How Oklahoma Department of Transportation got 30 years of inspection records ready for AI

Oklahoma has almost 23,000 bridges, which means it would take a mere human quite a while to answer questions like:

• How many bridges have been considered structurally deficient in the last five years?
• Which county has the most bridges?
• Which bridges have shown the biggest decline in structural integrity over the last decade?

With the help of Google Cloud partner North Highland, the Oklahoma Department of Transportation (ODOT) prepared 30 years of bridge-inspection data for an AI platform that can help public sector employees answer important infrastructure questions.

Prepping bridge report data required a governance effort, a cleanup of inconsistent formats, and human review.

Why do this? Infrastructure information existed in silos, according to ODOT Enterprise Systems and Services Director Lance Underwood. “Some of these questions just were taking a little bit too long to get the answer from the right group,” he told IT Brew.

Oklahoma had decades of systemic bridge inspection data—reports featuring IDs, locations, and condition scores. North Highland helped the government agency get that into an AI-ready form.

“Most organizations kind of have the current state, but to build a historical value, we need to be able to recreate that information in a way that’s standardized,” Doug Krauss, senior director of AI and data infrastructure at North Highland, said.

Over the course of about six months, Krauss and Underwood said, the North Highland technical team and ODOT bridge team went through these preparation steps: 

• Data governance: Data owners made decisions on standardized definitions for the reports—definitions like what determines a structurally deficient or at-risk bridge based on a bridge classification guideline. The data owners act as the “linchpin for communicating any changes within their data set,” Underwood said. A business glossary of definitions and calculation logic was then paired to physical data elements in database tables, Krauss wrote in a follow-up email.
• Cleanup: The US Department of Transportation Federal Highway Administration’s Bridges & Structures bridge inspections, which date back to 1992, act as a primary historical data source, Krauss shared in a post-interview reply. What can change in 30 years of reporting? Boundaries like district outlines get altered, and bridge IDs may move from alphanumeric to integer, for example. Teams leveraged processing tools to reduce manual effort and to carefully remap data over time to ensure consistent formatting for the LLM to read.
• Add guidelines: North Highland even fed the official “Recording and Coding Guide for the Structure Inventory and Appraisal of the Nation’s Bridges” to Google large language models Vertex and Gemini, according to Krauss’s follow-up email. The process extracted essential definitions, diagrams, and technical guidelines for the data catalog. After business-user checks, the metadata was attached to the core historical data, which the LLM uses as context. North Highland and ODOT staff also reviewed accuracy of the resulting dashboards and reports, Krauss wrote.
• Load it up: Project leaders used Google Cloud’s Dataplex cataloging tool and its centralized data warehouse BigQuery. Managers keep the data lake incrementally up to date throughout the year, informed by regular meetings from data-governance and steering committees, Underwood said.

The high stakes of an incorrect decision—nobody wants a structure to collapse on them—mean that AI in this context must be implemented as “decision support, not decision replacement,” according to Underwood.

In a follow-up email to IT Brew, he wrote: “We should never allow a model to be the sole basis for a safety-critical conclusion. The right approach is governed data, transparent models, confidence scoring, exception thresholds, human engineering review, and continuous validation against real-world outcomes. In other words, AI helps us identify risk sooner and prioritize better, but final safety decisions remain with qualified professionals.”

Who else is in? ODOT is one of many organizations trying to bring structured data to LLMs, with some groups more prepared than others. A Harvard Business Review survey cited in a report by AI platform Cloudera found that just 7% of respondents (those “involved in their organization’s data decisions” actively considering AI use for business purposes) said their data was “completely ready” for AI adoption.

CJ Combs, senior AI consultant at consultancy Columbus, spoke with us in January about his conversations with companies beginning their automation transformation.

“We start asking questions, not about, ‘What are your AI ideas?’ It’s like, ‘Where is your data? How are you managing that?’” he said.

Now ODOT is moving beyond bridge data sets to other infrastructure report repositories.

“I think eventually that is the goal, is to have as many of the production data sets into this one data lake as possible, and that you can start having that cross-pollination from a bridge data set and construction data set, traffic data set. You can use all of the different data sets to ask those natural-language questions,” Underwood said.