From the Local Government edition of September 23, 2026
Raleigh ends Daisy resident AI pilot after missing 90% correct-answer bar
City of Raleigh · Municipal resident services and customer experience innovation · Raleigh, North Carolina, United States
- Publisher
- StateTech Magazine
- Original publication
- September 23, 2026 StateTech Magazine coverage of Smart Cities Connect findings
- Source retrieved
- 2026-09-24
- Event date
- 2026-09-23
What happened
On Sept. 23, 2026 StateTech reported that Raleigh tested Daisy, an AI customer-service assistant with a humanlike avatar, from April through July and ultimately decided not to keep it. Officials set a firm bar: at least 90% correct answers before spending hundreds of thousands on broader deployment. Internal testing with 25 employees showed about 30% accuracy and ~4/10 satisfaction. An outdated strategic plan naming a prior mayor/council polluted answers (“garbage in, garbage out”). By pilot end only about 5% of answers were incorrect, but many replies were non-answers lacking information, so the 90% goal was still missed. Focus groups split on whether a human face helped; residents preferred people for emotional/complex needs and were more open to AI for queue-jumping simple questions or after-hours help. Takeaway emphasized staff-facing AI for tedious work and needs-based placement rather than slapping AI on everything. Cited 2025 PayIt index: 83% of governments implementing AI vs 50% consumer comfort.
Why it matters
Same-day cautionary measured pilot: explicit accuracy gate, data-hygiene failure mode, avatar UX split, and decision not to procure—contrasts with go-live announcements peers may over-weight.
Evidence and measured results
StateTech Sept. 23, 2026: Daisy April–July pilot; not retained; 90% correct bar; internal ~30% accuracy / ~4/10 satisfaction; outdated strategic plan caused errors; end-state ~5% incorrect but too many “don’t have information” replies; avatar face controversial; complex/emotional questions prefer humans; staff-facing AI may offer more value.
Limitations and uncertainty
Conference reporting of city officials’ briefing; exact vendor and public sample sizes beyond the 25-employee internal test are limited in the article; PayIt comfort statistic is secondary citation.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-09-24; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Resident service queues and language-access gaps create pressure to buy AI assistants before accuracy is proven.
- Stakeholders
- Strategy/innovation leads
- Customer service managers
- IT/content owners
- Mayor/council
- Residents/language-access advocates
- Discovery questions
- What correct-answer and non-answer thresholds stop the procurement?
- How is superseded content removed from the corpus?
- Is a human avatar required, optional, or banned?
- Where do complex/emotional cases hand off to people?
- Value hypothesis
- A gated pilot can reveal whether resident AI beats empty answers and stale data—without assuming production spend is justified.
- Potential engagement
- Pilot design with written go/no-go accuracy gate, corpus cleanup, and UX testing—explicitly allowing a stop decision like Raleigh.
- Unsupported claims
- Raleigh proved a successful production chatbot
- 30% early accuracy is acceptable for go-live
- Human avatars are universally preferred
- Guaranteed call deflection or FTE savings
Pre-sales engineering
Role takeaway
- Fit
- Grounded multilingual resident assistant with strict corpus governance and measurable non-answer handling.
- Architecture and integration
- Kiosk/web channel → retrieve approved city docs → answer or abstain → log correctness samples; optional non-human persona.
- Prerequisites
- Document inventory and retirement workflow
- Language test sets
- Accuracy sampling harness
- Human handoff path
- Constraints
- Missed 90% useful-correct bar
- Avatar polarization
- Pilot scoped to Parks info in testing narrative
- Conference-reported metrics
- Security
- Keep only current authoritative docs in the index; control who can publish to the bot corpus.
- Proof of value
- Correct-answer rate
- Non-answer rate
- Satisfaction
- Language-access success
Delivery
Role takeaway
- Work
- Clean corpus, run employee then public tests, facilitate focus groups on persona, and execute a documented stop/scale decision.
- Dependencies
- Content stewards
- Innovation team
- Vendor (if any)
- Communications
- Ownership
- Strategy and Innovation owns the pilot decision; departments own source content.
- Skills and adoption
- Train staff that fixing documents fixes answers; brief front-line teams on when to override.
- Governance checkpoints
- Gate memo before public pilot
- Mid-pilot corpus audit
- End go/no-go
- Postmortem on stop decision
- Acceptance criteria
- Written accuracy gate exists
- Stale docs removed after discovery
- Stop decision documented if unmet
- Human path for complex cases
- Risks
- Launching despite empty answers
- Avatar controversy
- Political pressure to buy anyway
- Ignoring language-access failures
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Corpus currency controls and non-answer rates matter as much as factual error rates; avatar UX is a product choice with polarized acceptance.
Governance
Who approves, reviews and stays accountable for outcomes?
Pre-commit published accuracy/usability gates and be willing to stop spend when unmet; treat shadow/outdated documents as deployment blockers.
Security and privacy
What data, permissions and controls need testing?
Stale authoritative documents create confident wrong answers; inventory and retire superseded plans before retrieval grounding.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Multilingual testing matters, but empty answers and avatar discomfort can still block equitable access; keep human service for complex needs.
Procurement
What should contracts, pricing and exit terms secure?
Write kill criteria into pilots before six-figure deployments; pay for measured learning, not assumed production.
Operating model
Which teams own the service once it runs?
Prefer staff-facing AI for tedious work when resident-facing tools fail gates; redesign service halls around human problem-solving.
Publication history
- 2026-09-23Local Government · Issue 184 resources
Stable resource ID: raleigh-daisy-pilot-not-retained-20260923