Domain Atlas / Benefits navigation & public-facing chat
Mass.gov Virtual Assistant
Massachusetts built a retrieval-grounded generative-AI Virtual Assistant in house in 83 days and launched it on motor-vehicle registry pages in April 2025 ahead of the May 7, 2025 federal identification deadline, describing it as a state-owned platform that replaced a vendor-managed rule-based chatbot; by 2026 the assistant's own support page listed motor-vehicle, toll, unemployment-assistance, tax, child-support, transitional-assistance, family-and-medical-leave and state-login content. The state reports more than 1,500 conversations a day and roughly 200,000 conversations since launch, a chat open rate rising from 1.29 to 2.92 percent, positive feedback rising from 10 to 56 percent, negative feedback down 60 percent, page abandonment down 40 percent, around-the-clock availability in English, Spanish and Portuguese, a knowledge base refreshed nightly from published state content, and a custom evaluation tool that flags answer issues within hours. Every one of those figures is an agency self-report; the technology agency withheld the cost and usage reports and the technical logs that would confirm its claims, and the underlying foundation model and hosting stack have not been publicly identified.[5]
What happened
Massachusetts put a generative-AI assistant on its public program pages in April 2025, after an 83-day build, ahead of the May 7, 2025 federal identification deadline. The state's digital service unit built it in house inside the technology agency, describes it as a "state-owned platform", and says it replaced a vendor-managed rule-based chatbot that had been integrated with the motor-vehicle registry since 2022 in twenty languages. The new assistant answers around the clock in three: English, Spanish and Portuguese. It is retrieval-grounded — it answers from a knowledge base rebuilt each night out of published state content — and the underlying foundation model and hosting stack have never been publicly identified. It extended to the state's toll service in October 2025, and by 2026 the assistant's own support page listed motor vehicles, tolls, unemployment assistance, taxes, child support, transitional assistance, family and medical leave and the state login service "and more". The date the transitional-assistance content went live has not been made public. The state health insurance program is not on that list, and the AI in that program's record is internal rather than public-facing.
The reported numbers are the state's own. More than 1,500 conversations a day and roughly 200,000 conversations since launch; a chat open rate rising from 1.29 to 2.92 percent; positive user feedback rising from 10 to 56 percent with negative feedback down 60 percent; page abandonment down 40 percent. An independent institute brief, reporting state figures, records motor-vehicle registry calls falling by about 1,000 a day and emails by about 200 a week after launch. The state says a custom evaluation tool flags answer issues "within hours", and that chat analytics drive prompt tuning, content updates and enhancements. It also says something quieter that matters more: the published pages themselves were "updated and simplified" to suit the assistant. The authoritative public statement of what a benefit rule says is being adapted to the channel that reads it, and the nightly rebuild carries each edit into the next day's answers.
The governance scaffolding around all of this is real and dated. Executive Order 629, signed in February 2024, created an AI Strategic Task Force whose members serve in an advisory capacity with a mandate to study and advise; no intervention by it in any deployment appears in the record. Policy AI.001, effective January 31, 2025, requires human fact-checking of generative output, conspicuous labeling of AI content, Chief Technology Officer approval for generative procurement, a generative-AI inventory, and consultation with legal and security teams on privacy. The phrase "privacy impact" appears in it zero times. The Enterprise Privacy Office told the legislature in February 2025 that it "has been using a Privacy Impact Assessment" and was "overseeing a pilot program in collaboration with our risk and security teams to assess privacy risks during the contracting and development stages". The inventory exists, and it carries fields for that assessment on every entry.
What a records investigation published in April 2026 found is that the fields were blank. Records obtained by an independent outlet showed at least forty AI use cases in the technology agency's internal survey; the agency withheld all detail on thirty-one of them, plus the cost and usage reports and the technical logs that would confirm its data-handling claims. Of the nine entries released, not one recorded a completed privacy impact assessment, and the fields were left blank without explanation — including for tools that process Social Security numbers and Medicaid data. The agency spokesperson did not directly answer questions about the missing assessments. A separate outside analysis, published the same month, counted about twenty AI use cases publicly reported by the state and just three facing the public, against forty in the internal survey, and recommended a formal public AI inventory and structured user feedback loops; it noted a comparator state with roughly 120 applications in use or development. After months of negotiation the Supervisor of Records ordered the agency to produce the withheld records for in camera review, giving it ten business days to produce and taking up to fifteen to review. The month of that order is an inference from the reporting's phrasing rather than a documented date, and no final determination in the matter was found as of July 20, 2026.
Two things the state said out loud frame the rest. The technology secretary committed publicly that "there will be a human reviewing it" before public distribution, and framed AI as relieving stretched agencies without adding headcount. A public-employee union representing about 9,000 state employees, including transitional-assistance call-center workers, argued that hiring more staff, not AI tooling, is what would let the agency serve more clients. Around the same portfolio sit deployments that are deliberately not this case: a vendor-built call summarizer that has processed roughly 400 food-assistance eligibility calls since December 2025 and writes machine-made summaries into the caller's benefits record with no transcript retained, and a third-party-liability call-center tool for the state health insurance program held on the vendor's servers with its data-interaction plan still "to be defined" while in production. Those are other use cases with other shapes, and none of their facts is asserted of this assistant.
The sociotechnical reading
Most benefits-navigation cases in this atlas have a person in the middle: a caseworker who reads a score, a screener who overrides a flag, an adviser who checks a citation before relaying it. This one has nobody. The assistant talks to the applicant directly, produces navigational and eligibility-adjacent guidance rather than a determination, writes to no case record, issues no adverse-action notice and creates no appeal hook. That is the shape's whole point and its whole difficulty. An error here is not a wrongful denial anyone can contest; it is a confident paragraph about a benefit rule, delivered at scale to people who have no way to know it was wrong and no artifact through which anyone could later count how often it was. Error visibility depends entirely on tooling the operator built for itself.
The second lesson is what the deployment does to its own ground truth. The state says the published pages were updated and simplified so the assistant would answer from them well, and the knowledge base is rebuilt from those pages every night. Read as a service story, that is genuinely good: plainer public pages help everyone who reads them, and the abandonment figure moved. Read as a governance story, it is a loop with the arrow pointing the unfamiliar way. In most retrieval deployments the corpus is curated for the model and held still; here the authoritative public record of what a benefit rule says is being rewritten for a machine channel, by the same institution that reports the channel's success, with the edits landing in tomorrow's answers automatically. The audience of a public benefits page has quietly changed, and nothing in the record measures what that did to a reader who is not an assistant.
The third lesson is about a control that exists and is not exercised. This is not the story of oversight being blocked — nobody severed anyone's access, and there was no obstruction. There is a privacy office, a stated assessment practice, a pilot at the contracting and development stages, an inventory the policy requires, and a field on every entry waiting for a completed assessment. Across the nine entries anyone outside has been allowed to see, that field is blank. That is a specific and unusual failure mode: capacity present, exercise rate at zero, and no statute violated by it, because the assessment is an internal commitment that the governing policy never names. It is the reason a cadence, rather than a new instrument, is the honest remedy on this board.
The fourth is where the oversight that worked came from. Of the layers around this deployment — an advisory task force, a legislature receiving annual reports with several AI bills pending and none enacted, an internal privacy office — the only one that demonstrably changed the agency's behavior was public-records law, and it reached the documentation about the system rather than the system's answers. The state publishes favorable figures it produced itself and withheld the logs and usage reports that would let anyone check them, so the one exercised check is aimed exactly at the layer where the asymmetry lives. And the human fallback that the record treats as the safety valve is narrowing by design: about a thousand fewer calls a day is what success looks like here, twenty languages became three, and the residents who cannot use a chat channel are, as everywhere in this collection, outside the model — documented in this file, never computed from a diagram.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library.