Purchasing gets less attention than it deserves

Most companies watch sales closely and buying loosely.

That is backwards. Money leaks quietly on the purchasing side — duplicate orders, prices nobody checked against what was agreed, deliveries that arrived short and were paid in full, bills approved because nobody could remember the terms.

None of these are dramatic. Each is small. Together, across a year, they are not.

Here is where automation genuinely closes those gaps, and where it does not.

1. Reading supplier documents

The clearest win, as it is everywhere.

Bills, delivery notes, order confirmations and certificates read and turned into drafts. Vendor, dates, amounts and line items extracted.

Why purchasing benefits most: buying generates more inbound paper than any other function. A company processing three hundred documents a month gets real hours back.

The check. Compare against the document before posting. Tax treatment and account allocation are the fields most often needing correction — extraction reads what is printed, and it does not know your accounting policy.

2. Making three-way matching practical

This is the one that matters most, and it is worth understanding why.

Three-way matching compares what you ordered, what you actually received, and what the vendor charged. It catches quantities billed but not received, prices different from the agreement, duplicate bills, and charges for things you never ordered.

It is the most useful control in purchasing. It is also normally skipped, because doing it manually across hundreds of bills is impractical.

Automating the extraction is what makes it practical to actually do.

The requirement, and it is not a software one: your warehouse must count receipts properly. If receipts are validated at the ordered quantity without counting, the match compares two copies of the same assumption and proves nothing.

That habit matters more than any setting in this article.

3. Better reordering

Standard reordering rules use a fixed minimum and maximum, usually set during implementation and rarely revisited.

Forecast-based reordering estimates what you will actually need over the coming lead time, accounting for trend, seasonality and how unreliable each supplier is.

Measure two numbers together: stockouts, and stock value held. Anyone can eliminate one by worsening the other. Improving both is the test.

It needs accurate stock records. If your system says 40 and the shelf has 12, no forecast helps. Most reordering disappointments are really counting problems.

4. Flagging unusual purchases

Transactions that do not fit your pattern. A duplicate bill. An amount well outside the normal range for that vendor. A price different from the last three orders of the same item. A new bank account on an existing supplier.

That last one is worth its own mention. Payment redirection fraud — where someone impersonates a supplier and asks for bank details to be changed — is common and expensive. A flag on changed bank details is a simple control that catches it.

Expect false positives early while it learns what normal looks like for you. That is the learning period, not a failure.

5. Ordering supplier evaluation

Which vendors deliver on time, which are consistently late, which quietly raise prices, which have the most quality issues.

None of this needs prediction — it is your own history. But it is rarely looked at, because assembling it manually across hundreds of orders is work nobody has time for.

Why it pays: a supplier who is often ten days late needs a bigger buffer than one who is reliably on time. Knowing which is which turns an argument into a measurement.

What it does not fix

Four things, and they are the ones that actually cost most companies the most.

Buying outside the system. Someone orders by phone and the goods arrive with no purchase order. No amount of automation sees a transaction that was never recorded. This is a habit problem and it is fixed by policy, not software.

Receipts validated without counting. Everything downstream depends on this one human action.

Approval limits nobody enforces. If the threshold exists but people work around it, automating the check does not help.

Not negotiating. No system negotiates your prices. Better data gives you a stronger position; it does not use it for you.

Setting it up

1. Clean your vendor records. Duplicates and inconsistent names cause matching failures that look like AI errors and are not.

2. Measure a baseline. Documents per month, hours spent, duplicate bills caught, stockouts, stock value.

3. Start with document reading. Set up an email alias for bills.

4. Turn on three-way matching and make counting receipts a firm rule. Train it as a habit, not a policy document.

5. Add anomaly flagging, including changed bank details.

6. Look at reordering once your stock accuracy is genuinely good.

7. Review supplier performance quarterly. The report only helps if someone reads it.

Measuring it

Hours on document entry. Should drop clearly.

Mismatches caught before payment. This is money, directly.

Duplicate bills prevented. Often more than people expect.

Stockouts and stock value. Together, always.

Supplier on-time rate. Should improve once it is visible and discussed.

The rule that does not move

Nothing pays automatically.

Extraction can be very good. Payment is where a mistake becomes money leaving the company, and where fraud does its damage. A person reviews and approves.

The review costs seconds. One duplicate payment costs more than the feature saved in a year.

The short version

Automation in purchasing is genuinely valuable, and the value concentrates in two places: reading the paperwork, and making three-way matching practical enough to actually do.

The biggest leaks, though, are behavioural — buying outside the system and validating receipts without counting. Software does not close those.

Fix the counting habit first. Then automate the paperwork. And never let anything pay itself.