AWS anchors its document-processing accelerator to an $11,000 loan and a 44-day close

AWS has published a walkthrough for automating mortgage document intake using two pieces: the Generative AI Innovation Center's Intelligent Document Processing Accelerator and Amazon Quick Automate ^[1]. What makes it worth reading is not the architecture but the arithmetic it hands to buyers, because the ROI framing rests on third-party industry figures that any internal business case will then be measured against ^[3][4].

The cited numbers: the Mortgage Bankers Association puts the total cost to originate a single loan at over $11,000, spanning sales, fulfillment, production support and overhead ^[4]. ICE Mortgage Technology's Origination Insight Report puts the average close at 44 days ^[3]. The MBA Mortgage Finance Forecast has the US market at roughly 4 to 6 million loans a year, though the post prints that as "$4-6 million loans per year" ^[2]. None of these are AWS's own measurements, and none are decomposed. The post says document intake and processing make up a significant part of the fulfillment workload without quantifying how much ^[5].

Then the example lender. Summit Mortgage is fictional and AWS says so, describing the metrics as illustrative and indicative rather than observed ^[6]. Summit runs about 50,000 loans a year, with processors spending 15 to 20 minutes per file sorting documents, verifying completeness and keying data into the loan origination system ^[7], which AWS totals at more than 15,000 hours a year ^[8]. Check the multiplication: 50,000 files at 15 minutes is 12,500 hours, at 20 minutes it is 16,667 hours ^[1]. The 15,000-hour figure therefore assumes roughly an 18-minute average, near the top of the stated range ^[2]. The stated goal is under 6 minutes per file, with fewer keying errors and peak volume absorbed without added headcount ^[9]. The previous answer to peaks was temporary staffing, which AWS describes as expensive, slow to onboard and inconsistent in quality ^[12].

This is where the framing and the mechanism come apart. Going from 18 minutes to under 6 saves about 12 minutes per file, or roughly 10,000 hours a year at 50,000 loans ^[3]. Set against a 44-day close, 12 minutes is about 0.02 percent of elapsed cycle time ^[4]. Touch time and cycle time are different clocks. If document work moves the 44 days at all, it does so through the two failure modes AWS names in passing: manual keying errors that trigger downstream rework, and incomplete packages that sit in queues until someone notices what is missing ^[14]. Those are queue effects, not minutes-per-file effects, and the post does not size them.

The $11,000 has the same problem in reverse. At Summit's volume that is about $550 million of annual origination cost ^[5], and the post does not say what share of it is document intake ^[5]. Anyone who carries the $11,000 into a steering committee is implicitly promising to move a number they cannot yet attribute.

The mechanics are conventional and stated clearly enough: an open source serverless pipeline, billed per document processed, using Amazon Textract and Amazon Bedrock foundation models ^[10] to convert documents to machine-readable text, classify them, extract fields, validate against expected schemas and flag anomalies for human review ^[11]. That last step is the honest part of the design, and it is where residual headcount lives.

Worth watching: whether AWS follows with measured customer results rather than a fictional lender ^[6]; what per-document billing actually totals at 50,000-package volumes ^[10]; and the exception rate, since a pipeline flagging one file in five for review has a very different labour profile from one flagging one in fifty ^[11]. Also whether the same accelerator survives the move into banking, insurance, healthcare and the public sector that AWS says it adapts to ^[13].