Number7AI — Docs
Why AP teams get stuck at 60% STP and how to break through to 90%+
Straight-through processing (STP) is the metric finance teams feel in payroll and close quality — not headline OCR scores. AIdaptIQ reaches 90%+ STP on production-grade Indian AP and BPO workloads by pairing extraction with reconciliation-first validation and self-healing corrections before anything hits the ERP.
Last updated: April 2026
TL;DR
- STP here means: document upload → automated validation and repair where possible → ERP-ready export without human correction on that document under your approval rules.
- Industry AP programmes often plateau around ~60% STPwhen extraction stops at "good enough" fields and mismatches still land in the exception queue.
- Self-healing closes the gap: detect total / tax / line-item inconsistency, infer root cause (e.g. category subtotals, missing credits, balance forward), apply corrective actions, re-validate, and only then post or escalate with diagnostics.
- Fair comparison requires the same document mix, the same STP definition, and honest handling of vendor claims from different document types — do not interpolate medical forms or curated demos into mixed-format AP invoices without adjustment.
Why most "industry standard" AP stacks stall near ~60% STP
AP automation vendors and analyst surveys often report touchless or straight-through rates in the ~50–65% band for mixed real-world AP: multi-page PDFs, handwritten notes, tax-inclusive totals, credit memos, and vendor-specific layouts. That is not a failure of ML — it is what happens when the pipeline stops after field extraction and basic rules: every structural ambiguity becomes a human ticket.
AIdaptIQ's published production benchmarks (see Production benchmarks) describe a move from roughly ~60% to 90%+ STP in documented BPO AP conditions — with field-level accuracy and error-volume reductions measured on the same workload, not a cherry-picked subset.
What real self-healing looks like when invoices break in production
Self-healing is not "better OCR." It is a closed loop between extraction and accounting truth:
- Detect — mathematical and structural checks (line totals vs header, tax allocation, duplicate line patterns, impossible quantities).
- Diagnose — map mismatch signatures to known failure classes (subtotals mistaken for SKUs, missing deposits, balance forward omitted, tax double-counted, etc.).
- Act — apply deterministic corrections where confidence and policy allow (remove subtotal rows, insert missing credit lines, fix tax rows, merge split pages).
- Re-validate — rerun reconciliation until the document is internally consistent or escalates with a machine-generated explanation for a human.
That loop is what lifts STP without sacrificing auditability: every automated repair is traceable. For a full technical walkthrough with examples, see the markdown source at /docs/our-capabilities.md.
Extrapolate and interpolate without misleading your CFO
Buyers constantly compare percentages from different vendors. Most mismatches are definition errors, not product errors. Use this checklist before you interpolate anyone's STP or accuracy into your own ROI model:
| Question | Why it matters |
|---|---|
| What counts as "straight-through"? | Some teams count STP after operator OK on a pre-filled screen; others require zero touches before ERP post. Align definitions before comparing numbers. |
| Which documents are in the denominator? | Curated "golden" invoices inflate STP. Mixed PDFs, photos, and multi-invoice files collapse it. Always compare on the same production mix. |
| Are you mixing document classes? | High STP on narrow form types (e.g. healthcare intake) does not transfer to AP. Treat cross-industry numbers as non-comparable unless you explicitly model a correction factor — and disclose that assumption. |
| What happens to silent errors? | If a system auto-posts wrong GL splits with high "confidence," measured STP can look excellent while downstream rework explodes. Pair STP with silent-error sampling and exception-queue volume. |
| Interpolation rule of thumb | When only a headline number exists, model best / base / worst cases by shifting the assumed exception rate ±10–20 pts based on layout complexity — and run a 2-week parallel on your files to replace the guess with data. |
Where typical AP solutions break and why STP flatlines
These are architectural patterns, not digs at a single competitor. Most teams hit one or more of them in production:
- OCR / cloud vision only — great character readout, no guarantee that totals, taxes, and line semantics reconcile to an ERP-safe row.
- Extract-and-stop IDP — fields populate a UI, but structural errors (subtotals, missing credits) still require manual rework — STP caps out.
- LLM-only workflows — strong on paraphrase, weak on audited arithmetic and deterministic repair unless grounded in validation rules (see Why ChatGPT fails for AP).
- Rules without diagnosis — binary pass/fail checks fire exceptions but do not generate the minimal corrective action set, so operators re-type fixes.
- Weak audit trail on auto-changes — finance cannot defend automated adjustments to internal audit or tax — so automation gets throttled even if accuracy is high.
Self-healing directly targets the middle three: it is not replacing humans with a black box; it is reducing exception volume by fixing classes of problems that humans were already fixing by hand — faster and with receipts.
Related reading
- Production benchmarks — STP, accuracy, error reduction, cycle time tables.
- Reconciliation framework — three-pass validation before export.
- Failure modes taxonomy — where extraction silently fails AP.
- Competitor analysis (real invoice data)
- How to evaluate AP automation — buyer checklist aligned to these metrics.