How HDR Scaled Its SSES Program Without Scaling Staff

About HDR

HDR is one of the largest employee-owned engineering and architecture firms in North America, with deep expertise in water and wastewater infrastructure. HDR's Utility Management practice helps municipalities design and execute sewer condition assessment programs, managing the full workflow from field data collection through capital planning and rehabilitation design.

As part of the ongoing Sanitary Sewer Evaluation Survey (SSES), HDR manages 30 to 40 miles of contractor-collected CCTV data annually for utility clients. That data flows through a rigorous engineering review process before informing Preliminary Engineering Reports (PERs) and capital improvement programs.

The Challenge

HDR's sewer assessment practice faced three compounding problems that 100% manual human review alone couldn't solve.

Fragmented, Inconsistent Data

CCTV data routinely arrived across 60+ NASSCO exchange databases, collected by multiple contractors under different specifications with no shared naming conventions. Identifying duplicate videos, reconciling inconsistencies, and standardizing outputs before engineering review consumed significant capacity before any coding work began.

Selective Coding by Field Crews

Internal utility CCTV crews focused on the most serious structural defects, the Grade 4s and 5s. Minor structural issues, early-stage O&M defects, and nuanced conditions were frequently missed or undercoded. The result was a dataset biased toward severity rather than completeness, an incomplete picture feeding directly into capital prioritization models.

Unsustainable 100% Review Burden

HDR's standard workflow required 100% engineer review of all contractor CCTV footage before submission to the utility and development of the PER. With 30 to 40 miles of inspections per year and a growing backlog, the model was reaching its limits. There was no way to scale throughput without proportionally scaling staff.

The Solution

Beginning in 2023, HDR partnered with SewerAI to evaluate AutoCode, SewerAI’s AI-powered NASSCO PACP coding platform, through a two-phase study centered on two key questions:

1. How accurately does AutoCode code CCTV mainline data?
2. How does AI-coded data impact capital planning outcomes compared to human-coded data?

Phase 1 — Accuracy Validation

HDR processed 100,000 linear feet of existing CCTV data through AutoCode, then manually reviewed approximately 15% of that footage,across 64 inspections and 896 individual observations, covering a broad range of pipe sizes, materials, and defect types.

In parallel, HDR conducted a capital comparison study across 335 assets totaling 58,000 linear feet. Both the AutoCode output and manual-coded output were independently reviewed for accuracy, then processed through a rehabilitation decision algorithm to compare capital planning outcomes.

HDR's independent accuracy audit confirmed AutoCode met their engineering standard, giving them the confidence to deploy on live SSES projects.

• 87.95% codes required no modification

• 2.01% codes required only observation detail changes (clock position, % cross-section)

• 0.56% codes with minor modifier changes, with no impact on structural or condition grade

Only 9.6% of codes were incorrect, and critically, less than 1% of all codes involved defects that would actually change a rehabilitation recommendation.

Phase 2 — Live Deployment

Based on Phase 1 results, HDR deployed AutoCode into its active SSES workflow, shifting from 100% engineer review to a 20% targeted QC model. Uncoded contractor CCTV was submitted directly to AutoCode; human review was focused only on AI-flagged windows of interest, clarifying the operational shift in review effort.

By the CS&SW 2024 Conference, Phase 2 had processed over 190,000 linear feet, with 100,000 additional linear feet in the pipeline and a goal of 200,000 linear feet in the following fiscal year, showing the scale of operational throughput.

Results

80% Reduction in CCTV Review Effort

The shift from 100% engineer review to 20% targeted QC review reduced HDR's review burden by 80%. Engineers stopped watching every frame and focused on AI-flagged windows of interest, where their expertise added the most value.

21% More Defects Identified

AutoCode identified 21% more defects than field-coded contractor data, and 15% more than QC'd engineer-reviewed data across the same inspection set. The greatest differences appeared in Structural (+28.6% vs. field crews) and O&M (+28.8% vs. field crews), the defect families that drive rehabilitation urgency and cost.

$1.82M in Hidden Rehabilitation Costs Surfaced

Because AutoCode identified more defects without increasing the max defect severity, the structural risk score's defect count component increased, shifting more pipes from point-repair territory into manhole-to-manhole rehabilitation territory.

When rehabilitation costs were calculated for both datasets:

• $664,100 more in rehab costs identified in the conventional CCTV dataset
• $1,160,000 more in rehab costs identified in the 360-degree video dataset

This wasn't an over-recommendation. It was evidence that human coding had been systematically under-identifying defects, and capital programs built on that data underestimated what the infrastructure actually required.

Consistent, Complete Data Across All Sources

AutoCode processed data from 60+ NASSCO databases, multiple contractor crews, and inconsistent video naming conventions, producing a unified, consistently coded output. It flagged duplicate videos regardless of Video ID and applied consistent media naming based on uploaded specifications. Legacy inspections previously coded to outdated standards were brought into compliance with current NASSCO standards.

Accurate Grade Distribution

Historically, field crews produced skewed grade distributions, with a strong bias toward Grades 4 and 5. AutoCode restored the full spectrum of NASSCO grades, giving engineers and planners accurate risk profiles across the network rather than a list of only the most urgent assets.

"Using AI significantly reduces the amount of human review needed on the CCTV data. AutoCode consistently identifies more defects than our field crews, and less than 1% of incorrect codes were actually decision defects. That's a level of accuracy we can build capital programs on."

— Kelly Alexander, PE, Utility Management Practice, HDR