We Asked AI to Find Our Duplicate Payments. Here’s What Actually Happened.

Have you ever asked an AI tool to find the duplicate payments in your ERP system? It sounds like the perfect job for AI: mountains of transactions, patterns to spot, money to recover. So go ahead and ask it. Here’s how it usually goes.

First, it wants into your ERP

Before it can do anything, the AI needs access to your data. So the first thing it asks is whether it can connect to your ERP system. Stop and think about that for a second. Do you really want to hand a general-purpose AI engine direct access to your financial system of record? For most finance and IT teams, that’s a hard conversation before you’ve recovered a single dollar.

Then it hands you 30-year-old algorithms

Say you get past the access question. The AI goes to work, searching the web and its training data, and comes back with the same basic duplicate-detection logic that has been around for 30 years. Match on vendor, invoice number, amount, and date. Maybe a fuzzy match on the invoice number. It’s not wrong, exactly. It’s just not new, and it’s not deep.

And the results? A pile of false positives

Then it gives you a list. You start reviewing, and it’s mostly noise: legitimate payments flagged as duplicates, near-matches that aren’t, recurring charges that only look suspicious. So you refine the prompt. Tighten the rules. Try again. You get a slightly different pile of false positives. You refine again. You never quite land on an algorithm you can trust.

We know, because we tried it

We’re not guessing here. We ran this experiment ourselves. And to be fair, it does find some duplicates. But they tend to be the obvious ones: the kind a seasoned AP team can often catch on their own. Finding the duplicates that are genuinely hidden takes more than a general-purpose model and a good prompt.

The problem was never the AI. The problem is that finding real duplicate payments isn’t a generic pattern-matching exercise. It’s a domain problem. And that’s exactly where we come at it differently.

How we use AI differently

We start with experience, not algorithms

Our team has spent careers inside these systems. Some of us implemented ERP platforms; some came from large consulting firms. We know how invoices actually flow, where controls quietly fail, and why the same payment can go out twice. That context is the thing a general AI model simply doesn’t have.

We reverse-engineered the real recoveries

Over the years we’ve reverse-engineered thousands of duplicate payments and open vendor credits, many of them sitting right on the supplier’s own statement, and, just as important, we figured out how each one happened. Was it a data-entry error? A duplicate vendor record? A data feed that reprocessed the same invoices? Understanding the root cause is what lets you catch the next one.

We built targeted routines

That knowledge is baked into targeted routines that flag the specific transactions worth looking at, not a blind scan of everything. We’re not asking a model to guess what a duplicate looks like. We already know, and we’ve encoded it.

Then we use AI to remove the noise

This is where AI earns its place. We use it to analyze vendor patterns and weed out the false positives, leaving only the results that matter. Very targeted. Instead of “AI, go find duplicates,” it’s “here are the real candidates. Now help us confirm which ones are worth your team’s time.”

And AI that actually learns from you

Here’s the part we’re most proud of. As you move through our workflow, you mark each item as truly a duplicate or not. We use that feedback to learn. If you mark a certain type of item as not a duplicate for a particular vendor, our routines learn from it and stop surfacing similar items for that vendor going forward. The system gets sharper the more you use it, tuned to your data and your decisions.

The difference is AI plus experience

Pointing AI at your ERP and hoping for the best gets you recycled logic and a review queue full of noise. Pairing decades of hands-on AP and audit experience with AI that removes the noise and learns from your team gets you something else entirely: real, recoverable duplicates, with the root cause attached.

That’s the difference, and it’s why our clients recover money that generic tools, and generic prompts, leave on the table.

Curious what it would find in your data? Start a no-cost Proof of Value. We only get paid a percentage of what we recover.