5 Years Later: How Generative AI Could (or Couldn’t) Have Helped My Chief Data Officer Work

Five years ago, I wrote a blog post on my first three years as Chief Data Officer for the City of Syracuse. It’s wild to think about — it feels both like a distant memory and just yesterday. A monumental shift since then? The explosion of accessible Generative AI technologies.

Given this technological leap, my continued professional journey, and my unwavering interest in data and tech within local government, I’ve been pondering how things might have unfolded differently had we wielded today’s AI tools back then. More than just a thought experiment, this is an opportunity for me to re-examine some of the specific use cases we tackled, and explore alternative approaches using these new technologies. So, let’s revisit some key themes from my original post and explore how generative AI could have been a game-changer, or perhaps, a non-factor.

I’ll mirror the original structure of my post, dissecting the challenges and then delving into how these technologies might have influenced our work, and hopefully sparking a lively conversation in the process!

I. Counting Things: More Complicated Than It Seems

My initial point was simple: “counting things is hard.” It should be easy, right? But in practice, it’s often a complex undertaking. It hinges on crystal-clear definitions, robust data collection systems, and the willingness of people to… well, actually count things. Ultimately, you rely on staff, developers, and users to provide the “how” — and that’s rarely straightforward.

Interestingly, this isn’t a uniquely local government problem, as some might assume. I’ve seen the very same hurdles when consulting with large corporations across various industries.

So, how could generative AI have potentially streamlined these challenges? Let’s break it down with some real-world examples:

Potholes

Original Challenge: Counting filled potholes was surprisingly arduous. Crews weren’t prioritizing documentation, and the existing systems were clunky. We tried spreadsheets, an app (which wasn’t well-adopted), and finally settled on truck-mounted sensors to log locations automatically.

How GenAI Could Have Helped:

• Image Recognition Power-Up: Back then, image recognition for potholes was in its infancy. Shadows and snow were constant hurdles. Today, a GenAI model, trained on a robust dataset of pothole imagery, could easily analyze street views, photos from crews and say, “Yup, that’s a pothole!” Visual confirmation would be automatic with an image captured by the truck sensor at each fill.
• Voice-Activated Location Logging: Forget clunky apps. A driver could simply say, “I’m filling a pothole now,” and a tool like ChatGPT Voice Mode or a similar model could automatically log the location, time, and pothole using GPS data. This would remove friction and allow the crew to keep their eyes on the road.
• Intelligent Data Cleaning: The sensor data had its quirks. AI could have swiftly identified and cleaned patterns in this data, distinguishing between a “legit” fill versus a multi-step process, ultimately improving data quality and accuracy.
• Coding Sidekick: Building tools solo was a major lift. Generative AI could have been my coding sidekick, significantly accelerating the prototyping process and allowing me to test different approaches more rapidly.

Limitations: GenAI is not magic. It wouldn’t have fixed poor sensor data, nor would it have fixed clunky tech. It’s a powerful tool, but it still requires reliable inputs and robust tech foundations.

Discussion: Would these AI-driven methods have resulted in higher accuracy? Faster turnaround? Lower costs? I’d love to hear other’s thoughts on how they’ve used tech and GenAI to improve counting practices.

Sidewalk Curb Corners

Original Challenge: Gathering data on sidewalk curb corners for ADA compliance was a massive undertaking. Walking and surveying the city was unrealistic, so we relied on Google Maps/Street View and a team of student interns to manually collect the data.

How GenAI Could Have Helped:

• Automated Image Analysis: Today’s GenAI image analysis is incredibly powerful. It could readily analyze Google Maps images, classifying curb corners based on features like color changes, raised bumps, and ramp presence.
• Natural Language Classification: Instead of manual categorization, interns could describe the corner using natural language, and a GenAI model could determine its compliance based on that description, removing a degree of interpretation from the equation.
• Data Validation Powerhouse: Because we were relying on student interns, we had concerns about data inconsistencies. GenAI could have been a diligent data quality checker, identifying errors and flagging problematic entries.
• Coding Support: Developing maps and dashboards to visualize this data was time-consuming. A coding partner, powered by GenAI, would have streamlined the process considerably.

Limitations: GenAI models require training data, and might not be pre-trained on specific ADA standards. Also, even with AI, data quality and reliability are not guaranteed.

Discussion: Could we have deployed the students more effectively, allowing them to focus on higher-level analysis instead of manual data entry?

Overall, I believe generative AI has the potential to transform how we approach “counting.” It could even assist in establishing the “how” by providing automation that cuts to the goal much more quickly.

II. Trends: What’s Really Going On?

Once we’re confident in our counts, we can begin analyzing trends. Are we seeing more of ‘X’ this month than last?

Road Ratings

• Original Challenge: Our road rating data revealed that our roads were in rough shape. However, a major obstacle was human bias in ratings. Staff sometimes hesitated to drop a road rating below “6”, even when it warranted a “5” or lower.

How GenAI Could Have Helped:

• Anomaly Detection: GenAI models could have scrutinized historical data, pinpointing anomalies and unusual rating patterns, making it easier to identify rating bias and areas for closer inspection.
• Hypothesis Generator: AI could generate hypotheses around why a road might be deteriorating. For example, was it due to a harsh winter? A specific type of traffic? This would have focused our investigation on actionable information, rather than speculation.
• Granular Image Ratings: Instead of the 1–10 scale, an image recognition GenAI could have created much more detailed ratings, going beyond a standard number rating system by examining the road surface in a granular way.

Limitations: Subjectivity will always be a factor in data collection, even with AI.

Discussion: How might we have detected and corrected the bias more swiftly, especially with a new tool in our arsenal?

Performance Dashboard

Original Challenge: We wanted to track code violation compliance, but this was complicated by increased proactive enforcement. It was tough to discern whether compliance was genuinely improving or if we were simply uncovering more violations due to the proactive effort.

How GenAI Could Have Helped:

• Causal Inference Power: AI could have been used to analyze how the proactive effort was influencing our results.
• Predictive Insights: GenAI could have helped build predictive models to show the impact of both proactive and reactive efforts on overall compliance, giving us a more nuanced picture of performance.
• Visualizing Insights: AI could have generated improved dashboards and visualizations to communicate this data in a more clear manner.
• System Integration: GenAI might have also helped us bridge the gaps between our legacy systems, which were difficult to work with.

Limitations: Data quality is crucial, and AI doesn’t magically fix dirty data.

Discussion: How can we move beyond just identifying trends to exploring the why behind them?

III. Prediction: Looking to the Future

Armed with an understanding of trends, we can then venture into prediction — what might happen next?

Water Main Risk

Original Challenge: Predicting water main breaks was tough due to incomplete data, institutional knowledge residing in the minds of veteran staff nearing retirement, and a lack of coordination across city departments.

How GenAI Could Have Helped:

• Data Gap Identifier: GenAI could have identified gaps in our data sets as we were building models, and even helped generate synthetic data to plug those gaps and reduce bias.
• Feature Engineering Assistant: Generative AI could have made it much easier to experiment with different approaches for model building, and do it quickly.
• Institutional Knowledge Extraction: We could have analyzed interviews with veteran staff to extract their tacit knowledge and structure it for wider use within the city, ensuring it wasn’t lost.
• Model Enhancement: The original model only focused on whether a pipe broke, and not how. Generative AI may have helped build a more complex model including additional variables.
• Research Assistant: GenAI could have served as a research agent, finding relevant research articles and other information that may have helped, and summarizing it in ways that were relevant to our work.
• Image to Data: We had old engineering books filled with hand drawn diagrams; GenAI could have been used to convert this data into a structured format for model inclusion.

Limitations: AI is only as good as the data you give it, and our century-old water system had limited historical data.

Discussion: How could AI help us build models faster and with a more complete picture of the situation?

IV. Partner: Together We’re Stronger

With limited resources, collaboration was essential.

Hackathons

Original Approach: A way to engage the community and generate solutions.

How GenAI Could Have Helped:

• Idea and Prototype Accelerator: GenAI could have helped participants come up with ideas faster and create early prototypes more quickly, reducing barriers for non-technical participants and making the events more inclusive.
• Synthetic Data Creation: We could have generated synthetic datasets to provide a starting point.
• Community Engagement: We could have created chatbots for participants or provided translation services for non-English speakers.

Limitations: GenAI is an enhancement tool, but not a substitute for real-world engagement and in-person connection.

Discussion: Could GenAI have made our hackathons more productive and welcoming to diverse skillsets?

Student Consultants/Class Projects

Original Approach: Leveraging university resources.

How GenAI Could Have Helped:

• Project Scoping: GenAI could have helped to create clear, well-defined scopes for class projects, and more importantly, help the students understand the business problem.
• Coding Support: GenAI could have served as a coding assistant for students who weren’t as skilled at coding, allowing them to experiment and test their ideas more quickly.
• Research Resource: GenAI could help the students find relevant resources and articles for their projects.
• Rapid Feedback: GenAI could have provided faster feedback, enabling faster iteration.

Limitations: GenAI should be seen as a tool to augment learning, not replace the need for human engagement.

Discussion: Could GenAI have elevated the quality and expedited the pace of student-led projects?

Local Government Organizations/Internal Partnerships

Original Approach: Collaboration across public/private sectors, and across city departments.

How GenAI Could Have Helped:

• Knowledge Hubs: GenAI could have been used to document institutional knowledge across departments and build databases that allow for information sharing.
• Better Communication: GenAI could have summarized documents or drafted communications, keeping stakeholders informed.

Limitations: GenAI isn’t a substitute for human connection and collaboration.

Discussion: How can we foster stronger public and private partnerships through shared knowledge and improved communication?

V. Being Open and Transparent: Access For All

Original Approach: A core goal was to make city data public and available.

How GenAI Could Have Helped:

• Data Summaries: GenAI could have summarized complex data sets in ways that are easier to understand, making it more accessible to all.
• Accessibility Power-Up: AI could have translated data into multiple languages and formats to improve its accessibility.
• Data Quality Checks: GenAI could have helped identify and resolve data quality issues before publishing.
• Privacy Review: Generative AI could help us to flag and review potential privacy risks, making data disclosure more responsible.
• Automation: GenAI could have streamlined the code needed to automatically publish data to open data portals, lowering the time and energy needed to create a portal.

Limitations: GenAI is only useful if people are interested in data, and possess a foundational understanding of data principles.

Discussion: How can GenAI be used to make open data more impactful, accessible, and responsible?

Conclusion

Generative AI isn’t a magic wand that would have solved everything, but it could have been an incredibly valuable partner in many areas, acting as a coding assistant, project manager, or data synthesizer. It could have also helped us unlock the value of non-machine-readable data, as well as prompt us to think through strategic goals. Furthermore, the potential Return on Investment (ROI) and value derived from these AI-powered approaches are worth considering:

• Increased Efficiency: Metrics like reduction in time spent on manual data entry, faster project turnaround times, and decreased time to insight could demonstrate significant gains in operational efficiency.
• Improved Accuracy: Measuring the reduction in errors and data inconsistencies could provide data points for better and more reliable information.
• Cost Savings: Tracking metrics like reduced labor costs, optimization of resource allocation, and increased productivity could paint a clearer picture of potential cost-savings.
• Enhanced Citizen Engagement: Measuring the impact of improved accessibility through metrics like increased website traffic on open data portals and greater community participation in hackathons.
• Better Decision Making: Measuring improved accuracy in predictions, identification of hidden trends, and faster detection of problems could provide a clear ROI on implementing the tools.

We still must partner with humans and engage with the community. This technology is designed to amplify and enhance what we do, not replace it. The advent of generative AI expands the range of challenges a Chief Data Officer needs to consider: leveraging unstructured data, selecting the right models, understanding cost implications, and more.

The core functions of the job — counting things, analyzing trends, making predictions, partnering with others, and being open and transparent — remain paramount. Generative AI has the ability to smooth the process, but the work itself is never-ending.

I continue to believe that being a Chief Data Officer in municipal government is an incredibly rewarding role. It’s exciting to see how others are using these tools.

I’m eager to hear your feedback! What are your experiences, thoughts, or questions after reading this post?