Keeping automated books healthy

Two years after a successful implementation, a common state: the exception queue has four hundred items nobody works, nobody knows why the tolerance is 2%, the person who configured it has left, and a supplier has been miscoded since a rate change eighteen months ago.

Nothing broke. Nothing errored. The business changed and the configuration did not.

Automated finance functions degrade by drift, not by failure, which means the maintenance has to be scheduled rather than triggered — nothing will prompt it.

The four ways it drifts

The business changes and the rules do not. New product line, new supplier types, different transaction sizes, a restructure. Every one of these can make a threshold or a learned pattern wrong, and none produce an error.

Knowledge leaves. The person who set it up goes, taking the reasoning. Their successor either changes nothing out of caution or changes things without understanding why they were set that way.

The exception queue grows past the point of being worked. Gradual. There is no day on which it becomes unworkable; it just is, eventually, and then transactions are effectively processing unsupervised while everyone believes a review control exists.

Corrections migrate downstream. Someone starts fixing things with journals because it is faster than correcting in the workflow. The system stops learning, the same proposals return, and the audit trail stops explaining how balances were reached.

The schedule

Monthly — ten minutes

• Exception queue: is it shorter than last month, or longer?
• Is anything in suspense, and how old?
• Did every feed deliver? Check the date of the latest transaction on each bank account, not just the unmatched list.

Quarterly — two hours

• Sample 30–50 confidently-processed transactions and check them properly (sampling automated transactions)
• Review exception types by volume: which recur, and what is the upstream cause?
• Confirm every control account still decomposes into nameable items
• Check whether corrections are being made in the workflow or downstream

Annually — half a day

• Do the thresholds still fit the business? Volumes, transaction sizes, supplier mix
• Is the process documentation current, including the reasoning?
• Who owns each automated process, and are they still here?
• What configuration changed this year, and was it recorded?
• Are approval limits and hard constraints still right?

On any business change — immediately

New product line, new supplier group, acquisition, restructure, change in what you sell. These are the moments learned patterns stop applying, and a sample taken just afterwards is worth several taken during a stable period.

The five checks that catch most problems

If the full schedule will not happen, these five catch the majority:

1. Exception queue trend. Rising is the earliest warning available.
2. Suspense account contents. Anything sitting there is something nobody resolved.
3. Control accounts decompose. A balance that will not break into nameable items means something systematic upstream.
4. Feed freshness. A stopped feed is silent — reconciliation looks complete because there is nothing new to reconcile.
5. Where corrections happen. Downstream journals mean the system has stopped learning.

Fifteen minutes monthly. It is the difference between a function that stays healthy and one that quietly stops being controlled.

The ownership problem

The reason none of this happens by default: it is nobody's job.

Processing has a deadline. The close has a deadline. Reviewing whether a threshold set two years ago still makes sense has no deadline, no complaining stakeholder, and no visible consequence for skipping it — until the consequence arrives all at once.

The only reliable fix is naming someone. Per process, a person accountable for whether it still works, with the review dates in a calendar. See who is accountable for an automated entry.

The annual question

One question, asked once a year, catches most of the drift:

"If we were setting this up today, knowing what we now know about how this business operates, would we configure it this way?"

Usually the answer is no in one or two specific respects — and those respects are exactly where the drift has accumulated.

Common questions

How do automated accounting systems degrade?

By drift rather than failure. The business changes while the configuration does not, the person who set it up leaves with the reasoning, the exception queue gradually grows past the point where anyone works it, and corrections migrate from inside the workflow to downstream journals so the system stops learning. None of these produce an error message, which is why maintenance has to be scheduled rather than triggered.

What maintenance does automated accounting need?

Monthly, check the exception queue trend, the suspense account and whether every data feed is current. Quarterly, sample confidently-processed transactions, review recurring exception types for upstream causes and confirm control accounts still decompose. Annually, review whether thresholds still fit the business, whether documentation is current and who owns each process. Additionally, sample immediately after any significant business change.

What are the earliest warning signs?

A rising exception queue, anything accumulating in suspense, a control account balance that will not decompose into nameable items, a data feed that has stopped delivering without erroring, and corrections being made through downstream journals rather than inside the workflow. Those five checks take about fifteen minutes a month.

Why does maintenance get skipped?

Because it is nobody's job and has no deadline. Processing and the close have deadlines and complaining stakeholders, whereas reviewing whether a threshold set two years ago still makes sense has neither, and no visible consequence for skipping — until the consequence arrives all at once. Naming an accountable owner per process with review dates in a calendar is the only reliable remedy.

Related: sampling automated transactions · who is accountable for an automated entry · measuring whether AI accounting worked