Strategic PartnersIssue 019 records
AWS
AWS SLED edition, October 29, 2026. GovCloud model options widened with Claude Haiku 5.5 and Grok 4.7, but no public-sector evaluations accompany them. Texas presented its SAM regulatory assistant as an AI service in AWS GovCloud, while Richmond's Amazon Connect nonemergency line handles about 19% of calls without a dispatcher against a 50-60% goal. AWS guidance for HHS eligibility, EdTech and higher-ed assistants consistently keeps humans on final decisions and puts guardrails and evaluation before scale. Evidence is thin: seven of nine items are vendor material.
- Evidence records
- 9
- Cross-source patterns
- 3
- Evidence classes
- 8 vendor claim1 independent reporting
- Outcomes
- 8 emerging1 cautionary
- Source freshness
- 7 recent2 new this fortnight
- Research completed
- 2026-10-30
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
GovCloud model choice is widening faster than agencies can evaluate it
Within two weeks AWS announced Claude Haiku 5.5 and Grok 4.7 in GovCloud, while partners pitch multi-model sandboxes to compare foundation models. None of these sources include public-sector evaluation results; the burden of evaluating each model on agency workloads falls to the buyer.
Operating questionDo we have a standing, repeatable process to evaluate and approve each new GovCloud model on our own tasks before it reaches production?
Supporting evidenceAWS announces Claude Haiku 5.5 availability on Amazon Bedrock in AWS GovCloud (US)Amazon Bedrock adds SpaceXAI Grok 4.7 in AWS GovCloud (US), pitched for coding, agentic tasks and portal-navigation browser agentsIT-CNP (AWS partner), described on the AWS Public Sector Blog
Public-facing assistants: designs stress guardrails, while the only measured outcome falls short
Texas's SAM deck and AWS's student-assistant guide emphasize scoped guardrails, grounding and refusals, but report no outcome metrics. The one item with disclosed operating figures, Richmond's Amazon Connect nonemergency line, shows about 19% handled without a dispatcher against a 50-60% goal.
Operating questionWhat containment, accuracy and disclosure metrics will we commit to measuring from day one of a public-facing assistant?
Supporting evidenceTexas Regulatory Efficiency Office (Office of the Governor); presented at Texas DIR DiscoverCity of Richmond Department of Emergency Communications (reported by Hoodline)AWS guide to building a grounded self-service student assistant with Bedrock Knowledge Bases, S3 Vectors and Guardrails
Humans keep final decisions in AWS SLED guidance
AWS's eligibility guidance keeps denials and edge cases with caseworkers and expands autonomy only on measured accuracy; its EdTech framework puts governance before scaling; Sanoma's teacher assistant leaves grading decisions to teachers. Across HHS and K-12, AI drafts and humans decide.
Operating questionFor each AI workflow, which decisions remain human, and what measured evidence would justify changing that boundary?
Supporting evidenceAWS outlines a human-in-the-loop autonomy model for agentic SNAP and Medicaid eligibility processing on Bedrock AgentCoreAWS applies its CAF-AI framework to EdTech, urging governance, security and operations maturity before scaling learner-facing AISanoma Learning (described on the AWS Public Sector Blog)
Evidence ledger
Full record · every source keeps its link and limitations
Texas Regulatory Efficiency Office presents SAM, a citizen-facing regulatory assistant hosted as an AI service in AWS GovCloud
The October 6, 2026 DIR Discover deck describes SAM on efficiency.texas.gov as a public, no-login assistant that helps Texans find which agency, rule and form applies, citing a corpus of 44,000 regulations in the Texas Administrative Code, plus statutes, agency forms and guides. The security slide states the AI service runs in AWS GovCloud on a texas.gov site with FedRAMP noted; guardrails say SAM does not give legal advice or decide cases, and answers point users to the linked agency source. This is the first explicit public statement located tying SAM's hosting to AWS GovCloud; the May 2026 launch is already archived (texas-treo-sam-regulatory-chatbot-statescoop-20260515), where the director described orchestration with multiple commercial models.
Why it matters, evidence and limitations
- Why it matters
- State-government exemplar of a citizen-facing, citation-grounded regulatory navigation assistant on AWS GovCloud; directly comparable to permitting and licensing assistants in other states and large cities.
- Evidence and measured results
- Presentation slides only: architecture claims (GovCloud, FedRAMP, no accounts), scope (44,000 regulations) and guardrail lists. No usage counts, accuracy rates, evaluation method, cost or incident data are presented in the deck.
- Limitations and uncertainty
- Operator/vendor-presented material (file name references the implementing firm Vulcan); not an independent evaluation. Slide text is terse and partly image-based. No outcome metrics. Model provenance unclear relative to earlier statements.
Richmond's Amazon Connect nonemergency AI handled about 19% of calls without a dispatcher against a 50-60% goal, drawing council scrutiny
Hoodline, relaying Axios Richmond, reports the city's Amazon Connect system transferred nearly 81% of roughly 111,000 nonemergency calls to dispatchers last fiscal year, handling or routing elsewhere only 19.3%. The city disclosed a 50-60% goal in a council presentation, set an interim target of at least 25% by January 1, 2027, is working with Amazon Connect on changes, and says it will look at other options if goals are missed. Reported cost to date exceeds $52,000. A council member said residents may not know they are speaking with AI. The article also cites Virginia Beach's similar system transferring 40.6% of 237,207 calls to telecommunicators and 4.4% to 311 (April-November 2024). The underlying Axios report is already archived (axios-richmond-dec-ai-nonemergency-60pct-goal-81pct-transferred-20261014); this item adds the AWS-stream view, the Virginia Beach comparison and state disclosure-legislation context.
Why it matters, evidence and limitations
- Why it matters
- Directly relevant to local governments and 911/311 centers evaluating Amazon Connect AI agents to relieve dispatcher workload; shows the gap between ambition and realized containment.
- Evidence and measured results
- Secondary reporting of city-disclosed figures (81% transferred, 19.3% handled, about 111,000 calls, more than $52,000 cost, 25% interim target); Virginia Beach figures attributed to a cited source. No call-level accuracy evaluation; Axios's own review reportedly found inconsistent responses.
- Limitations and uncertainty
- Secondary source relaying Axios and other outlets; figures were not independently verified here, and definitions of 'handled' may differ between Richmond and Virginia Beach. Not a controlled evaluation.
AWS announces Claude Haiku 5.5 availability on Amazon Bedrock in AWS GovCloud (US)
AWS states Claude Haiku 5.5 is now offered in AWS GovCloud (US) through Amazon Bedrock, positioned for subagents and high-volume, cost-sensitive work such as voice agents, live support, classification, summarization and document field extraction. The announcement relays Anthropic's claim that it costs around 75% less than Claude Haiku 4.5 for most tasks and notes it is the first Haiku with effort controls. It frames multi-agent use where a larger model plans and hands work to Haiku.
Why it matters, evidence and limitations
- Why it matters
- Agencies and institutions requiring GovCloud hosting gain a lower-cost model option for contact-center, document and agent workloads.
- Evidence and measured results
- Vendor announcement only. Cost reduction is attributed to Anthropic, not measured on public-sector workloads. No benchmark data, FedRAMP authorization detail for this specific model, or pricing table in the page text.
- Limitations and uncertainty
- Vendor marketing claim; cost figures from the model provider; no independent evaluation; authorization status for this model not stated on the page.
Amazon Bedrock adds SpaceXAI Grok 4.7 in AWS GovCloud (US), pitched for coding, agentic tasks and portal-navigation browser agents
AWS announces that Amazon Bedrock supports SpaceXAI Grok 4.7 in AWS GovCloud (US), describing it as a frontier model for coding, agentic tasks and knowledge work. The page claims improvements over Grok 4.6 in mixed-document handling, repo-scale coding with planning and error recovery, and browser-use agents for form fills and portal navigation. Access is through the Bedrock console and APIs; Regions, pricing and features are deferred to documentation.
Why it matters, evidence and limitations
- Why it matters
- Adds another frontier model family to GovCloud, relevant to agencies piloting coding assistants and agents that navigate legacy web portals, while widening the evaluation burden.
- Evidence and measured results
- Vendor announcement; capability improvements are asserted without benchmark data in the page text. No pricing or authorization detail on the page.
- Limitations and uncertainty
- Vendor claim only; no independent evaluation; Regions and pricing not stated in the page text.
IT-CNP's AI Fusion Lab offers a Bedrock-based sandbox for government teams to compare foundation models before production
An AWS partner blog describes IT-CNP's AI Fusion Lab, a sandbox using Amazon Bedrock and other services where government teams test foundation models without affecting live systems. It lists plug-and-play access to models including ChatGPT-4o, Claude 3.5 Sonnet, Titan and Gemini, and describes a FedRAMP High and FISMA-compliant environment. An IT-CNP director quoted says the lab is offered to federal and DoD customers and helps avoid being locked into one AI pathway.
Why it matters, evidence and limitations
- Why it matters
- Illustrates the sandbox-before-production pattern that state and local agencies can adopt for model selection, though the offering itself is pitched to federal and DoD customers.
- Evidence and measured results
- Partner marketing; compliance claims are self-reported; no customer outcomes, usage figures or evaluation results are given.
- Limitations and uncertainty
- Vendor/partner claim; federal and DoD focus limits direct SLED transferability; model list appears dated; no independent verification of compliance claims.
AWS applies its CAF-AI framework to EdTech, urging governance, security and operations maturity before scaling learner-facing AI
AWS applies the Cloud Adoption Framework for AI (CAF-AI) to EdTech, arguing that Governance, Security and Operations should mature before use cases scale because users are children. It presents a progression from Explore to Experiment to Scale: basic content filters and intended-use statements; bias evaluation, model cards and age-specific guardrails; then quarterly fairness audits, a responsible AI committee with educator representation and explainability. Security spans encryption, de-identification, prompt-injection defense and automated PII discovery; operations include calendar-driven scaling and rollback runbooks.
Why it matters, evidence and limitations
- Why it matters
- Gives districts and EdTech vendors a maturity checklist that districts can use in AI product procurement and vendor due diligence.
- Evidence and measured results
- Guidance framework from the vendor; no case data, adoption figures or evaluation evidence presented.
- Limitations and uncertainty
- Vendor framework without independent validation; aimed at EdTech companies more than districts; no mention of specific student-privacy statutes in the extracted text.
AWS guide to building a grounded self-service student assistant with Bedrock Knowledge Bases, S3 Vectors and Guardrails
An AWS technical post walks universities through building a student assistant grounded in syllabi, policy PDFs, catalogs and transcripts using Amazon Bedrock Knowledge Bases. It recommends S3 Vectors for a scoped pilot with no idle cost and moving to OpenSearch Serverless only if hybrid or full-text search is needed. It advises building an evaluation set of question and expected-answer pairs scored with RAG evaluation jobs, and configuring a guardrail with denied topics such as medical, legal or mental-health advice before exposing the assistant to students. The post is tagged for AWS GovCloud (US).
Why it matters, evidence and limitations
- Why it matters
- A concrete, low-cost starting architecture for institutions wanting a grounded student-services or course assistant, with explicit evaluation and guardrail steps.
- Evidence and measured results
- Vendor how-to guidance; scenarios are described generically (career-focused universities, an HBCU) without named institutions or results.
- Limitations and uncertainty
- Vendor guidance without outcome data; GovCloud tagging does not establish that every component is available there; costs depend on usage.
AWS outlines a human-in-the-loop autonomy model for agentic SNAP and Medicaid eligibility processing on Bedrock AgentCore
AWS argues most production eligibility agents should not start fully autonomous, noting states face sanctions when SNAP determination error rates exceed 6%. It proposes three modes: Assist (agent drafts, caseworker approves each decision), Supervised autonomy (agent handles routine cases; edge cases, denials and low-confidence cases go to humans) and selective Full autonomy for case types with sustained accuracy. Cedar policies declaratively define what the agent may do, per program, such as authorizing SNAP determinations while keeping Medicaid recommend-only. Knowledge Bases, AgentCore Memory, Runtime, Gateway and Guardrails provide policy retrieval, case context, isolation, tool access and PII protection.
Why it matters, evidence and limitations
- Why it matters
- Directly applicable to state and county HHS agencies weighing agentic AI in eligibility, where error rates carry federal financial consequences.
- Evidence and measured results
- Vendor architecture guidance; no named state deployment or measured accuracy results. The SNAP 6% threshold is cited by AWS.
- Limitations and uncertainty
- Vendor guidance, published about two months before this edition; no field results; federal program rules and state policy determine what any autonomy level can legally do.
Sanoma Learning launches Sanna, a curriculum-grounded AI teacher assistant built on Amazon Bedrock, after a 1,500-teacher European pilot
An AWS customer story reports that Sanoma Learning built Sanna, a teacher assistant grounded in its own learning materials and national curricula, on Amazon Bedrock. Teachers in seven countries shaped it, and 1,500 used it in a spring pilot. Self-reported pilot results: 88 percent said output quality matched or beat their own, 75 percent used outputs directly in class, and more than 80 percent would be disappointed to lose it. Sanna soft-launched October 1, 2026 in selected European markets; a grading module will suggest scores against the teacher's criteria while the teacher makes the call.
Why it matters, evidence and limitations
- Why it matters
- Relevant to US districts evaluating publisher-embedded AI teacher assistants: it shows the curriculum-grounded, teacher-in-control design pattern and the kind of pilot metrics vendors will present.
- Evidence and measured results
- Vendor-published customer story with self-reported teacher survey results; no learning-outcome data, independent evaluation or methodology details.
- Limitations and uncertainty
- European context; US transferability depends on state standards, student-privacy laws and district procurement. Self-reported satisfaction, not learning outcomes. Vendor and customer marketing.
How to read this edition
Source findings, measured results and limitations come from the cited publications. Patterns, operating questions, role takeaways and implementation considerations are Lighthouse Advisory interpretation, stated as questions to validate locally rather than guaranteed outcomes. Vendor and operator claims are labeled as claims. Full research method.
- Vendor claim
- A supplier-provided assertion that has not been upgraded to independent evidence.
- Independent reporting
- Independent reporting with attributable sources but without a formal evaluation design.