How do governments adopt AI tools that truly serve communities’ needs? In a recent InnovateUS workshop, Thomas Gilbert, Founder and CEO of Hortus AI, offered public service professionals a practical guide to navigating the vendor landscape, evaluating artificial intelligence (AI) tools, and setting clear expectations for trustworthy implementation.
Gilbert, who founded his PhD program in machine ethics nearly a decade ago at UC Berkeley, began the workshop by highlighting AI deployments gone wrong. From Air Canada's chatbot promising refund policies the company never intended to honor, to New York City's MyCity bot suggesting small business owners commit arson for insurance money, he explained how the current approach to AI adoption needs to be improved.
"We are living in a Wild West moment with respect to AI development, AI adoption, and often AI procurement," Gilbert said. "And that's not sustainable."
Chatbots vs. AI Agents: Understanding the Difference
Gilbert then made the important distinction between chatbots and AI agents.
While a chatbot might help draft emails or summarize documents, an agent could directly engage citizens, manage public services, or even impact legal standing—creating entirely new categories of risk and opportunity.
The risks that come with the usage of both these technologies also look very different.
“If you're using AI agents to do [legal] work, they may start asserting or impacting the legal standing of certain constituents or stakeholders in ways that may or may not fit with your own understanding, which is a very different kind of risk than a chatbot that just might occasionally hallucinate or not respond well to your prompts,” he said.
Trust is About Vendors, Not Just Technology
The remainder of the workshop focused on reframing AI trust as fundamentally about vendor relationships rather than technical capabilities.
Gilbert outlined several approaches for evaluating AI vendors beyond their marketing claims:
-
Algorithmic Audits: Third-party assessments that reveal how systems perform across different populations and use cases
-
Algorithmic Impact Assessments (AIAs): Systematic evaluations of potential risks before deployment, helping teams anticipate and plan for edge cases and failure modes
-
Model Cards and Data Sheets: Standardized documentation that reveals training data sources, accuracy metrics, and recommended use cases
Rather than accepting what vendors offer "out of the box," Gilbert advocated for more collaborative approaches to AI procurement. He highlighted emerging pilot programs and community information exchanges that prioritize stakeholder needs over technological capabilities.
Community Information Exchanges, platforms that coordinate services across entire communities rather than just making individual processes more efficient. Cities like San Diego, Chicago, and Columbus are leading this movement, creating integrated systems that bring communities together rather than optimizing isolated functions.
The Future of AI Procurement
Looking ahead, Gilbert sees AI transforming not just what government does, but how it does it.
"AI itself is becoming the fulcrum around which tech procurement, pilots, and shared services rotate," he said. "AI is not just something that is being procured. AI is becoming procurement itself."
This shift creates both opportunity and urgency. While there is currently uncertainty in the field, there is real potential for the government to lead rather than simply react. This starts with empowering communities to identify what successful adoption looks like and prioritizing human considerations over technology.
"I do think there's a powerful vision there for communities to lead that conversation for stakeholders to play a much more active role and for them to decide what their own problems are before they seek out solutions,” Gilbert said. “And I think that's really where we should look for leadership now in a moment, where we will need to start local and need to start with communities in order to get this right."
To watch Gilbert’s full workshop, click here.
