evidoria

← Back to browse

Good practice Imported

New Jersey's AI Assistant — an Open-Source, State-Built GenAI Tool Reaches 20,000 Public Employees

United States of America · Trenton · See the United States of America profile

Evidence: Descriptive / self-reported Top 6% 87/100 · Ask Evidence Copilot about this practice

New Jersey built its own secure generative-AI assistant for state workers, now used by 20,000+ of its 70,000 employees for 300,000+ sessions and a million-plus prompts, with agencies reporting 50% more resolved tax-hotline calls and 35% faster constituent email replies.

20,000+ users
State employees who have used the AI Assistant (by early-to-mid 2026)
300,000+ sessions
Total sessions logged (since July 2024)
1,000,000+ prompts
Prompts submitted (since July 2024)
1 USD/user/month
In-house cost per user, per month (2024-2026)
20 USD/user/month
Equivalent commercial licence cost per user, per month (2024-2026)
50 %
Increase in resolved ANCHOR tax-hotline calls (2024-2026)
35 %
Reduction in constituent email response time (2024-2026)
>80 %
Self-reported user satisfaction (2024-2026)

Details

Maturity
Scaling
Promoter
New Jersey Office of Innovation (New Jersey Innovation Authority)
Period
July 2024–present
Keywords
state government, internal productivity, generative AI, digital government

Context

In July 2024, New Jersey became one of the first US states to give its own employees a state-run generative-AI tool, built and hosted in-house rather than licensed from a commercial vendor.

Objectives

The programme aims to give the state's roughly 70,000 employees a secure, low-cost generative-AI assistant for everyday tasks, without sending state data to a third-party model provider.

Activities

By early-to-mid 2026, more than 20,000 state employees had used the Assistant, logging over 300,000 sessions and more than a million prompts. In March 2026 New Jersey relaunched the tool on an open-source fork of the LibreChat framework, publishing the code on GitHub, hosting it on state infrastructure, pledging not to use state data to train third-party models, and surfacing the model's step-by-step reasoning to users.

Results

Individual agencies report measurable effects: the Division of Taxation used the Assistant to rebuild its ANCHOR property-tax hotline's self-service phone menus, lifting successfully resolved calls by 50%, and the Department of Labor used it to draft plain-language emails, cutting response times by 35%. The state reports user satisfaction above 80%, and estimates the in-house build costs about $1 per user per month versus roughly $20 per user per month for equivalent commercial licenses.

Conclusions

No independent, external audit of the tool's accuracy or bias has been published to date; the performance figures above are self-reported by the issuing agencies.

Implementation
Indicative cost
Low (< €50k) — Estimated at about $1 per user per month for the in-house build versus roughly $20 per user per month for equivalent commercial licenses, with savings estimated in the millions of dollars in the first year.
Time to results
Medium (1–3 years) — Launched July 2024; relaunched on an open-source LibreChat fork in March 2026, with adoption still scaling toward the state's roughly 70,000 employees.
Staffing & skills
New Jersey Office of Innovation (New Jersey Innovation Authority) team, In-house engineering team building and hosting the open-source LibreChat fork, Agency staff in the Division of Taxation and Department of Labor who adapted the tool to their own workflows

Conditions for success

  • In-house development and hosting to control per-user cost and keep state data out of third-party training pipelines
  • A public pledge not to use state data to train third-party models
  • Surfacing the model's step-by-step reasoning to users to support trust and verification
  • A paired AI-skills training programme for state employees

Common failure modes

  • No independent external audit of the tool's accuracy or bias has been published to date, so the reported gains rest on agency self-reporting.

Where it fits

Governance type
US state government
Scale
state-wide (roughly 70,000 employees)
Income level
high-income

Commonly funded by

National / regional programmes

Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.

Replication kit

Reusable artefacts from this practice — as published by their sources.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Similar practices you may find useful