"Our goal was to speed up our manual document review, but after a month, we realized our team never made any edits. We actually removed the human from the loop entirely."

— Mike Abner, CTO, HomeLight

About HomeLight

HomeLight is a real estate platform, powering over $1 billion in transactions annually, that helps people buy and sell homes more efficiently. One of their flagship offerings, Buy Before You Sell, lets homeowners purchase their next home before selling their current one. It’s a magical experience for their users, but behind the scenes, it involves complex contracts and heavy operational workflows.

Manual Review Bottleneck

Each real estate transaction involves contracts — 150+ page counteroffers or closing packets — that HomeLight’s operations team needed to manually review. Every file took about 20+ minutes of human time to process. More deals meant more documents, and more documents meant more humans.

The team had a clear target in mind: grow this product line without growing headcount linearly.

The Build-or-Buy Decision

After an initial prototype using OpenAI’s APIs, they reached ~85% accuracy extracting key fields from complex contracts — promising, but not good enough to remove humans from the loop.

“We thought we could eke out a few more % points from our prototype, but not enough to hit our goal of eliminating human oversight entirely.”

— Mike Abner

To close that gap, they realized they’d need more than just a model call. They’d need to:

• Build human review tooling to catch low-confidence extractions

• Show bounding boxes or citation trails to operators

• Create feedback loops and retraining pipelines to improve accuracy as the business evolved

“The level of effort to go from 85% to 99% is way more investment than you think it's going to be. There’s a ton of infrastructure around the actual processing of documents that you have to build.”

— Mike Abner

The team started exploring alternatives, and that’s when they came across Extend.

“It kind of all clicked, we got two tools for the price of one. More accurate document extraction, and a human-in-the-loop system. It was exactly what we thought we’d have to build, but it was already working and way more polished.”

— Mike Abner

On day one, they immediately gained access to:

• A state-of-the-art pipeline combining OCR, VLMs, and an advanced LLM engine to power document extraction, splitting, and classification

• A suite of configurable tools including semantic chunking, an advanced citations system, and battle-tested heuristics

• A robust human-in-the-loop system for review and corrections

• A UI for non-technical PMs and operators to configure, review, test, and iterate on data schemas

• An ergonomic API for their developers to plug into their existing product stack

HomeLight went live with their first use case — extracting 35+ fields from purchase contracts — in just two weeks.

Results

The impact was immediate:

• Accuracy jumped from ~85% to 93%+ in early tests; accuracy now exceeds 98%

• Manual review of documents was turned off entirely after a full month of 0 corrections

• PMs and operators own schema iteration and testing, freeing up engineers to focus on the data post-processing experience downstream

“We got two tools for the price of one. More accurate document extraction, and a human-in-the-loop system we didn’t have to build or maintain.”

— Mike Abner

HomeLight has since shipped 2 additional new use cases into production because of how easy Extend makes achieving high accuracy possible. Most importantly, HomeLight’s team was able to scale their operations workflows without scaling human review, meaning more families getting into homes, faster.

“Our goal was to go from 20 minutes down to 3 or 4. In the end, we got all the way to zero. After a month, we never made any changes. We actually removed the human in the loop.”

— Mike Abner

Engineering Impact

By shifting schema setup, evaluations, and pipelines into Extend, HomeLight’s engineering team now works in tandem with product managers and operators to own the full document lifecycle, from testing new formats to deploying updated configs.

“PMs and Ops are able to handle the configuration, prompt engineering, and evaluation. And engineers are freed up to now focus on how we can use the data downstream to build experiences for our users.”

— Mike Abner

That shift unlocked broader use cases and faster experimentation, freeing up months of engineering resources dedicated to core product development.

Looking Ahead

HomeLight now processes thousands of documents per month through Extend. With each new format or contract type, PMs and operators can onboard new workflows directly without creating tickets or fighting for engineering bandwidth.

More importantly, they avoided the infrastructure burden of building and maintaining document processing in-house, achieving 98%+ accuracy in days or weeks, not months.