Imagine a Wednesday morning in your conference room. The city manager has pulled together a dozen people: the IT director, the head of community development, two permit technicians, a budget analyst, a public works supervisor, and a few others. City councilmembers have been asking about AI. The city manager’s question is simple: “Where should we start?”

By the end of the hour, the whiteboard is full of ambitious ideas (chatbots and predictive analytics opportunities abound!) but there is no agreement on which ones are realistic, affordable, or worth the risk.

This scene plays out across local governments everywhere, and it mirrors a pattern that researchers have documented across industries: any teams skip structured ideation, jump to pursuing the first idea that sounds exciting, and end up working on the wrong problem. It’s no surprise that most AI projects fail to deliver value.¹ 

Too often, the people who understand the work and the people who understand AI aren’t together in the same room and the resulting brainstorming runs the risk of producing either moonshot ideas that can’t be built or technically elegant solutions to problems nobody has.²

Researchers at the University of Virginia’s School of Data Science, in collaboration with Carnegie Mellon University, have worked to develop and test a method for innovating AI use cases. Their insights may change how you think about leveraging AI in your organization.

“Good Enough” May Be Exactly What You Need

“Good enough” isn’t typically a winning slogan, but when it comes to AI, that may not be too far off the mark from what you actually need. When researchers analyzed 40 commercially successful AI applications across 14 domains (healthcare, manufacturing, finance, government, and more) they found that 25 out  of 40 required only moderate performance to be useful.³

This isn’t really too different from how many of us are already using AI. Consider automatic speech recognition—we would likely agree that despite only running at about 90–95% accuracy (meaning voicemail transcripts/video captions can get a meaningful share of words wrong), it’s still much more useful than no transcript at all.⁴ 

Even so, when considering where to start with selecting an AI project, teams frequently gravitate toward the hardest problems, situations where AI would need to be nearly flawless (high risk and high effort). 

Another, more relevant example: Your planning department receives 200 permit applications per month, and each one needs to be read, classified by type, and routed to the right reviewer. A staff member spends roughly 15 minutes per application doing this (about 50 hours per month). Imagine an AI tool that classifies applications correctly 85% of the time. It misroutes about 30 applications per month that a staff member needs to redirect, but it handles the other 170 instantly. That’s 40+ hours per month back for higher-value work. 

When selecting the “right” AI project, keep in mind: AI doesn’t need to be perfect to achieve a significantly positive outcome. It just needs to be good enough.

Google trains its internal product teams using a vivid image, popularized by data scientist Cassie Kozyrkov: think of AI as an island of drunk people.⁵ There are a lot of them, so they can handle enormous volume and work fast, but they make mistakes, so don’t expect precision from any individual. 

A helpful reframe of the AI question to apply for your team: instead of “where do we need a genius?” try “where would a fast, imperfect workforce still be an improvement over what we have now?” Instead of asking, “where could AI make a perfect decision?” ask, “where are my employees spending hours on tasks that AI could do pretty well in minutes?” That question, applied across departments, is where the real opportunities live.⁶ 

Running an AI  Design Sprint in Your Organization

Let’s go back to that Wednesday morning and think about how it might look with the adoption of a research-informed AI Design Sprint model for identifying/designing AI applications for your organization. 

The AI Design Sprint is a structured workshop (1–4 hours) that brings together a cross-functional team to simultaneously consider what AI can do and what problems staff actually face. It combines two approaches that don’t work well alone: starting from user needs (which tends to surface problems AI can’t solve) and starting from technology capabilities (which tends to produce solutions nobody wants). The method lends
both — the researchers call it combining “user-centered design” with “matchmaking.” No technology purchase is required for this, and no data science expertise is needed to make use of this method. 

The output is a prioritized list of use cases. Technical specs come later, and with a much smaller investment, when you’re confident you’re solving the right problem.

Who Should Be in the Room

• 2–3 department heads or senior staff who know the operational pain points.
• 2–3 frontline employees who do the daily work.
• 1–2 people from IT/data who understand what systems and data exist.
• 1 person from budget/finance who can reality-check cost implications.
• Total: 6–12 people. The research is clear that the right mix matters more than the right technology.

How It Works

1. Set the scope (pre-workshop, 30 minutes):

The city manager and a facilitator define a focus area. Not “AI for the whole city” but something concrete: “AI opportunities for our permitting and inspections workflow” or “AI to support internal operations across departments.” Broad enough to explore but anchored enough to be actionable.

2. Review AI capabilities (20 minutes): 

The facilitator presents 8–10 plain-language slides showing what AI can do, each paired with a familiar everyday example:

• Summarize text—auto-generated meeting recap.
• Extract and format data—pulling structured fields from a scanned document.
• Classify things—email spam filter.
• Answer questions about documents—searching a policy manual by asking a question.
• Cluster similar items—grouping 311 complaints by topic.
• Discover patterns—spotting seasonal trends in code violations.
• Forecast demand—predicting next month’s permit volume from historical patterns.

The point is not to be exhaustive. It’s to give everyone in the room a shared, non-technical vocabulary for what AI can realistically do.

3. Brainstorm (30 minutes): 

Everyone writes ideas on sticky notes individually, then shares with the group. The key prompt: Where in your work would any of these capabilities be useful (even if it only worked 80 or 85% of the time)? The facilitator actively pushes back on ideas that require near-perfect performance and probes for situations where “pretty good” is still an improvement.

4. Sort and cluster (20 minutes): 

Group the sticky notes into themes.

5. Detail top concepts (30 minutes): 

For the highest-rated ideas, the team works through a simple concept sheet:

• What data would this need? Do we have it?
• What’s the AI capability involved?
• What would the output look like?
• What would the staff member do with that output?
• How accurate does the AI need to be for this to be useful?

6. Assess impact and effort (30 minutes): 

Map the top concepts on a 2x2 matrix—high impact versus low impact, high effort versus low effort. The sweet spot is the upper-left: high value, low effort. The research consistently found that these “low-hanging fruit” exist in every domain, but teams almost never find them without this structured process.

What You Walk Out With

A prioritized shortlist of 3–5 AI use cases, each captured with enough detail to evaluate for a pilot or proof of concept. Increased AI literacy is an added bonus. A well-structured AI Design Sprint doubles as AI literacy training. Across every team the researchers worked with, participants reported that the brainstorming exercise itself was as valuable as the specific ideas it generated.⁷

Some Final Thoughts and Resources

Consider a half-day workshop that brings the right people together, grounds them in what AI can realistically do, and searches for the best places to use it, while also building the shared AI understanding your team needs to evaluate vendor pitches, write smarter RFPs, and have more productive conversations with elected officials about what AI can and can’t do.

The AI Design Sprint isn’t a silver bullet for AI adoption, but it can be a starting point to bring the right people together (grounded in what AI can realistically do) and identify where “good enough” can create great value for your organization and community. For local governments, this means the best AI opportunities probably aren’t in replacing expert judgment or automating trivial tasks; they’re in augmenting expert work with moderate-quality AI that handles the tedious, time-consuming parts.

A great resource for getting started was developed by the University of Virginia’s School of Data Science professor Dr. Nur Yildirim: an open-source toolkit containing plain-language AI capability cards with real-world examples, sample slide deck, concept assessment sheets, and impact-effort matrices. Access it for free. 

SIRI RUSSELL is associate dean for community and government partnerships at the University of Virginia School of Data Science.

To learn more about the AI Design Sprint or have a UVA researcher facilitate a free workshop, contact us at sds_caal@virginia.edu.

ENDNOTES AND REFERENCES:
1. Ermakova, T., Blume, J., Fabian, B., Fomenko, E., Berlin, M., & Hauswirth, M. (2021). Beyond the hype: Why do data-driven projects fail? In Proceedings of the 54th Hawaii International Conference on System Sciences (pp. 5081–5090). https://doi.org/10.24251/HICSS.2021.619
Joshi, M. P., Su, N., Austin, R. D., & Sundaram, A. K. (2021). Why so many data science projects fail to deliver. MIT Sloan Management Review, 62(3), 85–89.
Weiner, J. (2020). Why AI/data science projects fail: How to avoid project pitfalls. Morgan & Claypool. https://doi.org/10.2200/S01070ED1V01Y202012CAN001
2. Yildirim, N. et al. (2023). “Creating Design Resources to Scaffold the Ideation of AI Concepts.” Designing Interactive Systems Conference (DIS ‘23). The study’s Task Expertise–AI Performance matrix maps 40 real-world AI applications and reveals that the majority create value with moderate model performance. Full paper: https://doi.org/10.1145/3563657.3596058.
3. Ibid.
4. Munteanu et al., 2006, as cited in Yildirim et al., 2023.
5. Kozyrkov, 2021, 2022, as cited in Yildirim et al., 2023.
6. Ibid, note 1.
7. Yildirim, N. et al. (2026). “AI Design Sprints: Facilitating AI Innovation within Cross-functional Industry Teams.” Designing Interactive Systems Conference (DIS ‘26). Tested with cross-functional teams in insurance, accounting, and consulting — regulated industries with dynamics similar to local government. Full paper: https://doi.org/10.1145/3800645.3812991
 

Next in this series: Once you’ve identified the right AI use case, how do you evaluate whether the tool you’re buying is trustworthy? We’ll explore what recent research reveals about AI security, reliability, and the questions every city/county should ask before procurement.