# HCAD NLP Pilot: 40% Mean Review Time Cut, Not a Guarantee

Natalie Fletcher · August 17, 2026

> HCAD NLP Pilot: 40% Mean Review Time Cut, Not a Guarantee. The NLP Pipeline HCAD’s 2025 pilot logged a 40% reduction in administrative review time, bu...

## The NLP Pipeline

HCAD’s 2025 pilot logged a 40% reduction in administrative review time, but that headline number obscures the actual engineering: the gain is entirely a function of where the NLP pipeline routes each document. The system is not a valuation engine; it is a triage mechanism that separates machine-readable evidence from everything else, and it only works because the model was trained on HCAD protest files from 2022–2024, not on generic real-estate text. According to HCAD’s internal time-tracking logs from the pilot, the median time from filing to first human touch dropped from 18 minutes to 10.8 minutes per case—but that figure assumes the evidence is a clean PDF or TIFF at 300 DPI or higher. The moment a document falls below that threshold, the clock resets.

The pipeline begins with a fine-tuned BERT-based model that extracts three feature classes from each submission: property characteristics (square footage, year built, condition grade), evidence type (appraisal report, repair estimate, tax statement), and comparable sales. The model was trained on HCAD-specific protest files because the district’s forms and appraisal review board language have idiosyncratic formatting that a generic legal NLP model misreads. For example, HCAD’s condition grade codes (C1–C6) appear in a dedicated field on the protest form but are often buried in the remarks section of a third-party appraisal; the fine-tuned model learns to look in both places. OCR preprocessing with Tesseract 5.0 handles scanned documents, but only under strict conditions: resolution must be ≥300 DPI and the text must not be handwritten. Handwritten evidence—common in owner-occupied protests where the homeowner annotates a printout—is routed directly to human review, bypassing the NLP layer entirely. This is a deliberate design choice, not a limitation: the model’s confidence score on handwriting is unreliable, and a false extraction on a handwritten repair estimate would poison the entire comparable-selection process.

Once the model extracts the property’s features, it compares them against a dynamic database of 1.4 million HCAD sales records, generating a similarity score for each potential comparable. The similarity score is a weighted Euclidean distance over the three extracted features, with condition grade weighted most heavily because it is the strongest predictor of value variance in Harris County’s residential market. The model then assigns a confidence score (0–1) to the entire extraction—not just the property features, but the evidence classification and the comparable matches as a package. Scores below 0.85 trigger a manual review queue; scores above 0.85 auto-populate a protest summary that includes the top 3 comparable sales and a suggested value range. The 0.85 threshold is not arbitrary: it was calibrated during the 2025 pilot to balance recall against the cost of human review, and it is the canonical decision rule for the entire system. A score of 0.84 is functionally identical to a score of 0.30—both go to a human, and the human sees the raw extraction alongside the confidence score so they know which parts of the summary to distrust.

The infrastructure matters as much as the model. The pipeline runs on AWS Lambda with a 2-second latency per document, allowing batch processing of 500 appeals per hour without additional staff. This is what makes the 40% reduction operationally real: HCAD’s review team was previously opening each PDF manually, scanning for the property’s square footage, and cross-referencing it against a static sales table. Now the system does that in two seconds, and the human reviewer only sees cases where the model is uncertain. The 9% median absolute error rate for automated valuation models, cited in the peer-reviewed research on big data in real estate, is a useful benchmark here—it tells you the ceiling for automated comparables, but it does not tell you how to handle the 15% of cases where the model’s confidence falls below 0.85. Those cases are where the licensed appraiser earns their fee, and the pipeline is designed to surface them early rather than let them slip through with a false sense of precision.

| Pipeline Stage | Input Condition | Output | Time Cost |
| --- | --- | --- | --- |
| OCR preprocessing (Tesseract 5.0) | ≥300 DPI, typed text | Machine-readable text layer | ~0.5 sec |
| BERT extraction | Clean text layer | Property features, evidence type, comparables | ~1.5 sec |
| Comparable matching | Extracted features | Similarity scores vs. 1.4M HCAD sales records | ~0.5 sec |
| Confidence scoring | Full extraction package | Score 0–1; ≥0.85 auto-populates summary | ~0.2 sec |
| Manual review queue | Score

Canonical: https://lawr.io/blog/hcad-nlp-pilot-40-mean-review-time-cut-not-a-guarantee.php
Markdown: https://lawr.io/blog/hcad-nlp-pilot-40-mean-review-time-cut-not-a-guarantee.php/index.md