Killing the Inbox: How Centrix Reads Every Invoice Email, Extracts the PDF, and Posts to A/P Automatically

The least-loved hour of the accounting week

Every Friday afternoon, somewhere in your back office, a person is opening emails one at a time, downloading PDF attachments, typing invoice numbers, matching to purchase orders, coding to GL accounts, and clicking save. The work is correct, important, and expensive. The work is also ~70-80% pattern recognition that a model can do reliably.

The reason it doesn't already get automated at most carriers isn't lack of tooling — it's that the tooling is sold as expensive A/P automation suites that require IT integration projects. Centrix builds it on top of components the platform already has: an email scanner, a PDF extractor, an ML classifier, and the existing accounting workflow.

What the workflow looks like

A vendor — Pilot Travel Centers, Comdata, Shopmonkey, an insurance carrier, a tire shop — emails an invoice to the carrier's A/P address. From there:

1. Email scanner reads the inbox

The `email_scanner` service watches the A/P mailbox via IMAP. New messages are parsed: from, subject, body, and attachments. The metadata feeds into the `email_rules` engine.

2. Email rules route by sender / subject pattern

A rule fires when the message matches a known vendor pattern. Pilot's "Statement for Account #..." subject routes to the fuel-statement workflow. Shopmonkey's "Invoice INV-..." routes to the maintenance workflow. An unknown sender routes to a manual-review queue with a draft classification.

3. PDF extractor parses the attachment

The `invoice_extractor` service uses pdfplumber to pull structured data from the PDF: invoice number, date, vendor name, line items with amounts, totals, and any matching reference numbers (load number, truck VIN, driver name). The extractor handles the top 30 vendor formats out of the box and learns new formats with two labeled examples.

4. GL classifier proposes coding

The line items are classified to GL accounts: fuel to fuel COGS, oil/grease to truck-specific PM, tires to tire account, etc. The classifier uses prior coding decisions from the same carrier as training signal, so within 60 days on a new tenant the coding accuracy stabilizes at 95%+.

5. Match-to-PO (where applicable)

For maintenance and parts invoices, Centrix matches the invoice to the existing PO (issued through Shopmonkey or directly in Centrix). Variance exceeding the carrier's tolerance threshold (default 5%) routes to manual review. Within tolerance auto-approves.

6. Posting and audit trail

Approved invoices post to the A/P ledger with the original email message, PDF, and extraction confidence score attached. Every row carries a hash-chain entry so the audit history is tamper-evident.

The accounting role sees the queue: green-banded rows are auto-posted (just a record), yellow-banded need review (variance, low-confidence extraction, new vendor), red-banded are blocked (DLP-flagged, missing data, suspicious duplicate).

What this saves

On a 100-truck fleet with a typical vendor mix:

• 600 to 900 vendor invoices per month
• 8 to 14 minutes per invoice on the legacy manual workflow
• 80 to 200 hours of A/P data entry per month — call it $4,000 to $10,000

monthly fully-loaded labor cost

Centrix's automation cuts ~75% of this on stable vendor patterns. Net savings: $36K to $90K/year per 100 trucks. Plus the qualitative wins — cleaner vendor relationships (no "did you get that invoice?" chase), faster payment cycles, and audit-ready documentation chains.

The duplicate-invoice catch

A subtle but valuable side benefit: the extractor flags duplicate invoices across vendors. A common error is paying the same Pilot statement twice — once via the auto-deducted Comdata pull and once via the emailed PDF. The duplicate check cross-references invoice numbers and amounts across the account history, surfacing duplicates before the second payment posts.

Median duplicate-payment recovery on managed fleets: $8K to $24K/year. Not life-changing money but real money the back office didn't know was there.

What it doesn't replace

The system isn't designed to replace the controller. It's designed to remove the 70-80% of the controller's data entry so the controller can focus on:

• Variance investigation (why did fuel come in $4K over forecast?)
• Vendor negotiation (which contracts are due for review?)
• Cash management (which payments delay vs accelerate?)
• Audit preparation (year-end / quarterly close)

These are all higher-value uses of accounting time than typing invoice numbers from PDFs.

Where it fits with QuickBooks

If you run QuickBooks (Online or Desktop), Centrix's invoice automation sits in front of QB, not in place of it. Approved invoices flow into QB via the `qb_sync` service, with the original PDF and email message attached as QB attachments. The controller works in QuickBooks the way they always have; the data entry is just done.

For carriers running a different GL (NetSuite, Sage Intacct, Microsoft Dynamics), Centrix exports approved invoices in standard formats (CSV, XML, JSON). Direct connectors are case-by-case.

Where to start

If you're a 30+ truck fleet with one or more A/P clerks:

• Connect the A/P inbox first. Read-only at first — the system parses but

doesn't post. Compare extraction output against actual coding for 30 days. Most carriers see 92-96% accuracy on first 30 days.

• Turn on auto-post for the top 5 vendors (usually fuel, maintenance,

tires, insurance, and one or two others). These are the highest-volume, most-pattern-friendly cases. Watch the variance for 60 days.

• Expand to the long tail. By month 4, the classifier has enough labeled

examples to handle 90%+ of the vendor base.

Book an A/P automation review — bring three months of inbound vendor invoices and we'll show you the extraction confidence and projected automation rate.