Why “98% classification” doesn’t survive your vendor master

Every spend tool quotes a classification accuracy north of 98%. No vendor master we’ve ever seen is that clean. The number isn’t a lie — it’s just built in a way that doesn’t mean what you think.

Shop for a spend analytics tool and you’ll meet the same statistic everywhere: “98% classification accuracy,” sometimes higher. It’s meant to retire the question of data quality before you ask it. But if you’ve ever actually looked at your vendor master, your gut says what ours does: nothing you own is that clean. Your gut is right. Here’s how the number gets built.

Coverage is not accuracy

The first trick is in the word. “Coverage” means a category got assigned. “Accuracy” means it was the right one. A tool can report 98% coverage while a chunk of those assignments are generic or wrong — everything dumped into “Other Services” still counts as covered. When a vendor quotes one big number, ask which one it is. The honest ones split them, and the accuracy figure is always the lower of the two.

By spend, not by vendor

The second trick is the denominator. “98% of spend classified” is easy, because your top 50 suppliers are most of your spend and they’re the easy ones to identify. The mess lives in the long tail — thousands of small, oddly-named, one-off vendors — which is a rounding error in dollars but half your vendor list. So “98% of spend classified” can sit happily next to a vendor file that’s still full of junk. Measured by vendor, the number looks very different.

Anyone who promises you 98% hasn’t seen your vendor master.

The number is really people

The third thing the stat doesn’t mention: how it was hit. The enterprise players that guarantee it are software plus managed services — analysts and a managed taxonomy clean the data to reach the number, on enterprise files that were already prepped. It’s a real result, but it’s not what a raw two-file upload does on its own, and it’s not free.

What to look at instead

Stop shopping for the biggest number and start asking to see the work:

• Can it show you the gaps? A tool should tell you which fields loaded cleanly and which are thin before you trust a single total. SpendAware’s data-health view does exactly that — email at 33%, NAICS at zero — so you know what you’re working with.
• Do you review and override? The AI should do the first pass and flag what it can’t place, with the final call yours.
• How do you pay? We charge for enrichment only when we successfully assign a code. If we can’t classify a vendor, there’s no charge — which is a stronger accuracy guarantee than a footnote.

We won’t promise you 98%. We’ll show you the truth about your data and let you decide — which, on a file as messy as a real vendor master, is worth more than a number you can’t check.