Before any AI: find where the hours and the errors actually are. A wholesaler's invoice-to-purchase-order match as the worked case
Most conversations about AI in a business start with a tool. Someone has seen a demonstration, and the question is how soon it can be switched on. My first step is less exciting. It is a tally sheet. Before anyone chooses anything, I want to know where the hours in a process actually go and where the errors actually come from, counted on real work rather than remembered. This is step one of five, and the other four depend on it.
A typical case
Take a wholesaler with one accounts-payable clerk. Say 400 supplier invoices arrive a month, by email, as PDFs, a few on paper. Each has to be matched to its purchase order and to the warehouse's goods receipt before it is approved for payment. The numbers here are illustrative; the shape is common.
Most invoices match. The price is the order price, the quantity is what was received, and the clerk ticks it through. The rest are exceptions, and they are where the time goes. A price that is the supplier's new list price rather than the agreed one. A quantity invoiced in full when the warehouse received a short delivery. A unit of measure mix-up, cartons on the order and each on the invoice. Freight added that was meant to be included. Some are caught at entry. Others surface at month-end, or when the supplier's statement arrives and the balances do not agree.
Two weeks of counting would record three things. Minutes per invoice, split into the ones that matched first time and the ones that did not. Exceptions per hundred, by type and by supplier. What each exception cost, sorted into a few classes: an overpayment that went out, a credit that had to be chased, a late-payment charge or a lost early-payment discount, and the hour at month-end spent reconciling a statement.
How to count in two weeks
A tally sheet beside the work. One line per invoice: the time it was opened, the time it was finished, matched or exception, and if an exception, which kind. It costs a few seconds per invoice.
The ERP's own timestamps. Most systems record when an invoice was entered, when it was put on hold and when it was released. They miss the phone calls, but they cover months, not weeks.
A sample of fifty. Pull fifty paid invoices from the last quarter and check each against its order and receipt again. This finds the errors nobody caught, which the tally sheet cannot.
The person's own estimate, as a cross-check. Ask the clerk, before the count starts, how long a normal invoice takes and how many go wrong. Compare at the end. The gap is often the most useful number in the exercise.
Two things not to do. Do not ask for averages: "about five minutes each" hides the one invoice in ten that takes an hour. And do not trust a vendor's savings calculator, which starts from its own assumptions about your volumes.
What the count usually shows
The hours are rarely spread evenly. In a case like this one, the clean matches might take a minute or two each, and a minority of invoices, the exceptions, take most of the time the clerk spends on invoices. The errors concentrate too: a handful of suppliers, or one document type, such as invoices for part deliveries or anything with freight on it, produce most of them.
That changes the conversation. If one supplier causes a third of the price mismatches because its price file is never updated, the fix is a conversation with that supplier, or a price-update step on the buying side. If unit-of-measure errors come from one product range, the fix may be a validation rule in the ERP. Neither is AI, and both are cheaper. What is left after those fixes is the real candidate for automation, and now it has a size.
This is not just my habit. RAND's August 2024 report on why AI projects fail, based on interviews with 65 experienced data scientists and engineers, put misunderstanding or miscommunicating the problem to be solved first among its five root causes, ahead of missing data and inadequate infrastructure.
Why this is the plan
The count produces a ranked list: the tasks and exceptions in order of the hours and the money they cost. That list does three jobs. It decides which single problem gets attention first, which is step two. It becomes the baseline a small trial is judged against, which is step three: if exceptions take forty minutes each today, the trial has to show a number lower than forty, measured the same way. And it defines what "better" means before anyone has a stake in the answer.
Without the count, every later step is an argument. With it, most of them are arithmetic.
Checklist
- Two weeks of real work counted, not estimated.
- Minutes per item and errors per hundred, split by type and by supplier.
- A cost attached to each class of error.
- A ranked list of where the hours and the money go.
- The baseline written down before any tool is chosen.
This is step one of five. Next: Pick one AI idea, not twelve. All five steps are on the AI page.