From the State Government edition of October 15, 2026
StateTech (Oct 15): Maryland, Caltrans and Colorado officials describe moving GenAI from pilot to production, with Maryland citing ~55% employee adoption and a 97% SUN Bucks chatbot figure
Maryland (Senior AI Advisor), California State Transportation Agency/Caltrans, Colorado OIT (former CIO); reported by StateTech · State enterprise IT / GenAI productivity tools and constituent chatbots · Maryland, California, Colorado, Georgia (multi-state)
- Publisher
- StateTech Magazine
- Original publication
- 2026-10-15
- Source retrieved
- 2026-10-16
- Event date
- 2026-10-15
What happened
StateTech's Oct 15 feature quotes state officials on scaling AI past pilots. Maryland's Senior AI Advisor says Gemini and Copilot rollouts have reached adoption by about 55% of employees, one agency saved about $400,000 by doing work in Gemini instead of a vendor contract, and the SUN Bucks nutrition-benefit chatbot has had over 30,000 chats and resolves the question in 97% of cases. Caltrans' information security officer says two GenAI efforts (Road Safety Insights, Traffic Mobility Insights) moved into production and advises planning for production funding, support, cybersecurity, licenses and monitoring when the pilot is planned. Colorado's former CIO says Gemini went from a 150-plus-user pilot that deliberately included employees with disabilities to 18 agencies and about 2,000 initial users, gated by request, manager approval and responsible-use training. The Center for Democracy and Technology analyst cites unclear use cases and lack of scale-up funding as stall causes.
Why it matters
Fresh, quotable peer evidence for the pilot-to-production problem most state CIOs now face, with concrete operating practices (access gating, mandatory training, production budgeting). Usefulness is limited because metrics are officials' verbal claims, and the Maryland chatbot figure is framed differently from the state's earlier published description.
Evidence and measured results
Opened the StateTech article over HTTPS (curl, HTML text extraction) and read all state sections. For scrutiny, opened GovTech's Oct 29, 2025 Maryland report and Code for America's Jul 31, 2025 note on the Maryland Summer EBT chatbot. GovTech (2025) reported ~43,000 employees with Gemini access, ~12,500 active users, about 30% regular use, the same $400,000 Maryland Department of the Environment saving, and SUN Bucks Chat 'accurately answered questions in 97 percent of scenarios' with 16,000-plus users.
Limitations and uncertainty
Vendor-adjacent outlet (StateTech is published by CDW, an IT reseller); claims are officials' quotes, not audited results. The Maryland 97% figure appears here as a resolution rate across 30,000-plus chats, whereas GovTech in Oct 2025 reported it as accuracy across test scenarios by the state benefits team; the two are different measures and the methodology is unpublished. The '55% adoption' is not defined (access, monthly use, or regular use) and compares with about 30% regular use reported in Oct 2025. The $400,000 saving is the same example reported a year ago, not new. Caltrans production tools were archived earlier via NASCIO award coverage; only the pilot-to-production advice is new.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-10-15; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- States have many GenAI pilots and enterprise licenses but stall moving use cases into funded, supported production with credible metrics.
- Stakeholders
- State CIO or chief AI officer
- Agency program leads with successful pilots
- Budget office
- CISO and privacy officer
- HR and training lead
- Discovery questions
- Which pilots have shown value, and is there recurring funding for licenses, support and monitoring?
- How do you define adoption: access, monthly use or regular use?
- How are chatbot results measured: test-scenario accuracy, live resolution or containment?
- Who owns each AI tool once it is in production?
- Value hypothesis
- A structured production-readiness review may help a state move a few proven pilots into supported service with defined metrics; savings depend on the use case and are not established by this source.
- Potential engagement
- Fixed-scope pilot-to-production assessment of three to five existing GenAI use cases, producing funding, ownership, access-gating and metric definitions.
- Unsupported claims
- That enterprise AI adoption of about 55% is typical or comparable across states
- That state chatbots resolve 97% of live inquiries
- Repeatable vendor-contract savings like Maryland's single $400,000 example
- Any audited productivity gain
Pre-sales engineering
Role takeaway
- Fit
- States with Gemini or Copilot licenses and constituent chatbots that need production operations, monitoring and metrics.
- Architecture and integration
- Tie tool access to identity groups and training completion, connect domain tools to governed data sources, and add usage telemetry and answer-quality monitoring for chatbots before declaring production.
- Prerequisites
- Enterprise identity and license management
- Training-completion records linked to access
- Data governance for the sources tools use
- Logging and monitoring for usage and chatbot answers
- Constraints
- Officials' figures are unaudited and definitions unpublished
- Commercial suite features and data terms vary by contract
- High-risk uses such as benefit eligibility need heavier review
- Security
- Treat cybersecurity and monitoring as recurring production costs, restrict access by role and training status, keep sensitive data in state-controlled tenants, and log prompts and outputs for audit.
- Proof of value
- Report regular-use rate separately from licensed access
- Measure chatbot accuracy on a fixed scenario set and live resolution with repeat-contact checks as separate metrics
- Validate any cost-avoidance claim against the contract it replaced
- Track production support hours and costs per use case
Delivery
Role takeaway
- Work
- Inventory pilots, select candidates for production, assign operating owners and budgets, define metrics, set up training-gated access and schedule periodic audits.
- Dependencies
- Recurring budget authority
- Executive governance body or subcabinet
- Training provider and completion tracking
- Data owners for each use case
- Ownership
- State enterprise AI office or CIO organization, with each agency program owning its production use case.
- Skills and adoption
- Mandatory responsible-use training before access, plus pairing technologists with business experts to design use cases; include employees with disabilities in pilots for accessibility feedback.
- Governance checkpoints
- Risk tiering before production
- Metric definitions approved before launch
- Training completion verified before access
- Periodic audit for accuracy and bias on high-risk uses
- Acceptance criteria
- Proposed: each production use case has a named owner and recurring budget line
- Proposed: regular-use rate defined and reported quarterly
- Proposed: chatbot accuracy on a scenario set and live resolution reported separately
- Proposed: 100% of users complete training before access
- Risks
- Pilots stall without recurring funding
- Adoption and chatbot metrics framed inconsistently
- Governance left to IT alone
- Reliance on unaudited peer claims
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?
Maryland runs commercial suites (Google Workspace with Gemini, Microsoft Copilot) rather than custom platforms for broad productivity, plus a purpose-built benefits chatbot; Caltrans built domain GenAI tools over safety and mobility data. The article says production requires ongoing infrastructure, data management, monitoring and named operators, but gives no architecture detail.
Governance
Who approves, reviews and stays accountable for outcomes?
Practices cited: Maryland AI subcabinet and statutory governance process; Colorado request-and-approve access with mandatory responsible-use training; CDT/GovAI Coalition checklist emphasizing transparency and accuracy, extra scrutiny for high-risk uses such as benefit eligibility, and recurring audits for bias. Gap: no state described how production metrics are defined or audited.
Security and privacy
What data, permissions and controls need testing?
Caltrans frames cybersecurity and data management as recurring production costs. Georgia is described as running a chatbot where data never leaves the state environment, without detail. No security incidents or controls testing are reported.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Colorado credits including employees with disabilities in its pilot as one of its most valuable decisions; access is tied to completed training. Caltrans stresses pairing AI literacy with business experts who do the work. No workforce reduction claims are made.
Procurement
What should contracts, pricing and exit terms secure?
Maryland's example of displacing a vendor contract with in-house Gemini work signals that enterprise AI licenses can substitute for some small consulting or software buys, which procurement offices may need to track. Pilots funded as one-time projects need recurring license, support and monitoring budgets before production.
Operating model
Which teams own the service once it runs?
Recurring theme: production requires a named owner, recurring funding and enterprise governance spanning cybersecurity, privacy, legal, procurement and HR, not IT alone. Colorado's gated rollout and Maryland's subcabinet are two models for controlling scale-up.
Publication history
- 2026-10-15State Government · Issue 382 resources
Stable resource ID: statetech-pilot-to-production-maryland-caltrans-colorado-20261015