Most AP automation implementations take longer than the vendor promised and produce lower initial accuracy than the demo showed. This isn't because the technology doesn't work — it's because the first 90 days are an integration and calibration problem as much as an extraction problem, and that phase requires more from the customer team than a typical software rollout.
This post describes what a realistic AP automation implementation looks like at 25,000 invoices per month: what the three phases actually involve, where the time goes, what can go wrong, and what metrics you should be tracking from day one. We're writing this from the experience of being on the vendor side of these rollouts — we've seen implementations go smoothly and we've seen them stall, and the difference is almost always in phase 1 preparation, not extraction capability.
Phase 1 (Weeks 1–3): ERP Mapping and Field Schema Alignment
Before a single invoice runs through extraction, you need to define what "extracted" means in the context of your ERP's data model. This sounds obvious but is where most implementations lose 1–2 weeks.
Your AP system has a specific invoice record structure. The field names, data types, and validation rules in your ERP are not the same as the field names in a generic extraction output schema. If your ERP uses a 10-digit internal vendor code (VD-0000123456) and the invoice just shows the vendor's legal name, that mapping has to be defined explicitly: extraction output "vendor_name" → ERP lookup against vendor master → ERP field "vendor_code."
The field mapping session is typically 2–4 hours with your ERP admin and AP manager. It covers:
- Which extracted fields map directly (invoice number, invoice date, total amount)
- Which extracted fields require lookup transforms (vendor name → vendor code, GL description → GL account number)
- Which ERP fields have no extraction source (fields your team currently populates manually that aren't on the invoice)
- Which fields are optional in extraction but required in ERP (fields that cause ERP validation failures if missing)
The last category is critical. If your ERP requires a cost center code on every invoice record and cost center codes don't appear on invoices, that field either gets routed to exception queue for human assignment or is assigned programmatically based on GL rules. This is a business logic decision, not a technical one, and it needs to be made in phase 1 before you try to run production volume.
ERP integration setup — connecting Fieldiq to your SAP, NetSuite, or Dynamics instance, authenticating the API connection, testing the push with sample records — typically takes 4–8 hours of engineering time. It's not a day-long engagement, but it does require your IT team's participation, and scheduling that time is usually the actual bottleneck.
Phase 2 (Weeks 3–6): Threshold Calibration and Parallel Running
The extraction confidence threshold determines what goes to automated processing versus exception queue. Set it too high and you're routing 15% of your volume to exceptions — more than your team can clear daily and no better than manual processing. Set it too low and errors slip through to your ERP. Finding the right threshold for your specific document corpus takes 2–3 weeks of parallel running.
Parallel running means running Fieldiq extractions alongside your existing process (human or BPO) simultaneously, comparing outputs, and measuring disagreements. This phase is where you find which fields have systematically lower confidence on your specific document types — maybe your vendors use non-standard tax code formats that the model is less certain about, or many of your invoices are scanned at lower DPI which affects handwritten field confidence.
A concrete example: a building materials distributor processing 25,000 invoices monthly found during parallel running that their construction subcontractor invoices — roughly 30% of volume — frequently used handwritten quantity corrections where a printed quantity had been crossed out and replaced. Those corrections generated lower confidence scores on the quantity field. The fix wasn't changing the threshold globally; it was adding a specific validation rule that flagged any quantity field showing a significant deviation from the expected range for that product category — catching the corrected quantities for human review without flagging the 70% of clean invoices.
By the end of parallel running, you should have a calibrated threshold per field type, a set of validation rules that match your AP policy, and a clear exception queue workflow that your team has practiced.
Phase 3 (Weeks 6–12): Live Running and Team Transition
Going live means switching from parallel running to Fieldiq as the primary processing path. For most teams, this is when the anxiety peaks — because it's also when the AP team's job function changes from "process these invoices" to "review these exceptions."
The job change is real and requires deliberate management. Your AP specialists are being asked to stop doing the task they've been doing and start doing a different task: reviewing flagged items, approving or correcting extractions, handling edge cases. Some team members adapt quickly; others find the change disorienting. The first two weeks of live running typically see higher exception rates as the team gets comfortable with the review workflow and as the system accumulates signal from corrections.
We're not saying the team transition is easy to manage — it's the human side of the implementation and it deserves the same attention as the technical side. What makes it smoother: transparency about what the exception queue looks like before go-live, having the team do practice runs in the exception review interface during parallel running, and giving team leads visibility into the metrics so they can see the volume and accuracy improving over time rather than just feeling the change abstractly.
The Metrics That Actually Matter in the First 90 Days
Start tracking these on day one of parallel running, not after go-live:
Straight-through processing rate (STP rate): the percentage of invoices that go from ingestion to ERP without any human touchpoint. This starts at whatever your calibrated threshold produces during parallel running — typically 85–92% in the first weeks — and should trend toward 95–98% as validation rules are tuned. If STP rate isn't improving week over week in the first 8 weeks, something is wrong in your threshold or rule configuration.
Exception queue age: the median time from exception flagging to human review and approval. If your queue is aging more than 24 hours, you have a staffing or workflow bottleneck in exception review that will create payment delays. This metric is a leading indicator — it tells you about problems before they become vendor complaints.
Extraction error rate: the percentage of documents where the extraction output was wrong in a field that reached your ERP (errors that bypassed exception routing). This should be below 1% from week one; if it's above 2%, your confidence threshold is set too low. Track this separately from exception rate — exceptions are caught and handled, errors are not.
Average turnaround time: end-to-end time from document receipt to ERP record creation. Your SLA with vendors probably includes payment terms that start at invoice receipt date. A turnaround under 4 hours preserves same-day processing for most business-hours invoices. Turnaround above 12 hours starts creating payment scheduling issues for tight-terms vendors.
What Causes Implementations to Stall
In our experience, implementations that slip past the 90-day mark almost always stall on one of three things: IT scheduling delays for ERP API setup, unclear ownership of the field mapping decisions (who makes the call when the GL code can't be auto-derived), or exception queue backlogs that develop when the AP team doesn't have enough capacity during the parallel running phase.
The IT scheduling issue is the most common. ERP API access and authentication often requires tickets through IT security, and those tickets queue behind other priorities. If you're planning an AP automation implementation, get IT on the calendar before you sign the vendor contract — not after. The API setup itself takes hours; getting it scheduled can take weeks.
At 25,000 invoices per month, a successful implementation should reach 95%+ STP rate within 90 days and full parallel running overhead should be eliminated within 45 days of go-live. Those are realistic targets, not optimistic ones, assuming the ERP and field mapping work is done before live processing starts.