How static payment waterfalls work — and where they leak
A static payment waterfall is a fixed sequence: try VCard first, fall back to ACH if declined, use check as last resort. The logic is simple and the implementation is straightforward. At low volume with a small supplier base, it performs reasonably well.
At scale, it leaks. A static waterfall does not know that supplier X has a 94% VCard acceptance rate on Monday and a 12% rate on Friday because their AR team only processes cards on Monday. It does not know that payments above $25,000 to supplier Y are almost always declined on VCard and should route directly to ACH. It does not know that a newly enrolled supplier has not yet configured their AR system to auto-process cards, so every VCard attempt generates a manual processing call that reduces future acceptance. The static waterfall tries VCard, gets declined, routes to ACH, and leaves interchange revenue on the floor — repeatedly, at scale.
At $100M annual payment volume, a 5% improvement in VCard routing accuracy — directing payments to VCard where they will succeed and to ACH where they won't — generates approximately $875,000 in additional annual interchange at 175 basis points. That is the gap between a static waterfall and a model that learns from acceptance patterns.
What dynamic AI routing actually optimizes
Dynamic AI routing is not a monolithic AI system making opaque decisions. It is a model trained on transaction outcomes that improves specific, measurable routing decisions. The inputs are structured: supplier ID, payment amount, payment day/time, historical acceptance rate by supplier and rail, current supplier AR configuration status, payment urgency. The output is a routing recommendation: which rail to attempt first for this specific payment, with what confidence.
The model learns over time. A supplier that enrolls for VCard but has low initial STP rates gets routed to ACH while the enrollment matures. A supplier that achieves high STP rates gets prioritized for VCard on every eligible payment. The routing logic updates continuously rather than requiring manual waterfall reconfiguration.
Three specific optimizations that AI routing improves over static rules:
Supplier-level acceptance prediction. Rather than a blanket "try VCard first" rule, the model predicts acceptance probability for each specific supplier on each specific payment. Suppliers with low VCard probability get routed to monetized ACH directly — generating $1.50–$3.00 per payment instead of a failed VCard attempt followed by ACH at $0.
Amount-based routing. VCard acceptance rates decline on large payments for many suppliers — their AR systems have authorization limits. A model trained on acceptance data routes payments above a supplier-specific threshold directly to ACH or wire, avoiding failed VCard attempts on payments that would succeed via ACH.
Timing optimization. Some suppliers process VCard payments only during business hours or only on specific days. Routing VCard payments to those suppliers during their processing windows — and routing to ACH outside those windows for time-sensitive payments — improves acceptance rates without requiring supplier behavior changes.
The data infrastructure required before AI routing works
Dynamic routing optimization requires data that most AP payment programs are not yet capturing systematically. The model is only as good as its training data, and the training data requires deliberate instrumentation.
What needs to be captured at the transaction level: payment attempt outcome (success, decline, timeout, error), decline reason code, time to settlement, supplier processing confirmation timing, whether the payment required manual intervention at the supplier's AR level, and the specific VCard controls applied. This data needs to be stored at the transaction level, associated with the supplier ID, and accessible to the routing model.
Most AP platforms capture payment outcome at the program level but not at the supplier-transaction level with the granularity the model needs. Building the data layer is the prerequisite investment — before the routing model, before the optimization. Programs that have been capturing clean transaction-level data for 12+ months have a significant head start on programs starting from scratch.
What AI routing cannot do
Dynamic routing optimization addresses the decision of which rail to attempt. It does not address the underlying economics of each rail — BIN category, BaaS middleware share, monetized ACH rate. A well-optimized routing model running on a poorly structured program still leaves economics on the table through the structural gaps.
It also does not solve supplier enrollment. A model that predicts low VCard acceptance for a supplier is identifying a supplier who hasn't been fully enrolled — the correct response is supplier enablement, not permanent ACH routing. Dynamic routing should feed a supplier development program: suppliers below an acceptance threshold get flagged for AR integration support, not permanently excluded from the VCard waterfall.
The routing optimization and the program economics work in parallel, not in sequence. The best outcome is a well-structured program — direct bank relationship, commercial BIN, monetized ACH — with dynamic routing optimization running on top of it.
Want to model the revenue impact of routing optimization on your payment volume? The AP payments calculator lets you model different VCard acceptance rates and see the economics difference. Or talk with us about your specific program.