Reading a merchant statement by hand is over, not because automation is fashionable, but because the manual process was never as reliable as it looked. A careful analyst can spend the better part of an hour decoding one statement, and still miss a reclassified interchange line, mis-key a basis-point figure, or quietly anchor the whole comparison to the one pricing model they know best. The replacement isn't a faster spreadsheet. It's AI that parses, classifies, and re-prices a statement across all six pricing models, shows every figure it pulled and every assumption it made, cross-checks its own work with a second model, and hands the result to a human to approve. The goal isn't speed for its own sake. It's an analysis you can audit.
The old way was slow, and that was the smaller problem
Manual statement analysis has two costs. The obvious one is time: pulling fees off a multi-page PDF, mapping them to interchange categories, normalizing oddly named line items, and rebuilding the effective rate. Do that across a book of merchants and the hours add up fast.
The cost that matters more is the one nobody sees. Manual work is where quiet errors live, a transposed figure, a fee bucketed under the wrong category, a surcharge line read as a processing cost. None of these announce themselves. They surface later, in a renegotiation or a renewal, as a number that no longer holds up.
And there's a subtler bias baked into the manual method. Analysts re-price into the model they're fluent in. For most of the industry, that default is interchange-plus. It's a defensible model, but it isn't the only one, and treating it as the destination quietly forecloses the other five before the merchant ever sees them. The merchant doesn't get a choice they were never shown.
The new way: parse, classify, re-price, across all six models
- 1Parse. Pull every line off the statement, not just the headline rate.
- 2Classify. Tag each fee: interchange, assessment, or processor markup.
- 3Re-price. Rebuild the cost across all six pricing models, neutrally.
- 4Cross-validate. Two models read it independently; disagreements get flagged.
- 5Human in the loop. A person reviews low-confidence cases before it goes out.
Audit-grade AI changes the shape of the work, not just the speed. The pipeline runs in stages, and each stage leaves a record.
- Parse. The statement, whatever the format, whatever the processor's naming conventions, is read into structured data: every fee line, basis-point charge, per-item cost, assessment, and pass-through, captured with its source location on the page.
- Classify. Each line is mapped to what it actually is: interchange, assessment, processor markup, equipment, PCI, batch, or other. Ambiguous lines are flagged, not guessed past.
- Re-price. The normalized cost base is run through all six pricing models, neutrally:
| Pricing model | What it optimizes for |
|---|---|
| Interchange-plus | Transparency, cost pass-through with a fixed markup |
| Tiered | Simplicity, qualified / mid / non-qualified buckets |
| Flat rate | Predictability, one blended rate, easy to budget |
| Surcharge | Cost recovery, card fees passed to the cardholder where permitted |
| Dual pricing | A posted cash price and a card price |
| Interchange optimization | Lower effective cost via better interchange qualification |
No model is the assumed answer. The output is a side-by-side that lets the merchant see, in their own numbers, what each structure would actually cost them. Interchange-plus earns its place on the merits or it doesn't, same as the other five.
Showing the work is the point, not a footnote
A re-priced number you can't trace is just a more confident guess. What makes this approach audit-grade is that every figure is traceable back to the line it came from on the original statement. When the model reports an effective rate, you can follow it down to the fees that produced it. When it classifies a line, the classification is visible and reversible. Nothing is asserted without the context behind it.
That matters for a practical reason: the analysis outlives the conversation it was made for. Six months later, when a merchant asks why a recommendation was made the way it was, the answer isn't "trust the tool." It's the audit trail, the parsed inputs, the classifications, the assumptions, and the math, preserved exactly as they were. Intelligence with a paper trail behind it.
Dual-model cross-validation: two readers, not one
Single-model output has a failure mode worth naming. A model can produce a clean, plausible, well-formatted number that is simply wrong, and because it's fluent, it's hard to catch by eye. The defense is structural, not hopeful: two independent models read the same statement, and their results are compared.
Where they agree, confidence is high. Where they diverge, a fee classified two different ways, an effective rate that doesn't reconcile, the disagreement is surfaced rather than silently averaged away. A discrepancy is treated as a signal to inspect, not a rounding error to bury. This is how you catch the quiet mistakes that one careful reader, human or machine, would miss alone.
Human-in-the-loop, by design
Automation here doesn't mean unattended. The AI does the work that machines are genuinely better at, reading hundreds of line items without fatigue, holding six pricing models in view at once, never getting bored on page four. A person stays in the loop for the judgment that machines shouldn't own: confirming flagged classifications, weighing which pricing structure fits a given merchant's mix and risk tolerance, and approving the output before it leaves the building.
The result is a clean division of labor. The machine proposes, with its reasoning exposed. The human disposes, with the full picture in front of them. Neither is asked to do the part they're worse at. That's not a compromise between speed and trust, it's how you get both.
What actually ends
What ends is the hour spent transcribing fees off a PDF, the silent classification error nobody catches until it costs something, and the reflex of re-pricing every merchant into one model because it's the one the analyst knows. What replaces it is faster, yes, but more importantly, it's legible. Every number has a source. Every classification is visible. Every recommendation is checked twice and approved by a person who can explain it.
For a merchant, that means a pricing comparison built on their real costs, across every model available to them, that holds up under questioning. For an agent, it means walking into a renewal with an analysis you can stand behind line by line. Manual statement analysis is ending. What takes its place isn't a black box that's faster. It's a glass box that shows its work.