When AP invoice automation is worth it: the honest decision tree

A typical mid-market AP (accounts payable) automation evaluation: three vendors, around 50,000 GBP of three-year cost, and pricing tiers designed for customers larger than the firm running the evaluation. The controller in front of the quotes is processing around 200 invoices a month, one entity, all on Xero. The cheapest tier on her shortlist starts to make sense at five times that volume.

For her shape, a 12,000 GBP one-off integration into her existing Xero setup usually handles every claim those quotes make at a fifth of the three-year cost.

The decision turns on two numbers - invoice volume per month and entity count - and almost everything else is a marketing layer on top. This post is the decision tree we walk controllers through, with the public pricing and benchmarks behind it.

The honest decision tree

Three branches cover most of the mid-market range we see. Numbers below are GBP and assume a UK/Ireland or EU operator. US figures track similarly with a different VAT (value-added tax) layer.

Under 100 invoices per month, single entity. Build nothing. Use what is already in your ERP (enterprise resource planning system). Xero ships native bill entry, duplicate flagging, approval routes, batch payment and an Aged Payables report at no extra subscription, with Melio-powered in-app US bill pay since the acquisition. Tighten the process: a single bills inbox, a vendor master cleanup, and a three-stage approval rule. The marginal cost of a 90-invoice month is one half day of an AP clerk. SaaS (software as a service) at this volume pays for shelfware.

100 to 500 invoices per month, single entity, on a mainstream ERP (Xero, QuickBooks Online, Sage, NetSuite). This is the band where a custom integration becomes a defensible bet, primarily because you own the integration. The work is bounded: an inbox-to-ERP capture flow with OCR (optical character recognition), an approval gate keyed to your matrix, a duplicate-detection rule, and a payments handoff. Scoping cost in our pipeline sits at 8,000 to 18,000 GBP one-off, with 1,500 to 3,000 GBP a year of light maintenance. On pure three-year TCO the SaaS subscription is usually within a few thousand pounds (the worked table below shows the figures at 200 invoices a month); the differentiator is who owns the failure modes when the demo collides with real supplier data, multi-currency rounding and sub-account coding rules.

500+ invoices per month, OR multi-entity, OR multi-currency with audit-heavy industries. Buy the SaaS. The volume earns the seat math, the multi-entity consolidation is the bit you would not want to rebuild, and the audit posture is what you are paying the vendor for. Tipalti's Advanced tier at 199 USD a month plus transaction fees, Stampli's unlimited-user custom-quote model, or Bill.com Corporate at 89 USD per user per month all become defensible at this volume. The mistake is assuming you are in this band before you are.

The transition points are fuzzy. A 350-invoice firm running across three legal entities probably belongs in branch three. A 600-invoice firm with one entity, one currency, one ERP and an unusually patient AP team can still live happily in branch two for another year. The frame is volume times complexity, where complexity carries roughly equal weight.

A worked TCO comparison: 200 invoices a month, Xero, single entity

Numbers below are three-year totals. Source figures are linked. The "do nothing but tighten process" column assumes one AP clerk at 0.4 full-time equivalent (FTE), priced at 38,000 GBP fully loaded a year, of which AP processing consumes 0.15 FTE at 200 invoices per month per the cross-industry rate of about 15 to 20 minutes of touch time on a manual invoice.

Cost line	Tighten process on Xero	Custom integration into Xero	Bill.com Corporate (5 seats)
Software / build	0	12,000 GBP one-off	5 x 89 USD x 36 = 16,020 USD (about 12,500 GBP)
Maintenance / support	0	2,500 GBP/yr x 3 = 7,500 GBP	included
Transaction fees (200/mo ACH equiv)	bank fees only	bank fees only	200 x 0.59 USD x 36 = 4,248 USD (about 3,300 GBP)
AP clerk touch time	0.15 FTE x 38k x 3 = 17,100 GBP	0.04 FTE x 38k x 3 = 4,560 GBP	0.04 FTE x 38k x 3 = 4,560 GBP
Implementation effort (internal)	1,500 GBP	4,000 GBP	6,000 GBP
Three-year total	18,600 GBP	28,560 GBP	26,360 GBP

Two things to notice. First, the SaaS and the custom integration are within a few thousand pounds of each other over three years - the choice is not really about cost, it is about who owns the long-term integration risk. Second, the "do nothing" column is the cheapest right now and the most expensive at 400 invoices a month. The decision tree's branch points are where the FTE line starts to bend.

The figures lean on the APQC cross-industry AP benchmark of 5.83 USD median cost per invoice processed, with top quartile at 2.07 USD and bottom quartile at 10 USD or more. The 79% best-in-class processing-cost gap from Ardent Partners' State of ePayables 2025 sets the upper bound of what automation can buy you.

What the OCR layer actually does (and does not) buy you

The vendor pitch deck always leans on the OCR demo. Three points worth knowing before the demo lands.

Accuracy claims and what they leave out

Independent benchmarks of document-AI invoice extraction settle in the 80s and low 90s on real-world supplier invoices, with both AWS Textract and Google Document AI coming in well below the vendor-published headlines. GPT-4o-class extraction touches 98% on clean documents and drops sharply on faxed or photocopied scans. Vendor-published numbers like Rossum's 98% top-line claim sit at the top of that band; they are best-case figures from controlled corpora and rarely survive contact with a real mid-market supplier base.

In practice the extraction quality is rarely the binding constraint. The binding constraint is what happens to the 5 to 10% of invoices the model gets wrong. A 97% accurate extractor on 500 invoices a month produces 15 invoices a month that need human review. If the SaaS routes those into a queue with no context, your AP clerk now reviews 15 invoices manually plus audits 485 that the model touched. That nets out to more work for the clerk than they had before.

The custom-integration version of this is easier to tune because you own the failure routing. A flagged invoice goes straight to the original email thread, the supplier history, and the GL (general ledger) coding suggestion in one screen. The SaaS version is faster to ship and harder to bend.

The four mistakes mid-market controllers keep making

Mistake one: paying for seats they do not need. Bill.com's price ladder is per user, and the Corporate tier at 89 USD a seat compounds fast across approvers, AP clerks, finance leadership and the occasional auditor. Controllers buy ten seats because the org chart has ten finance-adjacent people, then discover seven of them log in twice a year. Tipalti's no-per-user model and Stampli's unlimited-user pricing avoid the trap; Bill.com forces it. Count active approvers when you size a seat-based plan, never org-chart headcount.

Mistake two: assuming the ERP cannot already do this. Xero, QuickBooks Online and Sage Intacct ship a meaningful chunk of AP automation in the base subscription, and most controllers running them at 200 invoices a month have not exhausted the native feature set. NetSuite is the partial exception - the core AP module handles vendor bill entry and 3-way PO (purchase order) matching, but OCR capture, AI GL coding and robust duplicate-payment detection sit in the NetSuite Bill Capture add-on or a third-party SuiteApp. The right question to start with is "what is already running in our ERP that we have not turned on?" - and only once that floor is established does "do we need AP automation" become a real question.

Mistake three: not pricing the integration risk. Every SaaS vendor demos a clean sync to Xero or NetSuite. Production reality is more textured: timestamp drift, vendor-master mismatches, sub-account coding rules the SaaS does not know about, multi-currency journal entries that round differently in two systems. The most common failure mode for AP automation rollouts is exactly this: integration that "works" in the demo and then forces manual reconciliation against the ERP every month-end. The SaaS does not own that risk. You do.

Mistake four: choosing the tool before tightening the process. AP automation deployed on top of a messy supplier master, inconsistent GL coding rules and informal approval routes encodes the mess. The discipline is to clean the process first, run it manually for a month with a tightened workflow, then automate. Mid-market firms that skip the cleanup pay full price for the SaaS and still spend the first quarter cleaning what they should have cleaned before signing.

Where Appify fits

Appify is a neutral builder. We do not resell SaaS. When the decision tree points at branch three, we will say so and help you scope the SaaS rollout. When it points at branch two, we build the integration into your existing ERP - typically 8,000 to 18,000 GBP, four to eight weeks, owned outright at the end. When it points at branch one, we will tell you to spend three days tightening your bills inbox instead.

The framing is the same we used in our note on what actually moves AI unit economics: a system that costs more than the analyst it replaces is a hobby with a budget code. AP automation at the wrong volume is exactly that hobby.

If you are sitting on a vendor shortlist and want a second pair of eyes before the signature, book a discovery call. Bring the quotes, your invoice volume, your ERP and your entity count. We will run the decision tree with you in 45 minutes and tell you which branch you are actually in.

The short version

Under 100 invoices a month, tighten the process. 100 to 500 single-entity, the custom integration and the SaaS land within a few thousand pounds over three years; the choice comes down to integration control rather than cost. 500+ or multi-entity, the SaaS earns its seat. The four mistakes - paying for seats you do not need, ignoring native ERP features, mispricing integration risk, automating mess - explain most of the 50,000 GBP quotes that should have been 12,000 GBP integrations.

The honest version of this decision is shorter than any vendor wants it to be. That is usually a sign you should read it twice before signing.