AI pilots tend to get attention when they launch. The more revealing moment may come later, when someone has to decide whether to keep paying for them.
Denver’s City Council provided a useful example on October 13 when it voted 9 to 3 against renewing the contract behind Sunny, the city’s AI chatbot for answering questions about city services. Sunny launched in 2024, operates in 72 languages, and the administration says it has a 90 percent satisfaction rating.
On the surface, those numbers sound like a reasonable case for renewal.
But that was not the question the council was asking.
The proposed agreement would have added nearly $3.9 million to a contract that had previously totaled $450,000, extended it through 2031, and expanded the scope to include a voicebot for lower-complexity 311 calls. Council members raised concerns about inaccurate answers, the effect on jobs, and why the agreement reached them after the previous contract had technically expired.
That creates an important distinction for anyone responsible for AI adoption.
Usage is not accuracy. Satisfaction is not reliability. And a successful pilot does not automatically justify a five-year expansion.
The administration and the council were effectively measuring different things. The administration brought satisfaction and usage information. Council members were asking about accuracy, workforce impact, governance, and whether the city should expand the system. The public reporting did not identify accuracy evidence offered in response.
That is the part worth paying attention to.
AI governance starts before renewal
Organizations often think about governance as the policy that determines whether employees can use an AI tool.
That is only one piece.
Governance also means deciding what evidence an AI service must produce to continue receiving money and organizational support.
If an AI system is going to answer residents, customers, employees, or constituents, its evaluation should probably include more than adoption statistics:
- How accurate are its answers?
- What kinds of errors does it make?
- How are those errors detected and corrected?
- What happens when the system does not know the answer?
- Has it changed workload or service levels?
- What has it actually cost to operate?
- What new risks appear when the scope expands?
- What happens to data and configuration if the organization changes vendors?
Denver’s situation also illustrates the danger of combining too many decisions. The council was not simply being asked whether Sunny should continue. The proposal bundled renewal, a substantially larger financial commitment, a five-year term, and a new voice channel.
A vote on expansion can quickly become a referendum on the entire AI program.
Build the renewal case while the system is running
For CIOs and program leaders, the practical lesson is simple: do not wait until contract renewal to figure out whether the AI system worked.
Define the evidence when the service begins.
Track accuracy alongside satisfaction. Sample the output. Document failures and corrections. Decide what success means before the vendor comes back with the next proposal. And separate major expansions from routine renewal when possible.
The original Squawk story put it well:
Ninety percent satisfaction answered a question the council did not ask.
That may be the broader AI governance lesson.
The question is not merely whether people are using the technology.
It is whether the organization has enough evidence to make the next decision.
Sources
Pingback: Nebraska's Statewide AI Assistant: How to Read Pilot Numbers - Brainsquawk