NJ AI Assistant

What this models

This example runs on the State-built drafting assistant with an operator-authored practice library network: 9 components and 21 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 10 assumed · 4 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

• assumed

  This models the state-built, training-gated drafting-assistant pattern documented in the NJ AI Assistant case file — a communication-layer copilot whose serving stack the deploying government owns, whose practice standard its own operators wrote, and whose one external dependency is a hosted commercial model service — and not a reconstruction of the actual tool.

• assumed

  Every effect size in this case is state self-reported, promotional-grade and direction-only. The 35 percent faster claimant-response figure has no published methodology and predates the statewide launch, having been cited at launch as a pilot result; the satisfaction figure above 80 percent comes from in-tool feedback; the avoided-cost figure is the state's own. No edge weight, demand value or capacity value here is calibrated to any of those numbers, and no independent audit, inspector-general evaluation or peer-reviewed causal study of this tool was located.

• assumed

  Two adjacent New Jersey programs are deliberately absent from this network. The call-center modernization figures (queue times, abandonment rates, connection rates, cost reduction) come from a separate cloud call-center program, and the Spanish form-completion and manual-review figures come from an earlier human-centered form redesign that predates the AI materials. Neither set is attributed to this assistant, and neither is what the demand and capacity values were derived from.

• assumed

  A modest workload against ample capacity is derived, not defaulted. Demand is a steady statewide correspondence and language-access load rather than a documented backlog: the queue and abandonment figures that would read as strain belong to the separate call-center program. High capacity records that the manual process worked and still carries the legal weight — the tool is advisory and drafting-only, it issues no scores and makes no eligibility determinations, the circular makes the employee responsible for the outcome and requires independent fact-checking of every output, and the state's own benefit framing is time saved rather than work made possible.

• assumed

  The 'no vendor in the serving path' claim is scoped exactly as the record supports it: no vendor-owned application or software-as-a-service layer sits in the serving path, while the foundation model remains a hosted commercial service under state data-control terms. That round trip is drawn as a present, privacy-sensitive egress rather than a latent one, because it is the deployment's architecture and not a pressure that might open. The state demonstrated the distinction by replacing the underlying model during the 2026 rebuild while the operator-facing tool stayed the same.

• baseline

  The practice library is drawn as this network's load-bearing structure because three separate reads come out of one operator-authored store: it anchors what the model is asked, it trains both frontline classes, and it supplies the rubrics the review chain measures translations against. The protective reading and the cautionary reading are both authored from the same fact — a standard the practitioners wrote is the reverse of a vendor-held one, and a standard everything reads from is also a single point of shared authority. The write side is drawn once, from the teams that authored the artifacts; the record describes the glossaries, prompts and rubrics as durable artifacts refined in use, and what the review chain finds against the rubrics is carried as part of that refinement rather than as a separate documented write-back.

• baseline

  The two frontline classes split on documented coverage and documented wiring, not on invention. The state reports the focal unemployment-insurance operation reaching full training and tool coverage of its entire staff, which is why its drafting channel is drawn at the top of the scale; the statewide class sits under an obligation the circular's own text words as 'should', with more than 4,000 course completions reported two months after launch against roughly 20,000 users by February 2026. The focal channel carries claimant detail by policy and is second-read by the review chain; the statewide channel is screened by the bounded filter and reaches the policy layer's directive, and its feedback is what the record shows reaching the builders. That should-language observation is grounded in the circular itself and is not attributed to any commentator.

• assumed

  The bounded output screen is drawn at a low level and placed on the statewide channel. It is drawn at all because the state describes content filters and anti-jailbreak protections and because the 2026 rebuild shipped a visible reasoning section for the stated purpose of helping employees catch errors, hallucinations and bias. It is drawn no higher because both are bounded by construction and no source measures what either catches. It is placed on the statewide class because that is the population the rebuild's in-tool aids were designed around after the user research, and the class with no second-human reader of its own.

• baseline

  Two pathways are drawn without flow because sources document them as absent rather than merely unobserved: machine-drafted text entering claimant correspondence without a person in between (the circular requires human review of all AI-generated content before use, and the deployment has not been reported to pass the clearance gate decisional use would require), and an independent read of the assistant's output (every performance figure in the record was produced by the institution that built the tool). Each is drawn so the gap is visible and a lever can reach it.

• assumed

  The in-house platform team that built, hosts and rebuilt the tool is carried in copy rather than drawn as a separate operator class. The record documents its instrument as user research — in-tool feedback and satisfaction ratings feed the team, and a 2025 survey of hundreds of users plus follow-up interviews produced the March 2026 rebuild with its reasoning section, prompt editing, retry, history and in-tool training guides — and what that research produced is already on the board as the output screen and inside the practice library. The team receives no model output and adopts none, so drawing it separately would not change how error moves; the PAN organization record keeps it as a fourth user class, and this diagram deliberately narrates it instead. The other half of the consultation loop, the statewide worker survey that preceded the November 2024 recommendations, is drawn, because the record documents it producing the policy.

• assumed

  The resident feedback corpus is carried in copy rather than drawn as a separate external feed, and nothing about it is marked privacy-sensitive. The state's own anniversary post attributes the large public-feedback volumes to a cross-site feedback-widget program deployed across state websites over two years, with the assistant helping to analyze that corpus to inform plain-language content and service redesign; those volumes are never treated as this tool's own, the analysis sits on none of the paths by which this deployment's errors reach a claimant, and no source describes identifiable personal detail inside the corpus, so the silence is recorded rather than converted into an asserted exposure.

• assumed

  The republished curriculum and translation materials are carried in copy and in the case file rather than as an egress node. An egress drains the privacy gauge by construction, and publishing a training course and a glossary for other states to reuse is a diffusion of practice rather than an exposure of anyone's data. Drawing it as a boundary crossing would have priced a public good as a privacy risk.

• baseline

  This deployment's governance record is unusually dated and unusually self-graded at the same time. A statewide worker survey preceded the November 2024 report to the Governor, the interim 2023 policy was superseded by the fuller 2025 circular, and user surveys and interviews drove the 2026 rebuild whose headline feature exists to help operators catch hallucinations. Against that, no compliance measurement of the review-everything directive, no audit, and no external evaluation of the tool were located, and an independent June 2026 analysis credits the worker-consultation-first approach while placing the rollout in a national pattern where deployment outpaces governance.

• assumed

  Served claimants are not in the dynamics. The tool touches no eligibility determination and issues no score, and the state's report of a higher proportion of Spanish-language applications and better follow-through after the translation program is an unquantified state self-report recorded in the case file, never computed from this diagram. No harm incident, union grievance or litigation specific to this tool was located, which is an absence of evidence rather than a clearance.

What this example does not show

• Served claimants are not modeled here, including the limited-English-proficiency claimants the translation program serves. The Lab models institutional propagation only; the state's report of a higher proportion of Spanish-language applications and better follow-through is an unquantified state self-report recorded in the case file, never computed from this diagram.
• Every effect size in this record is state self-reported, promotional-grade and direction-only. The 35 percent faster claimant-response figure has no published methodology and predates the statewide launch, having been cited at launch as a pilot result; the satisfaction figure above 80 percent comes from in-tool feedback; and no inspector-general evaluation, external audit or peer-reviewed causal study of this tool was located.
• New Jersey's call-center transformation figures and its Spanish form-redesign outcomes belong to separate programs - a cloud call-center modernization and an earlier human-centered form redesign - and are never attributed to this assistant. The large public-feedback counts sometimes quoted alongside it come from a cross-site feedback-widget program that this tool helped analyze.
• The state-built claim is scoped exactly: no vendor-owned application or software-as-a-service layer sits in the serving path, while the foundation model remains a hosted commercial service under state data-control terms. This scenario describes capability class only and names no commercial model or interface product.
• The observation that the training obligation is worded as 'should' rather than 'must' comes from the governing circular's own text, and is not attributed to any outside commentator; no harm incident, union grievance or litigation specific to this tool was located, which is an absence of evidence rather than a clearance.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

• 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.

  empirical
  • Government NJ Office of Innovation, NJ AI Assistant (project page) (2026) https://innovation.nj.gov/projects/ai-assistant/
  • Trade press Route Fifty, How New Jersey's AI assistant saves the state time and money (2025) https://www.route-fifty.com/artificial-intelligence/2025/08/how-new-jerseys-ai-assistant-saves-state-time-and-money/407538/
  • Trade press StateScoop, New Jersey launches generative AI assistant and training tool for state employees (2024) https://statescoop.com/new-jersey-ai-assistant-training-tool-state-employees/
  • Government NJ Office of Innovation, Launched one of the nation's first AI tools specifically built for State employees (2024 Impact Report) (2024) https://innovation.nj.gov/impact-report/2024/ai-assistant/
  • Investigative Sofi, State AI Rollouts Are Outrunning Their Own Governance (TechPolicy.Press) (2026) https://www.techpolicy.press/state-ai-rollouts-are-outrunning-their-own-governance/

• New Jersey's joint policy 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 the NJ AI Assistant, with Agency CIO approval; and requires State Chief Technology Officer clearance plus registration for resident-facing or decisional generative systems, a gate the staff-facing assistant has not been reported to pass. The same circular's text says all state employees 'should' take the responsible-AI course, a should-language obligation stated in the governing document itself, while the focal unemployment-insurance agency is reported to have reached full training and tool coverage of its staff. A statewide public-workforce survey preceded the November 2024 AI Task Force recommendations to the Governor, and user surveys and interviews drove the March 2026 rebuild, which added a visible reasoning section intended to help employees catch errors and hallucinations.

  empirical
  • Government NJ Office of Information Technology and New Jersey Cybersecurity and Communications Integration Cell, Joint Circular 25-OIT-001: State of New Jersey Guidance on Responsible Use of Generative AI (2025) https://nj.gov/it/docs/ps/25-OIT-001-State-of-New-Jersey-Guidance-on-Responsible-Use-of-Generative-AI.pdf
  • Government State of New Jersey AI Task Force, New Jersey AI Task Force Report to the Governor (2024) https://innovation.nj.gov/news/NJ-AI-Task-Force-Report.pdf
  • Government NJ Office of Innovation, New Jersey Upgrades its AI Assistant for State Workers (2026) https://innovation.nj.gov/blog/2026-03-24-new_jersey_upgrades_its_ai_assistant_for_state_workers/
  • Trade press Route Fifty, How New Jersey's AI assistant saves the state time and money (2025) https://www.route-fifty.com/artificial-intelligence/2025/08/how-new-jerseys-ai-assistant-saves-state-time-and-money/407538/

• New Jersey's unemployment-insurance and TDI/FLI teams built plain-language glossaries, reusable prompt libraries, and quality-evaluation rubrics for AI-assisted 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, without quantifying either. US Digital Response gave the department a 2025 SEED Award and republished the materials for reuse by other states, the associated responsible-AI curriculum is used by 25 or more states and local partners, and the March 2026 rebuild moved training guidance inside the tool itself.

  empirical
  • Government New Jersey Department of Labor and Workforce Development, New Jersey Honored by US Digital Response for Leading AI Solutions to Improve Residents' Access to Critical Benefit Programs (press release) (2025) https://www.nj.gov/labor/lwdhome/press/2025/2025122_USDR.shtml
  • Advocacy US Digital Response, Social safety net 2.0: how New Jersey is forging a new path with language access and generative AI (2025) https://www.usdigitalresponse.org/resources/social-safety-net-2-0-how-new-jersey-is-forging-a-new-path-with-language-access-and-generative-ai
  • Trade press Route Fifty, How New Jersey's AI assistant saves the state time and money (2025) https://www.route-fifty.com/artificial-intelligence/2025/08/how-new-jerseys-ai-assistant-saves-state-time-and-money/407538/
  • Government NJ Office of Innovation, New Jersey Upgrades its AI Assistant for State Workers (2026) https://innovation.nj.gov/blog/2026-03-24-new_jersey_upgrades_its_ai_assistant_for_state_workers/
  • Trade press InnovateUS, Responsible AI for Public Professionals: Using Generative AI at Work (course page) (2024) https://innovate-us.org/course/responsible-ai-for-public-professionals-using-generative-ai-at-work/

Where this connects

Institutional pressures in this domain

• Workload surge — Demand outruns staffing; per-case attention shrinks and review becomes triage.
• Deadline pressure — Statutory or managerial timeliness rules reward fast approval of machine output over slow disagreement.
• Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.
• Staff turnover — Experienced skepticism leaves; new staff calibrate their trust on the tool itself.
• Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
• Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.

All of them in context on the Caseworker documentation & copilots domain page.

Levers available here and the patterns behind them

• Keep skills sharp — Deskilling-arrest mandate
• Keep prompts neutral — Framing and mirroring reduction
• Mark AI-written records — Provenance labeling
• Review on schedule — Oversight cadence & retrospectives
• Peer sharing rules — Peer-edge governance
• Gate vendor updates — Vendor quality gate
• Gate record entries — Human-in-the-loop write gating
• Vet connections — Connection authorization
• Store less data — Data minimization
• Check with a second model — Cross-model verification
• Upgrade model — Improve the model

Documented case histories

• NJ AI Assistant
• Magic Notes (Beam)
• Minute / Local Transcribe
• Massachusetts DTA call summaries
• Justice Transcribe
• Illinois DCFS Augintel
• GDS Microsoft 365 Copilot cross-government experiment
• DWP Whitemail Insights and Vulnerability Scanner
• UK Home Office asylum AI copilots: interview summarisation and policy search
• Learned Hand AI clerk pilot (LA and Riverside courts)
• SSA Insight
• CDTFA Axyom Assist
• VA claims automation (automated survivor-benefit decisions)
• Trelleborg's Welfare Robot
• Amsterdam Smart Check