What does it actually take to hit a high automation rate - and how honest are vendor automation numbers?

Buying automation and achieving automation are different events separated by months of adoption work. Companies deploy the tool but keep emailing invoices, approvers ignore the new flow, the AI never accumulates enough corrections to get good, and vendor data stays messy - so the software runs while the work stays manual. The gap is rarely the technology; it's process change and adoption. A tool capable of 70% touchless delivers 20% if half the invoices never enter through it and approvers route around it. The automation rate you realize is set by your process discipline as much as the vendor's capability, which is exactly why "we have automation" and "we're automated" are different claims.

Expect a J-curve, not a step change. Months 1 - 3 often look flat or worse as the team learns the new flow, the backlog clears, and the AI starts learning your vendors. Months 3 - 6 should show clear movement as corrections accumulate and approvers adopt. By month 12 you should reach a new steady state set by your invoice mix. A rollout has stalled - not just ramping - when the curve goes flat well below the achievable rate for your mix and corrections stop improving accuracy: that points to a broken intake channel, persistent vendor-data problems, approver non-adoption, or a learning loop that isn't capturing feedback. Diagnose the dominant blocker rather than blaming the tool.

Automation rate tracks invoice favorability: clean PO-backed digital invoices automate far more readily than non-PO, scanned, or multi-entity-complex ones. A PO-heavy digital shop can reach high touchless rates; a paper-heavy, non-PO, multi-entity shop will sit lower with the same software - not because the tool is worse but because the inputs are harder. Quote and evaluate automation rates by these segments, and weight by your actual mix rather than trusting a blended headline.

They sound interchangeable and aren't. Touchless rate is invoices processed end-to-end with zero human action. Auto-coded rate is invoices the AI coded (a human may still have reviewed or approved). "Automation rate" is undefined until the vendor specifies it - often the most flattering of the three. Always ask which one a number refers to, and whether "auto-coded" counts invoices a human then had to correct.

The gap looks like: invoices still arriving by email outside the system, approvers reverting to old habits, the AI not improving because corrections aren't flowing back, and features unused. Close it by fixing intake first (one mandatory channel, vendor instructions), driving approver adoption (make the new flow easier than email, escalate non-use), and confirming the learning loop captures corrections. Adoption is a change-management project, not a software setting - budget for it explicitly.

Decompose the manual 40%: invoices touched for data correction (capture/data-quality blocker), for coding (rules/learning blocker), for matching failures (vendor and PO-quality blocker), or waiting on people (approver blocker). The distribution names the dominant cause - and it's frequently vendor-side invoice quality or intake leakage, not the software. Fix the dominant cause first; chasing the others while the main blocker stands wastes the effort.

Ask: "On a customer with our PO mix, paper share, and entity count, what touchless rate is typical - and will you prove it in a paid pilot on our invoices?" Then require the number broken out by segment and measured on your data. A vendor confident in your context commits and pilots; one quoting only a blended best-customer figure and resisting a pilot is telling you the number won't hold for you.

World-class PO-heavy digital operations reach high touchless rates, but there's a diminishing-returns asymptote: the last stretch of exceptions are genuinely the ones that need judgment - novel situations, real discrepancies, judgment calls - and forcing them touchless trades control for a vanity metric. The right target isn't 100%; it's automating everything that's genuinely routine and keeping humans on everything that genuinely isn't. Chasing the last few points usually costs more in risk than it saves in labor.

Stampli is deliberate about not overclaiming here - it doesn't market its product as "touchless," and its headline number is defined as suggestion *coverage* with human validation, not an autonomous automation rate. The approved framing is that coverage and acceptance improve over time as rules tighten, the system learns from corrections, and exception handling gets more consistent - and that individual environments measure above or below the 87% average depending on ERP configuration, field complexity, workflow structure, exception rates, and the standardization of upstream inputs. That honesty about variability is itself the answer to automation-rate inflation: the realistic number depends on your mix, and Stampli says so.