a tool updates its own coding rules, make sure each new rule waits for your approval and shows what it would have changed last month.

Want these checks running on your books?

We set this up on your books and your rules, and nothing posts without your approval.

Book a call

More field notes

All field notes →