{"resourceId":"indiana-dwd-gemini-plain-language-ui-notices-20261007","versions":[{"version":"external-295041f8c7945c983c548953389eec71d9973b24d0841ff45421d4dd58ef01e4","resource":{"id":"indiana-dwd-gemini-plain-language-ui-notices-20261007","title":"Indiana's workforce agency has run about half its documents through a Gemini-based plain-language tool, with 58 rewritten documents now public and human review of every draft","organization":"Indiana Department of Workforce Development (DWD), Office of Gov. Mike Braun; tool built with Resultant","sector":"State unemployment insurance (UI) claimant communications","geography":"Indiana statewide","publishedAt":"2026-10-07","publicationDate":"2026-10-07","eventDate":"2026-10-07","sourceName":"Indiana Department of Workforce Development","sourceLabel":"Gov. Braun Makes Government More Efficient with AI-Powered Tool (DWD newsroom, Oct 7, 2026)","sourceUrl":"https://www.in.gov/dwd/newsroom/blog/gov-braun-makes-government-more-efficient-with-ai-powered-tool","evidenceClass":"government-evaluation","outcomeClass":"emerging","topics":["knowledge-work","accessibility-workforce","governance-procurement"],"finding":"On Oct 7 DWD said its AI plain-language tool, built on Google's Gemini model and configured for DWD over several rounds of testing, has processed about 50% of its documents, and 58 rewritten documents are now in public use. Staff generate a first revision with the tool and every document goes through human review for accuracy, meaning and compliance before publication. The first phase covers high-volume UI materials, including notices generated by the Uplink system, some of which have more than 100 versions depending on a claimant's situation. DWD is also moving decisions, benefit amounts, deadlines and next steps to more prominent places and adding plain question headings such as 'What if I disagree?' for appeals. DWD says the changes are expected to reduce avoidable calls, filing errors and unnecessary appeals and that it plans to track those outcomes, but it reports no results yet. An April 2026 DWD post said the tool was developed with Resultant, that DWD has 750 distinct documents, and estimated manual conversion at 13 years versus about two years with AI.","sledRelevance":"A concrete, narrow state use of generative AI on a high-stakes public channel: legal notices that tell residents about benefits decisions and appeal rights. It shows the drafting-plus-human-review model at production scale and gives a baseline (750 documents, 13-year manual estimate) against which progress can be checked.","evidence":"Opened DWD's Oct 7 newsroom post (article:published_time 2026-10-07) and its April 6, 2026 post on the same tool (curl, BeautifulSoup). WBIW republished the Oct 7 release text. Compared with an archived Nevada Taxation NASCIO entry on AI plain-language notices for context only.","architectureImplications":"A configured Gemini model produces draft rewrites outside the case system; reviewed text is then loaded back as templates, including templates generated by the Uplink UI system. Because a single notice can have 100+ variants, the template logic (which paragraphs appear for which claimant) matters as much as the rewritten wording.","governanceImplications":"Every rewritten document gets human review for accuracy, meaning and compliance before release, which matters because UI notices carry legal rights and appeal deadlines. DWD doesn't say who reviews (legal staff, UI policy staff), how disagreements are resolved, or whether review findings are recorded. Outcome tracking is planned, not reported.","securityPrivacyImplications":"Limited relevance: the tool rewrites document templates, not individual claimant files, per the release. DWD doesn't say whether any claimant data reaches the model or what data terms apply to the Gemini configuration.","caveats":"Agency press release with no independent check. '50% of its documents' has no stated denominator; if it refers to the 750 documents in the April post, most processed documents haven't been published yet (58 are public). The 13-year versus two-year estimate is a projection, not a measurement. No readability scores, call volumes, error rates or appeal data are given.","streamIds":["state-government"],"roles":{"sales":"Indiana shows a bounded, low-drama generative AI use: rewriting benefits notices in plain language with people reviewing every draft. The opening for other states is help with the hard parts DWD hasn't reported yet: a document inventory, review workflow, and measuring calls, filing errors and appeals before and after. Don't promise call reductions or faster processing; Indiana gives an estimate and an intent to track, not results, and the 50% figure has no clear base.","engineering":"The pattern is model-drafted, human-approved templates, not a live chatbot. The engineering work is in the template layer: notices generated by the UI system can have more than 100 variants, so rewritten text has to map onto the conditional logic that assembles each letter. Check whether any claimant data reaches the model, version every template, and keep the legal source text linked to its plain-language version so reviewers can compare them.","delivery":"Delivery is a content program with an AI drafting step. Set up an inventory of document families, a reviewer roster with legal sign-off, and a record of what reviewers changed in each AI draft. Before release, capture baselines for calls about each notice, filing errors and appeals, since DWD says it plans to track those. Publishing 58 documents out of roughly half processed suggests review and system loading, not drafting, is the bottleneck."},"roleDetails":{"sales":{"customerProblem":"Benefits notices written in legal language drive confusion, calls and avoidable appeals, and rewriting hundreds of templates by hand would take years.","stakeholders":["UI operations chief","Agency general counsel","Agency communications and content staff","State CIO or IT office for the UI system","Governor's efficiency office"],"discoveryQuestions":["How many distinct notice templates and variants do you have?","Who must approve rewritten legal notices?","Do you have baseline call, filing-error and appeal data by notice type?","Can rewritten text be loaded into your UI system's templates without code changes?"],"valueHypothesis":"AI-drafted plain-language templates with legal review could shorten a multi-year rewrite, if review capacity keeps up.","potentialEngagement":"A pilot rewriting one notice family with baseline and follow-up measures of calls and appeals.","unsupportedClaims":["That Indiana's rewrites reduced calls, errors or appeals: none reported yet","That the 13-year to two-year estimate has been achieved","That AI output is accurate without human review"]},"engineering":{"fit":"Fits agencies with large libraries of templated legal notices and a separate review step.","architectureIntegration":"Generate drafts outside the case system, then load approved text into the UI system's conditional templates with version control.","prerequisites":["Inventory of templates and their conditional variants","Plain-language style guide","Legal review capacity"],"constraints":["Notices carry legal rights and deadlines, so wording must stay legally accurate","Template systems may need changes to support new layouts","Translation depends on separate services"],"security":"Keep claimant records out of the drafting step and confirm model data terms.","proofOfValue":["Readability scores before and after for a sample of notices","Reviewer correction rate on AI drafts","Call and appeal counts for a rewritten notice versus its old version"]},"delivery":{"work":"Run a phased rewrite of notice families with AI drafting, human review and template loading.","dependencies":["Legal and policy reviewers","UI system template access","Baseline service data"],"ownership":"UI operations, with general counsel approving legal content.","skillsAdoption":"Content staff need plain-language and prompt skills; reviewers need a shared checklist.","governanceCheckpoints":["Legal sign-off for each document family","Accessibility check before publication","Outcome review after each phase"],"acceptanceCriteria":["Proposed: each published notice has a recorded reviewer approval","Proposed: readability target met for each rewritten notice","Proposed: calls and appeals tracked for six months after release"],"risks":["AI rewording changes legal meaning","Review becomes the bottleneck","Outcome tracking never gets set up"]}},"retrievedAt":"2026-10-07T22:18:00-05:00","enrichedAt":"2026-10-07T22:18:00-05:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Plain language and clearer layout aim to make UI decisions easier to understand, and DWD says simpler English text could support future Spanish and other translations. The work keeps staff as editors and reviewers rather than replacing them; staffing effects aren't discussed.","procurementImplications":"The tool was built with a consulting firm (Resultant) on a commercial model (Gemini). Agencies buying similar work should expect to pay for configuration and testing rounds, and should require the vendor to support measurement of downstream outcomes, which DWD hasn't yet reported.","operatingModelImplications":"The UI operations team owns the rewrite-and-review cycle in phases: UI determinations first, then the claimant handbook, employer materials and web content. That phased, document-family approach is easy to copy but needs a content owner for each family.","sourceVerification":{"responseId":"curl-in-dwd-oct7-20261007","openedUrl":"https://www.in.gov/dwd/newsroom/blog/gov-braun-makes-government-more-efficient-with-ai-powered-tool","referenceExcerpt":"Staff use the tool to produce an initial revision, but every document still goes through human review for accuracy, meaning and compliance before publication.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}