Not all of them — just the ones the OCR tool cannot read, which somehow account for most of the work and all of the errors.
Straightforward invoices flow. The 20% that fail a three-way match consume the AP clerk's entire week, and nobody measures how long they sit.
Without cost per invoice, cycle time, and touchless rate, there is no baseline — so there is no way to prove the automation worked.
Approval routing takes longer than the discount window, and the savings quietly leave the business every month.
It captures invoices well and then hands you a CSV, because the last mile into your ERP was never in scope.
ERP module
Point solution
Custom build
Time to live
3–9 months
6–12 weeks
6–8 weeks
Fits your exact approval rules
Only if your rules match the vendor's model
Configurable within limits
Built to your rules
Handles unusual invoice formats
Poorly
Well, up to the template limit
Extraction is trained on your actual documents
Integration with the rest of your stack
Native to that ERP, weak elsewhere
The usual failure point
The starting point, not an afterthought
Cost shape
Bundled into the ERP contract
Per invoice or per user, forever
One-time build, then hosting
Who owns it
Your ERP vendor
The platform vendor
You — code, data and models
Best when
You run one ERP, standard processes, and volume is modest
Your process is genuinely standard and you want it working next month
Your volume justifies the build, your process is not standard, or the integration is the hard part
Say the quiet part. A point solution is often the right answer. If you process a few hundred standard invoices a month against a single ERP, buy the product — you will be live sooner and cheaper, and we will tell you so on the first call. Building makes sense when the per-invoice fee starts to look like a salary, when your approval logic is genuinely yours, or when the integration everyone treats as an afterthought is actually the whole project.
.webp)
[01]
[02]
[03]
Document extraction and exception handling are the hard parts of AP, and they are also the hard parts of everything next to it. Once the layer exists, extending it is a fraction of the original build.
We establish these as a baseline before the build starts, so the result is measured rather than claimed.
.svg)
Cost per invoice — fully loaded, including the exception handling everyone forgets to count
Touchless rate — the share of invoices that reach payment with no human intervention
Cycle time — receipt to approval, and approval to payment, tracked separately
Exception rate and exception ageing — how many fail, and how long they sit
Duplicate and overpayment rate — the number that pays for the project on its own
Early-payment discount capture — offered versus actually taken
First-pass match rate — invoices that match a PO without anyone touching them
A distributor processes around 4,000 supplier invoices a month against purchase orders in NetSuite. Roughly 40% arrive as PDF attachments in layouts that change without warning; the rest come through EDI and a supplier portal. An existing OCR tool reads header totals reliably and line items almost never, so three-way matching is done by hand. Two AP clerks spend most of their week on the 20% of invoices that fail a match, and nobody can say how long an exception has been sitting.
Line-item extraction trained on the client's own historical invoices rather than a generic template set, with per-field confidence scoring so low-confidence values are routed to a person instead of guessed. Three-way match against the PO and goods receipt with the client's own tolerance rules. An exception queue designed around how the clerks actually work — grouped by cause, aged, and prioritised by discount deadline. Write-back into NetSuite through its REST API, with a full audit trail.
Cost per invoice and touchless rate established as a baseline in week one, before anything is built. Field-level extraction accuracy benchmarked on a sample of the client's worst invoices and reported before the build is committed. Post-launch, the same metrics tracked weekly, with autonomy extended vendor by vendor only as measured accuracy allows.
In a scenario like this the gain is concentrated in two places: line-item capture removes most of the manual keying, and a properly designed exception queue removes most of the waiting. Header-only OCR typically leaves straight-through processing in the low tens of percent; line-item extraction with confidence routing is what moves it. The discount capture is often the line finance notices first, because approval delay — not disagreement — is usually why early-payment discounts expire.
Six weeks: one week process review and baseline, one week extraction benchmark, two weeks prototype on the highest-volume vendor group, two weeks integration, controls and handover.
A 300-person services firm processes about 900 invoices a month, almost none of them backed by a purchase order. Every invoice has to be coded to a project and a cost centre, then routed through a delegation of authority that lives partly in the ERP and partly in a spreadsheet the finance director maintains. The firm already bought an AP platform. It captures documents well, charges per invoice and per user, and hands finance a file that someone still has to reconcile into the ERP by hand. The bill has grown with headcount for three years.
We would keep the capture tool — it works — and rebuild only the three layers that stalled. GL and project coding suggested from 24 months of the firm's own coding history, surfaced with confidence and confirmed by the approver rather than applied silently. Approval routing that encodes the real delegation of authority, including the exceptions currently held in the spreadsheet. Direct write-back to the ERP, which is the step the platform never covered.
Coding suggestion accuracy against the firm's historical decisions, measured before launch on invoices the model has not seen. Approval cycle time by band. And a straightforward three-year cost comparison: current per-invoice and per-seat fees at projected headcount, against build plus hosting.
The extraction and matching layer built for payables is the same layer receivables needs. In a firm like this the natural second phase is remittance matching and cash application — a fraction of the original build, because the hard part already exists.
Five weeks, shorter than Scenario A because capture is already solved and there is no three-way match to model.
Our procurement RFP case is the closest published work See how we productionize document AI
We build custom AI systems for companies that have decided the off-the-shelf answer does not fit. Every engagement gets a named task force: a solutions architect, an AI/ML engineer, an integration specialist and a security engineer. You own the system, the code and the intellectual property. No per-invoice fee, no seat licence, no rented black box.
We work with organizations across the United States.
Document extraction and vision models
LLMs with private deployment
Agentic frameworks and vector databases
Core backend: Python, FastAPI, PostgreSQL
Workflow and approval UI: self-hosted low-code
Native integrations: SAP, Coupa, NetSuite, Workday, QuickBooks, Shopify Plus.
