{"resourceId":"awwa-llm-utility-communications-human-validation-2026","versions":[{"version":"external-f6a0c85f721465413e9a7cc25e80c1fce2322e9379d5bdcd17af16349c473810","resource":{"id":"awwa-llm-utility-communications-human-validation-2026","title":"Large language models are a tool, not a replacement, for utility communications","organization":"American Water Works Association","sector":"Water utilities (public communications and customer information)","geography":"United States","publishedAt":"30 Oct 2026","publicationDate":"2026-10-30","eventDate":null,"sourceName":"AWWA","sourceLabel":"Association article summarizing an AWWA webinar","sourceUrl":"https://www.awwa.org/AWWA-Articles/large-language-models-are-a-tool-not-a-replacement-for-utility-communications/","evidenceClass":"public-sector-association","outcomeClass":"emerging","topics":["knowledge-work","governance-procurement","accessibility-workforce"],"finding":"AWWA summarizes its webinar 'Applying AI in Water Utility Communications' (more than 800 registrants). Panelists say all LLM output must be validated and edited by a human because fluency does not guarantee accuracy, and drafts stay drafts until verified for accuracy and voice. Recommended practices: specific prompts describing situation, task and audience; configuring tools with the utility's brand guide and past communications; and feeding newly published material back so outputs stay aligned. Panelists also describe a 'visibility gap'. AI search answers about a utility may be inaccurate and the utility cannot see who asked or what was answered. They advise controlling the consistency of public online content (generative or answer engine optimization), since AI draws heavily on news, Q&As and earned media.","sledRelevance":"Water utility communications, including boil-water notices, rate changes and incident updates, depend on public trust. This is the sector association's current practice guidance for staff already using general-purpose LLMs.","evidence":"Association news article summarizing expert panel advice (a PR consultant, an AI consultant and a university communications professor). No measured outcomes, survey data or evaluation of accuracy are reported.","architectureImplications":"Utilities using general-purpose LLMs for drafting should prefer configurations that keep brand guides and prior communications in controlled workspaces. Published content should be structured and consistent because AI search systems reuse it.","governanceImplications":"Supports a policy that AI-generated public communications are drafts requiring human verification and approval, consistent with the archived AWWA finding that many respondents lack an AI policy.","securityPrivacyImplications":"Uploading internal communications or incident details into public LLM tools can expose sensitive information; the article does not address data handling, which utilities should cover in policy.","caveats":"Practitioner advice, not evaluated evidence. Panelists include consultants with commercial services. Emergency communications and accessibility requirements (plain language, translation) are not addressed in the article. The full webinar recording was not reviewed.","streamIds":["utility-management","citizen-engagement"],"roles":{"sales":"Sales framing: this is a low-risk entry point into utility AI. AWWA's panel tells communicators to treat LLM output as a draft that a human verifies. Discovery with water utility communications and customer service leads should ask what tools staff already use, what policy exists and how the utility monitors AI search answers about it. A bounded engagement is a communications AI policy and workflow review. Do not claim time savings or accuracy improvements, which the article does not measure.","engineering":"Engineering framing: the technical work is configuration and content hygiene, not model building. Options include enterprise LLM workspaces holding brand guides and approved past messages, retrieval limited to approved content, and audit of what is uploaded. Separately, structured, consistent public web content (FAQs, outage pages) improves what AI search systems reuse. A proof of value tests drafted notices against fact-checks and samples AI search answers about the utility before and after content cleanup.","delivery":"Delivery framing: the owner is the utility's communications manager, with legal and IT for policy and tooling. Work covers a short policy defining human review and approval, a prompt and style kit, and a routine for monitoring AI search answers. Acceptance would be proposed full human sign-off on AI-assisted public messages and a periodic accuracy check of AI search results. Risks include unreviewed posts during incidents, sensitive data in public tools and inconsistent web content."},"roleDetails":{"sales":{"customerProblem":"Utility communicators use LLMs for speed but risk publishing fluent, inaccurate content and cannot see how AI search describes the utility.","stakeholders":["Communications or public information manager","Customer service manager","General manager","Legal counsel"],"discoveryQuestions":["Which AI tools do communications staff use today, and under what policy?","Who approves AI-assisted public messages, including during incidents?","Have you checked how AI search tools answer questions about your utility?"],"valueHypothesis":"A defined review workflow and consistent public content may reduce accuracy risk in AI-assisted communications; benefits are not measured.","potentialEngagement":"A communications AI policy, prompt kit and AI-search accuracy review for one utility.","unsupportedClaims":["Measured time savings","Improved accuracy of AI search answers","Suitability for unreviewed emergency messaging","Accessibility compliance"]},"engineering":{"fit":"Good fit for drafting support with human approval; poor fit for unattended publishing.","architectureIntegration":"Enterprise LLM workspace with approved brand and message content; structured web FAQs and outage pages as the public source of truth.","prerequisites":["Brand and style guide","Approved message library","Web content management access"],"constraints":["No control over third-party AI search answers","Public LLM data handling","Incident timing pressure"],"security":"Prohibit uploading sensitive incident, customer or system details to non-enterprise AI tools.","proofOfValue":["Fact-check a sample of AI-drafted notices","Sample AI search answers about the utility before and after content cleanup","Verify approval records exist for AI-assisted posts"]},"delivery":{"work":"Policy drafting, prompt and style kit, approval workflow and AI-search monitoring routine.","dependencies":["Communications staff","Legal review","IT tool configuration"],"ownership":"Communications or public information manager.","skillsAdoption":"Train communicators on prompt specificity, verification and data-handling rules.","governanceCheckpoints":["Policy approval by leadership","Human sign-off on every public AI-assisted message","Quarterly AI-search accuracy review"],"acceptanceCriteria":["Proposed: 100% of AI-assisted public messages carry an approval record","Proposed: quarterly AI-search accuracy sample completed","Proposed: data-handling rule acknowledged by all communications staff"],"risks":["Unreviewed posts during emergencies","Sensitive data in public tools","Inconsistent web content feeding AI answers"]}},"retrievedAt":"2026-10-30T22:09:10-05:00","enrichedAt":"2026-10-30T22:16:33-05:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Communications staff need verification and prompting skills; plain-language and multilingual requirements should be checked in any AI-assisted message.","procurementImplications":"Prefer enterprise AI tools with data-handling commitments over consumer tools for utility communications.","operatingModelImplications":"Keeps human authorship and approval as the control point while using AI for drafting.","sourceVerification":{"openedUrl":"https://www.awwa.org/AWWA-Articles/large-language-models-are-a-tool-not-a-replacement-for-utility-communications/","referenceExcerpt":"Whatever AI provides you must be validated and edited by a human. Just because it sounds fluent does not mean it is factual","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}