How to Use This Deck
This is prep for our March 20 meeting. Take a look, find your department, and see if anything clicks.
- Browse this deck — Jump to your department using the nav on the left. See which DMV tools might apply to your shop.
- If something resonates — Try to have a quick conversation with someone in your department before March 20. Even five minutes with the right person will surface more than a week of desk research.
- March 20 — We'll talk it through as a group.
- After — A survey will capture everyone's ideas in writing.
Fair warning: the department suggestions in this deck are AI-generated starting points. Some will be on target; others will miss completely. That's fine — you know your department better than any model does. The goal is to spark ideas, not prescribe answers.
Why This Matters
California DMV is running six distinct AI/automation systems in production — more than most state departments have even piloted.
The critical pattern: DMV uses AI as a decision-support layer (human stays in the loop), not full automation. This approach:
- Requires no new legislation
- Has survived legal challenge (Personalized Plates: zero lawsuits since GenAI implementation)
- Produces measurable, auditable results
The Governor's Innovation Fellows Program puts 21 Fellows across 15 departments in 9 agencies — every one of which has a direct opportunity to benefit from DMV's proven approach.
The Master Pattern: AI Assists, Humans Decide
Every one of DMV's 6 AI systems follows the same architecture:
- AI processes — Extracts, classifies, recommends, routes
- Human decides — Reviews flagged items, makes final call
- Automation executes — Carries out the approved action
This isn't a coincidence. It's a deliberate design choice that:
- Satisfies legal requirements — Human accountability for decisions
- Builds staff trust — AI augments, doesn't replace
- Survives scrutiny — Personalized Plates has had zero lawsuits since implementation
- Improves over time — ML learns from human corrections
This is the pattern other departments should look at. Not "AI replaces workers" — but "AI handles the 60-90% that's routine so workers can focus on the cases that actually need judgment."
The 6 AI Systems in Production
| # | System | What It Does | Key Metric |
|---|---|---|---|
| 1 | Miles Chatbot | Answers phone and web inquiries about DMV services — renewals, appointments, title transfers — using natural language processing. Escalates to a live agent when it can't resolve. | 13.77M calls/year, 49% self-service |
| 2 | RADV | Scans uploaded identity documents (birth certificates, passports, EAD cards) using OCR, extracts data, validates it against DMV records, and auto-approves clean submissions. | 5.4M docs/year, 62% auto-approved |
| 3 | MyDMV Identity | Verifies that online users are who they claim to be using a two-tier system — first automated checks, then fallback to enhanced verification. Catches fraud before it reaches a human. | 90-95% verification success |
| 4 | Service Advisor | AI-powered search on dmv.ca.gov that returns structured answers and form-filling guidance instead of just links. Detects scam reference numbers and surfaces fraud alerts. | 1M+ queries/year |
| 5 | GenAI Plates | Reviews personalized license plate requests for offensive or inappropriate content, then auto-assigns approved plates to vehicles 24/7. Replaced subjective human review that led to lawsuits. | 90% auto-processed, 0 lawsuits |
| 6 | Disaster Recovery | Mobile GIS app used by investigators to locate and document vehicles destroyed in wildfires — mapping clusters, collecting field data on iPads, syncing to central databases in real time. | 195 vehicles mapped (Sacramento) |
Each system follows the same philosophy: AI assists, humans decide.
4 Focus Areas for March 20
For this meeting, we're focusing on the four DMV systems with the broadest applicability across Fellowship departments:
| Focus Area | What DMV Built | Why It Matters to You |
|---|---|---|
| Miles Chatbot | Voice & chat AI handling 13.77M calls/year — 49% now fully self-service | Every department has a call center or help desk. What would your Miles answer? |
| RADV Document Verification | OCR + ML auto-verifies 62% of documents — saving 22,050 hours/year (28 FTE) | Every department processes paper. Which forms eat the most staff time? |
| MyDMV Identity Verification | Multi-tier identity proofing at 90-95% success (up from 35-40%) | Every department with online accounts needs identity verification. Where does proving identity create friction? |
| Service Advisor | AI-powered search returning structured answers + guided form-filling — 1M+ queries/year | Every department has a website. Every website has a "How do I...?" problem. |
The remaining 2 systems (GenAI Plates, Disaster Recovery) are covered at the end of this deck as additional reference.
Miles Chatbot
Scale: 13.77 Million Calls in 2025
The bot deflected 6.8M calls to self-service in 2025 (up from 6.0M in 2024) — a 13% increase in automation without adding staff. Of the remaining calls, 5.6M requested a live agent and 1.4M left a message.
Top call reasons: VR Renewal (13%), Title Transfer (5%), DL Renewal (5%), Drive Test Appointments (4%).
RADV Document Verification
Technology: ABBYY OCR/ICR + Machine Learning
How It Works
Citizens upload identity documents (birth certificates, passports, EAD cards, I-94s) through the online portal. The ABBYY engine:
- Extracts printed text and handwritten entries via OCR/ICR
- Validates extracted data against DMV's database
- Auto-approves documents that pass (~62%)
- Routes failures to a human technician for manual review
Five-Year Average Savings
- Time saved: 22,050 hours/year
- Staff equivalent: 28 FTE/year
- Documents processed: Millions monthly
MyDMV Identity Verification
Technology: Multi-tier automated identity verification via MyDMV + biometric exploration
The Problem → The Fix
| Before | After | |
|---|---|---|
| Approach | Single verification attempt | Two-tier (Level 1 → Level 2 fallback) |
| Success Rate | 35-40% | 90-95% |
| Failure Routing | All failures → call center | Only double-failures → agent |
| Document Flexibility | Rigid single pathway | Multiple verification pathways |
High-Risk Application Protections
For sensitive operations (phone number reset, replacement DL), the system adds external identity verification services — validating against third-party databases in addition to DMV records.
Future: Biometric Verification
DMV is exploring selfie + camera comparison against the DL photo on file — for both MyDMV account creation and CA DMV Wallet (mDL) issuance. Same face = verified identity = credential issued.
Service Advisor
Technology: Semantic search + guided form-filling assistant
Integrated into dmv.ca.gov at the three highest-traffic touchpoints: homepage search, site-wide search, and the appointments page.
Not Just Search — It's an AI Assistant
Contextual answers — Returns structured guidance, not just links. Searching "statement of facts" identifies the exact form (REG 256), explains all 6 uses, and offers a guided Form Filler.
Form Filler integration — Walks users step-by-step through DMV forms. Enter your license plate, answer questions, and it generates a completed PDF.
Scam detection — Recognizes known scam reference numbers and immediately surfaces fraud alerts: "The DMV will never ask for personal or financial information by text."
mDL Q&A assistant — Handles mobile driver's license questions with corrective actions when something goes wrong.
This is the most universally transferable pattern. Every department has a website. Every website has a "How do I...?" problem.
GovOps Agency
CDT (Dept of Technology) — Jennifer Uyeda Issertell
| DMV Tool | What This Could Look Like |
|---|---|
| Miles-style Chatbot | Statewide IT help desk bot — password resets, service tickets, system status. CDT's ServiceNow platform already has structured data; a bot layer on top could deflect 40%+ of routine calls. |
| Identity Verification | CDT could offer identity proofing as a shared service for all state departments. One implementation, every department benefits. The EDD Strike Team demonstrated what happens when identity verification fails at scale — $10-32B in fraud losses and call centers that couldn't keep up. |
DGS (General Services) — Lorna Brisco
| DMV Tool | What This Could Look Like |
|---|---|
| Miles-style Chatbot | DGS has 4 trainers for 4,000 employees (1:1000 ratio). An internal knowledge bot handling policy questions, onboarding guidance, and procedure lookups could bridge the training gap that can't be staffed. Externally: vendors constantly call about bid status, contract questions, and Cal eProcure — a procurement status bot could deflect significant volume. |
| Document Verification | SB/DVBE certification processing and K-12 school construction plan reviews are high-volume document workflows where automated extraction could speed review cycles. |
CDTFA (Tax & Fee Admin) — Jeremiah Oakden
| DMV Tool | What This Could Look Like |
|---|---|
| Miles-style Chatbot | CDTFA collects $90B+ annually across 42 different programs — every retail business in California files with them. Taxpayer inquiries — "When is my return due?" "What's my balance?" — are the exact same query types Miles handles for DMV. The Customer Service Center already publishes live wait times, signaling call volume is a known problem. |
| Document Verification | Sales tax returns, exemption certificates, and cannabis compliance filings across hundreds of thousands of businesses. CDTFA already digitizes returns; adding ML validation could flag mismatches before they reach auditors. |
Health & Human Services
DDS (Developmental Services) — Henri Aghaei Baradaran
| DMV Tool | What This Could Look Like |
|---|---|
| Document Verification | 21 regional centers each process eligibility paperwork and Individual Program Plans for 491K+ consumers — but with no unified data system across centers. Automated verification of standardized eligibility forms could reduce the 120-day intake timeline. The inconsistency between centers in how documents are formatted and processed is where the biggest gains hide. |
DCSS (Child Support) — Bryanna McAdams
| DMV Tool | What This Could Look Like |
|---|---|
| Miles-style Chatbot | DCSS already has "Customer Connect 24/7" — but it's keyword-based. The real gap is handling complex, sensitive inquiries: not just "What's my balance?" but "I'm being released from custody, what happens to my child support order?" or "I'm deployed overseas, how do I modify my support obligation?" 1.04M+ open cases, $2.55B/year distributed across 47 county agencies. |
| Document Verification | Income declarations for child support orders across 47 counties — each county submits slightly different formats. ML normalization is the high-value target. |
| Identity Verification | Parent identity verification for the child support portal across 47 counties — reliable identity proofing reduces fraud and call center burden. |
OTSI/CalHHS — Kattya Trinh
| DMV Tool | What This Could Look Like |
|---|---|
| Miles-style Chatbot | The All-Hazards Dashboard (NASCIO award winner) already aggregates data across CalHHS agencies for emergency response. The same coordination challenge exists for Californians navigating health and human services — they bounce between DHCS, DSS, DDS, and others trying to figure out eligibility and status. A benefit navigation chatbot coordinated across agencies is a natural extension of that cross-agency platform work. |
| Identity Verification | CalHHS agencies each verify identity independently — a shared identity layer (like DMV's model) serving DHCS, DSS, DDS from one platform would reduce duplication. Medi-Cal alone has 14M+ beneficiaries. |
Transportation Agency
CHP (Highway Patrol) — Lt. Marc Peachey
| DMV Tool | What This Could Look Like |
|---|---|
| Document Verification | CHP processes hundreds of thousands of collision reports (Form CHP 555) annually, plus commercial vehicle inspection reports and CAD records. Automated extraction from standardized collision report forms could speed data entry and improve the SWITRS collision database. |
| Miles-style Chatbot | CHP's primary call volume is 911 emergency dispatch (LA alone handles 2.7M calls/year) — that's fundamentally different from DMV's Miles model. But non-emergency public inquiries (collision report requests, commercial vehicle permit status, tow company licensing) could be deflected to self-service. |
Caltrans — Ben Bressette
| DMV Tool | What This Could Look Like |
|---|---|
| Miles-style Chatbot | Encroachment permits went mandatory online in January 2025 — contractors and utilities now navigate CEPS digitally at $173/hr processing cost. A chatbot guiding applicants through the permitting process could reduce processing errors and call volume across all 12 districts. |
Natural Resources & Environment
DWR (Water Resources) — Nikki Hatcher
| DMV Tool | What This Could Look Like |
|---|---|
| Miles-style Chatbot | The procurement branch handles thousands of contracts with 3-18 month timelines across DWR's 29 State Water Project contractor agencies. Vendors and internal staff ask the same questions: "What's the status of my contract?" "What documents do I need for this solicitation?" A procurement chatbot could reduce the repetitive inquiries the team fields daily — and the 5% Contract Reduction Pilot is already measuring exactly this kind of efficiency gain. |
| Document Verification | Contract compliance documents, dam safety inspection reports, and environmental impact reports are high-volume document types where automated extraction could flag errors before they stall the pipeline. |
SWRCB (Water Board) — Brent Vanderburgh
| DMV Tool | What This Could Look Like |
|---|---|
| Document Verification | The water rights division has 7M+ paper records dating to the 1800s. The $60M UPWARD initiative is already digitizing them; AI verification is the logical next step. Water quality lab data (chemistry samples across thousands of monitoring sites) needs anomaly detection — flagging outlier results before they hit compliance reports. |
OEHHA — Kannan Krishnan
| DMV Tool | What This Could Look Like |
|---|---|
| GenAI Content Review | OEHHA's chemical assessment process is methodical by necessity — but the volume of new research outpaces manual review capacity. GenAI could triage incoming research papers, flag chemicals needing expedited Prop 65 listing review, and cross-reference new findings against the existing 900+ chemical list. This is the Personalized Plates pattern (AI reviews, human decides) applied to scientific literature screening. |
| Service Advisor | For businesses and attorneys querying the Prop 65 list: a semantic search tool that interprets plain-English queries ("Is BPA in baby bottles covered?") is more appropriate than a full chatbot for an agency of ~120 staff. |
CARB (Air Resources Board) — Christina Marin-Fitzhugh
| DMV Tool | What This Could Look Like |
|---|---|
| Miles-style Chatbot | With 1,700+ employees, CARB's HR branch fields constant questions about hiring timelines, classification specs, leave policies, and health/safety protocols. An internal HR knowledge bot — "What's the process for an out-of-class assignment?" "How do I request a reasonable accommodation?" — could reduce repetitive inquiries. Department-wide: regulated entities (fleet operators, manufacturers, cap-and-trade participants) navigate fragmented compliance portals — a public-facing chatbot is a separate but large opportunity. |
| Document Verification | The new climate disclosure filings (SB 253/261, thousands of companies starting 2026) will create a surge of document processing on the regulatory side. |
CNRA (Natural Resources Agency) — Elizabeth Betancourt
| DMV Tool | What This Could Look Like |
|---|---|
| Document Verification | The ADA document remediation initiative — AI-assisted accessibility compliance for ca.gov websites — is the most broadly applicable project of any Fellow: a cross-cutting solution for all 170+ departments. For CNRA specifically: bond fund administration across 26+ entities involves massive document tracking and environmental review document summarization at the policy level. |
Military, Corrections & Other
CDCR (Corrections) — Samantha Kissane & Chris Siino
| DMV Tool | What This Could Look Like |
|---|---|
| Document Verification | CDCR processes 20,000+ grievances/month, thousands of classification chronos (Form 128-B, many still handwritten), and workers' comp claims currently tracked in individual Excel spreadsheets with no centralized database. Employee files don't transfer between institutions — pattern claims can't even be identified. Automated document extraction could tackle the grievance backlog alone. |
| Identity Verification | DMV's kiosk-based biometric approach could transform officer and inmate check-in/check-out across 31 institutions. |
CMD (Military Dept) — Jai London, SSgt Kelton Pisano & SMSgt Blake Carter
| DMV Tool | What This Could Look Like |
|---|---|
| Document Verification | Military paperwork is highly standardized (DD-214s, SF-180s, deployment orders) — ideal for OCR with near-zero customization. The A1 Connect platform already handles action item workflows across 5 Wings; AI document verification could extend it. Wildfire cost dashboards could integrate automated verification of reimbursement claims against historical norms. |
| Miles-style Chatbot | With dual state/federal pay systems and complex personnel rules, Guard members likely have repetitive questions about benefits, deployment status, and personnel actions. Centralized procurement could also benefit from a vendor/internal FAQ bot. |
LCI (Land Use & Climate Innovation) — Will Robinson
| DMV Tool | What This Could Look Like |
|---|---|
| GenAI Content Review | LCI's legislative portfolio spans complex, overlapping policy domains (land use, climate, housing, tribal consultation). GenAI-assisted bill analysis — flagging conflicts with existing CEQA Guidelines, summarizing amendments, cross-referencing related bills — is the Personalized Plates pattern applied to legislative work. LCI is a small policy shop (~50-80 staff), not a service-delivery agency, so the fit is targeted review tools, not high-volume automation. |
| Service Advisor | The Site Check mapping tool could integrate AI-driven search to help local governments and developers understand CEQA requirements for specific locations. |
CA Volunteers — Molly Linares
| DMV Tool | What This Could Look Like |
|---|---|
| Miles-style Chatbot | CaliforniaVolunteers runs the nation's largest statewide volunteer matching network — 68+ AmeriCorps programs, ~7,000 members at 1,000+ locations, initiatives targeting 10,000 mentors (Men's Service Challenge) and 55 campuses (College Corps). The "How do I find the right program?" question has real volume and real complexity. |
| Document Verification | Grant management across 68+ programs involves applications, renewals, performance reports, and financial reimbursements — a document processing pipeline where automated extraction could reduce staff review time. |
Also in Production: GenAI Personalized Plates
Technology: Generative AI (content review) + RPA (plate assignment)
The Origin Story
A 2020 lawsuit (Ogilvie v. Steve Gordon) exposed that plate configuration decisions were subjective and inconsistent across reviewers. The numbers told the story:
- 167,000 orders/year (~668/day)
- Only 4 reviewers + 5 assignment staff
- 36,000 plate backlog
- Up to 9 months wait time
- Inconsistent approve/deny decisions
GenAI Plates: The AI Solution
Two-Part Fix
Generative AI (Nov 2022): Reviews plate configurations for offensive/inappropriate content. Makes approve/deny recommendations. Eliminated human bias and created consistency.
Intelligent Automation (July 2023): Assigns approved plates to vehicles automatically. Runs 24/7 — no breaks, no sick days, works beyond business hours.
End-to-End Flow
Intake → GenAI Assist → Staff Review → IA Assist → Fulfillment
Online, Mail, Field Office, AAA → Machine review, flags & routing → Final human decision → Assigns plates to vehicles → CALPIA Manufacturing
Results
| Metric | Result |
|---|---|
| IA processing rate | 90% of all plate assignments |
| AI approval rate | 90% of AI recommendations confirmed by staff |
| Wait time | 2-4 months (was 9 months) — 50% reduction |
| Lawsuits | 0 since implementation |
The human-in-the-loop design is why there are zero lawsuits. GenAI recommends, staff decides.
Also in Production: Disaster Recovery App
Technology: GIS-based mobile application for field data collection
Real-World Application
Used by DMV investigators to locate and document vehicles destroyed in wildfires and other natural disasters. The mobile app:
- Maps investigation clusters using GIS (195 vehicles in Sacramento region alone)
- Enables on-site data collection via iPad
- Tracks investigation status in real-time
- Syncs field data with central databases
Field Operations
Investigators physically locate burned vehicles at disaster sites — crawling under wreckage to read VINs, photographing damage, recording GPS coordinates — all feeding back through the app.
Less "AI" in the traditional sense, more intelligent field data collection — but the GIS clustering, mobile-first workflow, and real-time sync represent a model for any department with field operations.
DMV Contacts
DMV Innovation Team
| Name | Area | |
|---|---|---|
| Adrian Monteon | Document Processing (RADV) | adrian.monteon@dmv.ca.gov |
| Sonia Huestis | Miles Chatbot | sonia.huestis@dmv.ca.gov |
| Randolph L. Fernandez-Gonzalez | Service Advisor | randolph.gonzalez@dmv.ca.gov |
| Stefan Schoy | MyDMV Identity | stefan.schoy@dmv.ca.gov |
| Amy Burks | Disaster Recovery App | amy.burks@dmv.ca.gov |
| Angela Marbray | GenAI Personalized Plates | angela.marbray@dmv.ca.gov |
Your Internal Bridge
Liyuan Guo — DMV's own Innovation Fellow (EEO Officer) — is already in Cohort 1. She can serve as the liaison between DMV's innovation teams and Fellows from other departments.