ShippedFintechFinance OperationsAI / MLB2B SaaS
// build study

AP Copilot

An AI accounts-payable copilot that reads invoices, matches them to POs, and routes clean approvals

Client:Mid-market finance team (NDA)
Duration8 weeks
ScopeDesign and build of the AP Copilot MVP front end — invoice inbox triage plus the human-in-the-loop review and approval flow

// AP Copilot — an AI accounts-payable copilot that reads invoices, matches them to POs and routes approvals.

// the numbers

Figures from the engagement, shared under NDA.

We publish them unchanged and can’t break them down any further here. The reference build you can open runs entirely on sample data.

82%Invoices auto-matched to PO
98%Avg field extraction confidence
1,284Invoices processed / month
1.7%Invoices needing human review
01In brief

If you read one part of this page, read these 6 lines.

  • Exceptions-first inbox that only surfaces invoices needing a human
  • Every extracted field shows a confidence ring so reviewers know where to look
  • Line-by-line invoice-vs-PO reconciliation with an in-tolerance verdict
  • Anomaly panel explains duplicate, price-variance and budget reasoning in plain language
  • Approval timeline captures ingestion, extraction, matching and routing for audit
  • Hand-authored inline-SVG charts, rings and sparklines — no chart libraries
02Overview

What it does — and which part of it we built.

AP Copilot is an AI accounts-payable workspace that turns a noisy invoice inbox into a controlled, largely touchless approval pipeline.

Invoices arrive by email, the copilot extracts every field with a per-field confidence score, matches line items against the originating purchase order within a price tolerance, and runs duplicate, price-variance and budget checks before routing to the right approver under policy. Only genuine exceptions surface for a human, and each one carries the model's full reasoning and an auditable trail. We designed and built the MVP front end as a premium light SaaS product — an "Invoice inbox" triage surface and a human-in-the-loop invoice review — to a Linear/Ramp bar of polish.

03The screens

Every screen here is from the reference build, not the client’s system.

The captions say what you are looking at. Every name and record on them is sample data.

04Outcomes

What the engagement left the client with.

  • 82% of invoices matched to a PO and cleared without manual touch
  • Exceptions cut to 1.7% of monthly volume, focusing reviewers on real risk
  • Duplicate and price-variance checks run against the full quarter of history before payment
  • Confidence-scored extraction gives controllers a defensible, auditable approval trail
05How it is built

The stack was picked per constraint, not per house habit.

9 technologies across 4 layers — the shape of the stack follows the problem.

AI & ML
LLM + OCR extraction pipeline
Frontend & Mobile
Next.jsReactTypeScriptTailwind CSS
Data & Infrastructure
PostgreSQLPrismaVercel
Backend & APIs
Node.js
Every feature in the build (6)
  • Email-to-inbox invoice ingestion with PDF source capture
  • AI field extraction with a confidence ring on every field
  • Line-level PO matching with configurable price tolerance
  • Duplicate, price-variance and budget anomaly checks
  • Policy-based approval routing with a full audit trail
  • Spend analytics by category and exceptions-by-type breakdown
// see for yourself

Don’t take our word for it — it is running in the frame below.

This is the same page the provenance block links to: our reference build for AP Copilot, with its own look, copy and motion. Scroll inside it, or open it full-screen.

// more from the lab

A team that ships across the board.

36products in production
8industries served
68+technologies in play

Closest to AP Copilot

// next

Building something in this space?

A 30-minute call is enough for us to understand your situation and tell you whether we're the right team. Usable version in the first two weeks if we are.