Automating construction invoice processing with Claude
A construction company with about $30M in annual revenue was processing roughly 500 vendor invoices a month by hand: open the email, read the attachment, find the invoice details, type them into QuickBooks. At about five minutes each, that was more than 41 hours a month spent moving information from one system to another. Anvil AI Systems put Claude between the inbox and QuickBooks so the reading and typing happen automatically and people handle only the exceptions.
Client
A $30M construction company
Industry
Construction
Company size
$30M annual revenue
Volume
500 invoices a month
Built with
Claude by Anthropic
Vendor email
Claude
Validation
QuickBooks
90%Reduction in invoice processing time, from five minutes to about thirty seconds
95%Correct processing rate in production, on 500 invoices a month
450hrsEmployee hours returned every year, about $22,500 of capacity
90% faster invoice processing, at 95% accuracy, with QuickBooks still the system of record.
01The challenge
02The approach
03The results
01 · The challenge
Every invoice was read and retyped by a person.
Vendor invoices arrive by email all month, in whatever layout each supplier uses. For each one, an employee opened the message, inspected the attachment, worked out the vendor and invoice details, entered them into QuickBooks, and checked the entry against the document. Around 500 invoices a month at five minutes each came to 2,500 minutes, or 41.7 hours, of administrative work every month. The work was necessary, but almost none of it was accounting judgment. It was repetitive interpretation and data entry.
02 · The approach
Claude became the intelligence layer between email and QuickBooks.
The new workflow is vendor email, Claude, validation, QuickBooks. When an invoice arrives, the system finds the relevant document and asks Claude to interpret it, turning an unstructured invoice into structured accounting data: vendor, invoice number, invoice date, due date, amount, line items, and supporting details. Software validates the extraction before anything is written to QuickBooks. Invoices are not standardized, and rules-based automation breaks down as supplier layouts, terminology, and templates pile up. Claude reads the variety and returns a consistent shape, so the workflow moved from a person reading, interpreting, and entering every invoice to Claude reading and structuring, software validating, and a person handling exceptions. The accounting system itself did not change.
03 · The results
A process that took 500 employee hours a year now takes about 50, and the team reviews 25 exceptions a month instead of typing 500 invoices.
500 invoices a month processed through the new workflow
95% processed correctly in production, so about 475 invoices a month need no touch and about 25 are routed for review
Processing time down 90%, from about five minutes an invoice to about thirty seconds
37.5 employee hours returned every month, 450 a year
About $22,500 a year in labor capacity created at a $50 hourly administrative rate, before any gains from faster invoice visibility, fewer corrections, or duplicate detection
No change to the accounting software or to how employees already work; QuickBooks remains the system of record
KPI
Before
After
Impact
Monthly invoice volume
500
500
—
Processing time per invoice
5 min
~30 sec
90% reduction
Monthly processing time
41.7 hrs
4.2 hrs
37.5 hrs saved
Annual processing time
500 hrs
50 hrs
450 hrs saved
Correct automated processing
0%
95%
475 invoices a month
Annual labor requirement at $50/hr
$25,000
$2,500
$22,500 capacity created
500 invoices. 95% accuracy. 90% faster processing. Claude was inserted between two systems the company already used, so AI handles the interpretation and transcription while QuickBooks keeps the records and people keep the exceptions.