Massachusetts DTA call summaries

This network follows the pattern the Massachusetts DTA call summaries case file describes. It does not rebuild the actual pipeline. DTA withheld the summarization prompt as proprietary. So how the tool turns a call into a summary is not public, and the network does not guess at it.

The network includes a check, Summary checked against the call, that would compare each saved summary with the call it came from. DTA's design keeps the editable summary and discards the full transcript. So the check has nothing left to read. That is the documented design, not a review someone skipped. DTA presents not keeping transcripts as a privacy protection. So a privacy control and an accountability control rest on the same choice.

The network draws no link from BEACON back into the summarizer. The summarizer reads the live call. The sources describe no use of past case records in a summary and no retraining on them. The tool's documentation listed its plan for interaction data as still to be defined while it was in production. Most documentation tools in the Lab draw that link back, and a few do not. Drawing it here would claim a reuse the sources do not describe.

The network shows two groups of workers because the sources describe two groups with different access to the facts. The caseworker took the call, heard it, and may edit the summary. Determination, recertification, and hearings staff later read that summary as the account of an interview they did not hear. That difference between them is the case.

The network assumes the whole two-party call goes into the summarizer unredacted, Social Security numbers, medical history, and immigration status included. The sources describe no filter or redaction between the call and the summarizer. For the same reason, the network marks that link, and the saving of the summary into BEACON, as privacy-sensitive.

The network draws the two links between the caseworker and the summarizer from the documented steps, not from call volume. The draft summary stands in for the note the caseworker would otherwise write, which is DTA's stated purpose for the tool. Two steps limit it: the caseworker may edit the draft, and use is voluntary under the union agreement. The caseworker's own link to the summarizer carries two choices. They choose which calls go through the tool, and they edit before saving. The caseworker does not shape how the summary is written, because the prompt is withheld.

The network assumes every call that goes through the summarizer ends in BEACON. On save, the summary becomes part of the client's case record as the account of the interview. Staff also write their decisions back into BEACON. What they write is a decision reached from the summary, not the interview account itself.

The network includes two further links that the documented design does not use. One is the summarizer saving into BEACON without a caseworker's save. DTA's design sends every summary through that save. It is included because the pressure Autonomy expands would act on it. The other is a later reader putting a disputed line back to the caseworker. The sources say no one downstream can compare the summary with the call after the save. They say correcting the record falls instead on the client's memory against the state's system of record.

The network assumes later decisions lean on the saved summary, from documented scale and documented reliance. The summary is the lasting account of the interview. Future determinations, recertifications, discrepancy checks, and fair hearings consult it. In December 2025 there were 31,390 SNAP application decisions and 56,454 recertifications due. Monthly churn was 24%, and a new application took 12 days on average to approve.

The network assumes the tool spreads slowly between caseworkers, because the pilot is small. Use is voluntary under the union agreement, so the network assumes the tool spreads by colleague example. About 400 calls had gone through it between the December 2025 rollout and the April 2026 reporting. In December 2025 alone, 45,703 callers were connected to staff. That is well under 1% of connected calls. This is a pilot with room to expand, not a deployment already in full use.

One pipeline and one withheld prompt handle every call in the pilot, and the pilot is small. So a weakness in the transcription or the prompt repeats across calls instead of averaging out. The callers speak many languages. After English, the most common are Spanish, Haitian Creole, Chinese, Portuguese, and Vietnamese. There are 263,828 recipients aged 60 or over and 308,652 with a disability. So a uniform pipeline's weaknesses fall on the same groups each time. That makeup of the caseload is documented. No difference in harm between groups is computed here.

The oversight part of the network is the Executive Office of Technology Services and Security (EOTSS), which owns the state's generative-AI policy and its internal AI inventory. The summarizer appears to EOTSS as an inventory entry, not as a feed of its summaries for review. It is one of nine uses disclosed from an inventory of at least 40. None of the nine, this one included, reported a completed privacy impact assessment, though the inventory had fields for one. The tool's documentation listed its plan for interaction data as still to be defined while it was in production. So the assessment link from EOTSS does nothing in the deployment as documented.

The network does not show the union, SEIU Local 509, as a separate part. Its agreement with DTA made use voluntary for workers and added job protections. It reviewed no summaries and did not address record provenance or safeguards for callers. Its documented effect is the caseworker's choice of which calls go through the summarizer. So the network shows it as part of the caseworker's link to the summarizer, not as a reviewer. The sources do not establish that the agreement came before the December 2025 rollout, and the network does not assume it did.

The callers are not part of the network. DTA says callers are notified before connection and can opt out, but no opt-out figures are published. That step is recorded in the case file, not drawn as part of the network.

The network shows one link by which records leave the state. While the pilot ran, BEACON's records were under a demand from the U.S. Department of Agriculture (USDA) for personal data on SNAP applicants and recipients. In California v. USDA, No. 3:25-cv-06310 (N.D. Cal.), preliminary injunctions of October 15, 2025 and February 27, 2026 blocked USDA from cutting funding over the states' refusal. The court held that the proposed data protocol would likely allow sharing beyond the entities permitted under 7 U.S.C. 2020(e)(8). Both orders are preliminary, and the litigation is live. Whether AI summaries in BEACON fall within the demand is unresolved, and the state Attorney General's office declined to comment. Two other ways out are not drawn. The sources document no caseworker pasting call details into an unapproved tool. DTA says summary data is held in state-owned systems, which is itself an agency claim. The state's separate enterprise AI contract is a different matter and is kept apart from this pilot.

The network draws no second model checking the summarizer. None is documented or planned, and no independent evaluation of the tool has been published. A second model would also face the same discarded transcript. So the summarizer's only link to itself is the one prompt used on every call. The network leaves out links the sources do not describe.

The network assumes the Assistance Line is under strain, and that caseworkers did the work competently before the tool. In December 2025 the Assistance Line connected 45,703 callers to staff. Each day, on average, 2,709 calls were connected, 5,427 were completed in the automated phone menu, and 6,454 callers could not connect at all. The program served 1,011,460 people in 622,837 households, one in six Massachusetts residents. The work without the tool is not assumed to have been better. Nothing documents the tool replacing a clearly better process, and no evaluation exists either way. The manual process also kept no verbatim record: caseworkers wrote notes from memory. What the tool changes is who writes the account, not whether a recording survives.

The agency's claimed benefits are unmeasured and are treated as claims. DTA says the tool cuts call handling times, makes case notes more consistent, and frees caseworkers to focus on the conversation. It also says data is stored in state-owned systems, follows existing access and retention policies, and callers can opt out. No independent evaluation, inspector-general audit, or completed privacy impact assessment of the tool is on record. Edit rates and edit depth at the caseworker's save are unpublished. The network computes no benefit figure from these claims.

The picture rests on one outlet. The finding that transcripts are not kept, the call count, the withheld prompt, and the blank assessment fields all come from The Shoestring's April 2026 investigation. It cites technical documentation and DTA statements on the record. The caseload, call, decision, language, age, and disability figures come separately from DTA's own December 2025 performance scorecard. The scorecard notes that its churn rate was recently updated after a computational error. No court or agency has found that any individual was harmed. The concern this network shows is structural and about the future.

Different effects on different groups of callers are documented, never computed. The caseload is disproportionately elderly, disabled, and non-English-speaking. Transcription quality across those groups has not been measured. Because the transcripts are discarded, the evidence that could measure it is discarded with them. The case file records this. The Lab shows how mistakes pass between the organization's parts, and estimates no harm to the people served.