Domain Atlas / Caseworker documentation & copilots
NJ AI Assistant
New Jersey built and hosts its own generative drafting assistant for state employees, with a state-owned interface, hosting and logs and a hosted commercial frontier-model service as the one external dependency in the serving path, an ownership arrangement that leaves the underlying model replaceable without reprocuring an application or moving employees onto a different product. The state reports that roughly 20,000 employees had used the tool across more than 300,000 sessions and more than 1,000,000 prompts by February 2026 at about one dollar per user per month, that access onboarding routes through a responsible-AI course whose curriculum is used by 25 or more states, and that unemployment-insurance staff rewriting claimant emails in plain language saw claimants respond 35 percent faster. Every outcome figure is state self-reported; the 35 percent figure has no published methodology and predates the statewide launch, and no inspector-general evaluation, external audit or peer-reviewed causal study of the tool was located.[5]
What happened
New Jersey launched the NJ AI Assistant in July 2024 as one of the first state-built generative AI tools for government employees, deliberately paired at launch with a responsible-AI training course developed with a nonprofit partner. The lineage runs through Executive Order 346 of October 2023, which created the state's AI Task Force, and an interim generative-AI policy for state employees issued in November 2023 that a fuller joint circular from the Office of Information Technology and the state cybersecurity cell later superseded. The tool's construction is the case's first distinctive fact, and it has to be stated precisely. The interface, the hosting and the logs are state-owned: built in house by the NJ Office of Innovation with the Office of Information Technology and the Civil Service Commission, tied to official state accounts, and rebuilt in March 2026 on an open-source chat interface. The underlying model is a hosted commercial frontier-model service held under state data-control terms, with state statements that state data is not used to train third-party models. There is no vendor-owned application or software-as-a-service layer in the serving path, and that is not the same as zero vendor dependency — what the arrangement buys is portability, because the underlying model can be changed without reprocuring an application or moving employees onto a different product, the state-run tool and the official accounts behind it staying put through such a change. Adoption grew from more than 4,000 users and more than 4,000 training completions within about two months of launch to roughly 14,000 monthly active users by August 2025. The Office of Innovation was codified into law as the New Jersey Innovation Authority on January 5, 2026, and by February 2026 the state reported roughly 20,000 employees — about a fifth of the roughly 70,000-member workforce — had used the tool across more than 300,000 sessions and more than 1,000,000 prompts, at a reported cost near one dollar per user per month against roughly twenty dollars for commercial licences.
The focal deployment is the Department of Labor's unemployment-insurance operation, which the state's chief innovation officer describes as having reached full training and tool coverage of its entire staff — the first agency-wide saturation of the assistant. Staff there used it to rewrite claimant-facing unemployment-insurance emails in plain language, and the state reports that claimants responded 35 percent faster, speeding benefit processing. That figure carries no published methodology and predates the statewide launch; it was cited at launch as a pilot result. The unemployment-insurance, temporary disability insurance (TDI) and family leave insurance (FLI) teams then built plain-language glossaries, reusable prompt libraries and quality-evaluation rubrics for translation into Spanish and Haitian Creole, with human review by professional translators, subject-matter experts and seven community organizations. The state reports a higher share of Spanish-language unemployment-insurance applications and better follow-through, unquantified; US Digital Response gave the department a 2025 SEED Award and republished the materials for other states. The governance regime around all of this is unusually documented. Joint Circular 25-OIT-001 requires human review of all AI-generated content for accuracy, bias, completeness, accessibility and style; permits sensitive personal information only inside state-approved tools, naming this one, with Agency CIO approval; and requires State Chief Technology Officer clearance plus registration for any resident-facing or decisional generative system — a gate this staff-facing deployment has not been reported to pass, and a line it has stayed beneath. On training, the circular's own text says all state employees "should" take the course. That is should-language rather than must-language, in the governing document itself, even as access onboarding routes through the course and the curriculum has spread to 25 or more states and local partners.
The oversight loop visibly fired more than once. New Jersey surveyed its own public-sector workforce about generative AI — described as a national first — before the AI Task Force's November 2024 report to the Governor shaped its recommendations. The interim 2023 policy was superseded by the fuller 2025 circular. And after surveying hundreds of users and conducting interviews, the state relaunched the assistant in March 2026 with a visible reasoning section added explicitly so employees can catch errors, verify logic and spot hallucinations and bias, alongside voice input, prompt editing, retry, history and in-tool training guides. What the record does not contain matters as much as what it does. Every outcome figure — the 35 percent response gain, the satisfaction figure above 80 percent from in-tool feedback, the reported multi-million-dollar annual avoided cost — is state self-reported, with no inspector-general evaluation, external audit or peer-reviewed causal study of the tool located, and no documented harm incident, union grievance or litigation specific to it either. An independent June 2026 analysis credits New Jersey's worker-consultation-first governance while placing the rollout in a national pattern where deployment outpaces governance. Two adjacent New Jersey successes are deliberately kept out of this ledger: the call-center modernization that cut queues from more than thirty minutes to about two, and the earlier human-centered form redesign that cut Spanish unemployment-insurance form completion from more than three hours to twenty-eight minutes, belong to separate programs and are not attributable to this assistant. The large public-feedback counts sometimes quoted nearby belong to a cross-site feedback-widget program deployed over two years, which the assistant helped analyze.
The sociotechnical reading
Most copilots in this atlas arrive as products: a vendor's model, a vendor's interface, a vendor's claims, and an agency deciding how far to defer. This one arrives as infrastructure the deploying government mostly owns, and the map changes shape accordingly. The vendor boundary shrinks to a single crossing — the hosted-model round trip — and everything on the near side of that crossing is the state's own to change. That is what serving-layer ownership buys: model portability, no procurement lock-in on the application, logs the state can read. It is also what it does not buy. The model layer still belongs to someone else, prompts carrying claimant personal information — admitted by the policy's own design, under agency approval — still cross that one edge, and "state-built" has to be said precisely: no vendor application layer, not no vendor.
The structural novelty is the practice library. The glossaries, prompt libraries and rubrics do four jobs at once: operators author them, the model's prompts are anchored by them, the workforce is trained out of them, and other states import them. Run forward, this is the deskilling dynamic with its sign flipped — the tool sits behind a curriculum, the curriculum is written by the practitioners, and after the rebuild the training lives inside the tool, so practice and product co-evolve instead of one hollowing out the other. The caution travels with the compliment. A store that steers every letter is a concentration as well as a strength, because a wrong glossary entry propagates statewide wearing the library's borrowed authority, and because the same rubrics that anchor the prompts are the instrument the reviewers judge the output against — the standard and its own check come out of one place. A mid-deployment engine swap quietly re-opens every judgement the library thought it had settled.
The second lesson is where the error channel sits. This system touches no eligibility determination and issues no score. Its output is claimant-facing text, so its failure mode is a badly translated notice or a confidently wrong paragraph in a benefits letter — and the deployment's strongest control, the translator-and-community review chain, stands exactly on that channel, which is where a control should stand. The third lesson is the honest one about evidence. This is a governance-rich record: a worker survey before the policy, a circular superseded as practice matured, a rebuild driven by user research whose headline feature exists to help operators catch hallucinations. It is simultaneously an evidence-poor one, because every number saying the tool works was produced by the institution that built it, the training obligation is a "should" in the circular's own text, and the statewide workforce beyond the saturated focal agency lives on the soft side of that word. A deployment can be genuinely well-governed and entirely self-graded at the same time; this case is the atlas's cleanest example of both at once. The served claimants whose letters got easier to read are, as everywhere in this collection, outside the model — the reported rise in Spanish-language applications is the state's own unquantified observation, recorded here as exactly that.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library.