What is touchless rate, and what AP quality KPIs should sit beside it?

Honest ranges, not vendor-slide numbers: many organizations sit below 30% truly touchless; well-run automated teams with strong PO discipline commonly reach 50 - 70%; the highest published claims concentrate in PO-heavy, EDI-heavy environments and often use looser definitions of "touch." Two cautions: definitional games (counting "human approved the AI's work" as touchless) inflate comparisons, and your invoice mix - PO share, vendor data quality, entity complexity - bounds what's achievable more than the software does. Demand the definition before believing any number, including your own.

Categorize every exception at creation - price variance, quantity variance, missing PO, coding failure, vendor-data mismatch, duplicate suspect - then Pareto by vendor, category, and entity monthly. The distribution is always concentrated: a handful of vendors or one entity's process typically drives the bulk. Fix at the source (vendor invoicing instructions, PO discipline in one department, a tolerance setting), not in the queue. Track exception rate *and* average exception age - a low rate with slow resolution still wrecks cycle time.

Exception rate is the share of invoices that fall out of the standard flow and require investigation. Standard categories: price variance, quantity variance, missing or unmatched PO, missing receipt, coding failure, vendor master mismatch, duplicate suspicion, and policy violations. Consistent categorization at creation is what makes the metric actionable.

First-pass match rate is the share of PO-backed invoices that match without human help on the first attempt. Well-run 2-way matching commonly reaches the high 80s to low 90s percent; 3-way runs lower because receiving data adds a failure mode. Below ~75%, the problem is usually upstream - PO quality and receiving discipline - not the matching engine.

Sample posted invoices monthly and score fields against source documents: amounts, dates, coding, vendor. Field-level error rates above 1 - 2% on financial fields, or any recurring systematic error (same field, same vendor type), justify process or tooling change - systematic errors compound at scale in ways random ones don't.

Track three numbers: duplicates blocked before payment (control effectiveness), duplicates paid and recovered (detection effectiveness), and net loss. Industry recovery audits commonly report duplicate payments in the range of a few hundredths of a percent up to roughly half a percent of disbursements depending on control maturity - your blocked-vs-paid ratio is the trend that matters.

Decompose the touched population: invoices touched for data correction (capture/data constraint), for coding (rules/learning constraint), for matching failures (vendor and PO-quality constraint), or waiting on people (approver constraint). The plateau almost always has one dominant cause, and it's frequently vendor-side invoice quality - which is fixed with vendor instructions, not software settings.

Late payment rate is invoices paid after due date as a share of payments. Split it by cause: invoices that were approval-stuck past due date (process-caused) versus invoices approved in time but scheduled late (treasury decision). Reporting them as one number lets a cash-management choice masquerade as an AP failure - or vice versa.

Show counts and value by status, stage age, owner, entity, and due date. Separate approval waits from missing vendor information, receipts, coding corrections, and ERP recovery. Include open-backlog aging as well as completed work; completed-only reporting can hide difficult unresolved invoices.

Define whether a touch means an edit, approval, follow-up, or another intervention. Reopening a record is not necessarily another unit of processing work. Report both invoices needing intervention and the number or duration of interventions.

For a defined stage and period, divide invoices completed without the specified manual intervention by all eligible invoices completed through that stage. Document exclusions, cancellations, duplicates, and reopened records. If 600 of 1,000 eligible invoices pass capture without correction, that capture rate is 60%; it does not establish touchless approval, payment, or time saved.

Group exceptions by cause, vendor, entity, and stage. Rank frequency and effort, assign an owner, and compare similar populations after the change. Separate approval bottlenecks from upstream data problems. Review rework and error outcomes to establish whether work disappeared or moved elsewhere.

Stampli connects AP activity and invoice context, while Stampli Deep Finance supports source-backed investigation of finance questions. Confirm which operational metrics, event fields, filters, and exports are available in your scope. Do not equate an AI work-completion claim with a process-wide automation rate or guaranteed time savings.