Value is clearest in bounded, routine knowledge work
Measured and reported gains cluster around search, summarization, email, drafting, and coding support—not autonomous replacement of public judgment.
Issue 01 · Evidence briefing
SLED AI Adoption Intelligence
Evidence, outcomes, and operating lessons for public-sector AI.
A decision-oriented read of what public institutions tried, what the evidence supports, and what leaders should design for next. Vendor claims are treated as claims, not outcomes.
10evidence records
5cross-source patterns
7topic lenses
Edition intelligence
Search the record over time
Enter one or more keywords to search the evidence record.
Synthesis
Patterns across the evidence
Public-facing assistants need bounded knowledge and continuous evaluation
GOV.UK’s source-bounded, measured rollout contrasts with MyCity’s audit dispute over accuracy, testing, and inconsistent answers.
Federated delivery works best with central assurance
Agency teams need room to develop use cases, while common intake, risk tiering, architecture review, training, and production gates remain centralized.
Inventory and procurement learning remain operational weak points
Growing portfolios are difficult to govern when use cases arrive late in inventories and acquisition lessons are not captured for reuse.
Education adoption is outpacing policy and learning evidence
Student and teacher use is rising faster than consistent rules, training, and evidence that distinguishes learning support from cognitive substitution.
Full record
Evidence ledger
Showing 10 of 10 records · All
Updated August 27, 2026
Large workforce trial finds measurable—but smaller—time savings
A 3,549-person Microsoft 365 Copilot trial paired staff feedback with econometric analysis, separating perceived gains from measured time savings.
Read full analysis
What happened
A 3,549-person Microsoft 365 Copilot trial paired staff feedback with econometric analysis, separating perceived gains from measured time savings.
Evidence read
Ninety percent of participants reported saving time, while the econometric estimate was 19 minutes per person per day. Seventy-three percent perceived better work quality.
Why it matters for SLED
The gap between reported and measured gains gives state and local agencies a realistic basis for benefits cases and broad productivity-tool rollouts.
Architecture implications
Instrument adoption and time saved at the workflow level; connect approved copilots to governed document stores and preserve usage telemetry for benefits analysis.
Governance implications
Pair licensing with role-based training, baseline measures, acceptable-use rules, and a named benefits owner who can review continuation or expansion.
Security and privacy implications
Validate tenant boundaries, data-loss controls, retention, and access permissions before assistants can retrieve or summarize sensitive case material.
Limits of the evidence
The evaluation was not a randomized controlled trial, and several outcomes rely on participant self-reporting.
Public-sector developers report strong utility from coding assistants
A cross-government GitHub Copilot trial examined adoption, code acceptance, developer sentiment, and reported time savings across a large license cohort.
Read full analysis
What happened
A cross-government GitHub Copilot trial examined adoption, code acceptance, developer sentiment, and reported time savings across a large license cohort.
Evidence read
Among 1,100 activated licenses, an average 418 users were active daily and participants averaged 2,298 chats. The code-line acceptance rate was 15.8%, and 58% said they would not return to working without an assistant.
Why it matters for SLED
Government technology teams can test assistants against delivery bottlenecks while monitoring actual usage, code acceptance, security, and developer experience.
Architecture implications
Treat approved coding assistants as part of the engineering platform, with repository boundaries, supported IDEs, code review, dependency scanning, and delivery telemetry.
Governance implications
Define permitted repositories and languages, require human review, and compare delivery and quality baselines before expanding access.
Security and privacy implications
Prevent sensitive source or secrets from entering unapproved models; validate data retention, prompt handling, and vendor training-use terms.
Limits of the evidence
Time savings were survey-reported, and code acceptance is not a direct measure of code quality or service outcomes.
A public assistant improved accuracy through repeated live testing
Two GOV.UK Chat pilots used real questions, benchmark testing, and adversarial probes to improve a retrieval-based public information assistant.
Read full analysis
What happened
Two GOV.UK Chat pilots used real questions, benchmark testing, and adversarial probes to improve a retrieval-based public information assistant.
Evidence read
More than 10,000 users asked 26,000 questions; 73% rated the service useful and 64% were satisfied. Benchmark accuracy rose from 76% to 90%, 508 jailbreak attempts were blocked, the in-scope answer rate was 88%, and average response time was 10.7-second latency.
Why it matters for SLED
Public-facing assistants need a service-standard mindset: authoritative content retrieval, scoped questions, accessibility, safety testing, and ongoing quality measurement.
Architecture implications
Use source-bounded retrieval, an evaluation harness, latency monitoring, failure logging, and a service path that can degrade safely to authoritative content or human help.
Governance implications
Publish service measures, maintain escalation and content ownership, and require benchmark and red-team evidence before each material release.
Security and privacy implications
Log and analyze abuse without retaining unnecessary personal data; isolate retrieval sources and test jailbreak, prompt-injection, and disclosure risks.
Limits of the evidence
The results reflect a bounded pilot and benchmark; they do not establish accuracy for every topic or user circumstance.
Federated governance moves approved state use cases into production
Ohio combined a central AI Council with agency participation, mandatory training, risk review, procurement checklists, and data classification for a growing statewide portfolio.
Read full analysis
What happened
Ohio combined a central AI Council with agency participation, mandatory training, risk review, procurement checklists, and data classification for a growing statewide portfolio.
Evidence read
The association case study reports more than 100 approved use cases, 33 in use, and seven accessibility use cases in production.
Why it matters for SLED
A federated model can create a repeatable path from agency idea to approved deployment without forcing all domain decisions into one central team.
Architecture implications
Create common intake, reference architectures, data classification, reusable platform services, and higher-risk architecture review across agencies.
Governance implications
Define decision rights across the AI Council, agency owners, security, legal, accessibility, data governance, and procurement, with one risk-tiered production gate.
Security and privacy implications
Use data classification and security review to determine approved hosting, model access, logging, and human oversight for each risk tier.
Limits of the evidence
Deployment counts are association-supplied and do not independently establish benefit, equity, reliability, or cost-effectiveness.
A large AI portfolio exposes inventory and workforce control gaps
GAO reviewed how the IRS manages a fast-growing AI portfolio spanning operational tools, voice bots, chatbots, and development-stage use cases.
Read full analysis
What happened
GAO reviewed how the IRS manages a fast-growing AI portfolio spanning operational tools, voice bots, chatbots, and development-stage use cases.
Evidence read
The IRS reported 126 use cases by June 2025, with about one-third operational. Eleven voice bots and two chatbots handled nearly 25 million sessions during the 2025 filing season; GAO also found inventory lag, mislabeling, and workforce gaps.
Why it matters for SLED
As portfolios grow, SLED leaders need an inventory connecting each use case to an owner, lifecycle status, data, risk tier, controls, and performance evidence.
Architecture implications
Make the use-case inventory a living system of record tied to architecture review, dependencies, production status, monitoring, and retirement.
Governance implications
Assign portfolio ownership, inventory-quality controls, workforce planning, outcome reporting, and deadlines for correcting lifecycle records.
Security and privacy implications
Link each use case to its data classification, privacy impact, authorization boundary, model monitoring, and incident owner.
Limits of the evidence
The federal tax environment differs from SLED operations, and session volume does not demonstrate answer quality or citizen benefit.
AI acquisition reviews rarely capture reusable lessons
GAO examined 13 AI acquisitions at four federal agencies and found that procurement processes did not systematically preserve lessons for future buyers.
Read full analysis
What happened
GAO examined 13 AI acquisitions at four federal agencies and found that procurement processes did not systematically preserve lessons for future buyers.
Evidence read
Four agencies lacked systematic lessons-learned requirements, missing reusable learning on data rights, testing, regional model accuracy, and discontinued solutions. All four concurred with GAO’s recommendations.
Why it matters for SLED
SLED buyers face similar risks when contracts omit evaluation data, performance thresholds, portability, audit access, or an exit path.
Architecture implications
Require pre-award test plans, integration boundaries, portability, performance acceptance criteria, observability, and a documented exit architecture.
Governance implications
Use AI-specific solicitation clauses and a shared lessons repository; assign procurement, legal, data, and technical owners to acceptance and renewal decisions.
Security and privacy implications
Contract for audit access, incident duties, data use and deletion, model-change notice, subcontractor controls, and security testing evidence.
Limits of the evidence
The sample is federal and limited to 13 acquisitions; local procurement statutes and market conditions vary.
City chatbot audit finds inconsistent answers and weak test evidence
An independent city audit identified inaccurate or inconsistent chatbot responses, performance delays, and insufficient detail behind reported accuracy claims.
Read full analysis
What happened
An independent city audit identified inaccurate or inconsistent chatbot responses, performance delays, and insufficient detail behind reported accuracy claims.
Evidence read
Auditors tested system behavior and requested structured red-teaming. The Office of Technology and Innovation reported 95–99% accuracy but supplied insufficient test detail and disagreed with recommendations.
Why it matters for SLED
Public assistants create direct service and trust risk when answers are authoritative in tone but weakly tested or inconsistently documented.
Architecture implications
Build reproducible evaluation sets, performance monitoring, authoritative-source retrieval, failure capture, and a fallback path into the service architecture.
Governance implications
Require public error reporting, red-team protocols, content ownership, release gates, and independent assurance for high-impact performance claims.
Security and privacy implications
Test prompt injection, data disclosure, abusive use, and log handling as part of an auditable preproduction and recurring assurance program.
Limits of the evidence
The agency contested parts of the audit; readers should review both the findings and the response in the source.
School AI use grows faster than policy and professional learning
Nationally representative panels found rapidly growing AI use among students and teachers while training, school policy, and shared expectations remained uneven.
Read full analysis
What happened
Nationally representative panels found rapidly growing AI use among students and teachers while training, school policy, and shared expectations remained uneven.
Evidence read
Fifty-four percent of students and 53% of core-subject teachers reported using AI in 2025, each increasing by more than 15 percentage points. Training and policy lagged while stakeholder risk perceptions diverged.
Why it matters for SLED
Districts need operational guidance distinguishing instructional uses, student support, assessment integrity, accessibility, privacy, and staff responsibilities.
Architecture implications
Use an approved-tool catalog with identity, age-appropriate access, accessibility, integration, logging, and data minimization requirements.
Governance implications
Align acceptable-use rules, professional learning, assessment guidance, procurement, family communication, and outcome review at the district level.
Security and privacy implications
Apply student privacy law, parental and age safeguards, vendor data-use limits, retention controls, and non-AI access pathways.
Limits of the evidence
Survey responses describe reported behavior and perceptions, not causal effects on learning or teacher productivity.
Students see convenience and critical-thinking risk at the same time
A survey of 1,214 youth tracked growing homework use and students’ own concerns about how generative AI may affect critical thinking.
Read full analysis
What happened
A survey of 1,214 youth tracked growing homework use and students’ own concerns about how generative AI may affect critical thinking.
Evidence read
Reported homework use rose from 48% to 62% during 2025, while 67% believed greater AI use would harm critical-thinking skills. School rules often varied by teacher.
Why it matters for SLED
District policy must be legible at classroom level; inconsistent rules can undermine trust, equitable access, and meaningful assessment.
Architecture implications
Provide approved learning tools that support assignment-level disclosure, citation, accessibility, and non-AI completion paths without silent model changes.
Governance implications
Set consistent district expectations while giving educators task-level guidance; monitor learning, assessment validity, access, and student experience.
Security and privacy implications
Minimize student data, prohibit unapproved accounts, document vendor use of prompts and outputs, and preserve age-appropriate protections.
Limits of the evidence
Perceptions of harm do not prove a measured decline in critical thinking, and self-reported use may be imprecise.
Agentic systems shift state risk from answers to actions
NASCIO frames agentic AI as a move from drafting toward systems that take limited action across approvals, anomaly detection, and citizen-service workflows.
Read full analysis
What happened
NASCIO frames agentic AI as a move from drafting toward systems that take limited action across approvals, anomaly detection, and citizen-service workflows.
Evidence read
The report identifies emerging opportunities and governance questions for multi-step automation; it does not present measured production outcomes.
Why it matters for SLED
The guidance is directly aimed at state technology leaders deciding where greater autonomy is useful and where it creates unacceptable operational risk.
Architecture implications
Use least-privilege tools, bounded action spaces, approval gates, tamper-resistant logs, transaction limits, rollback, and continuous evaluation.
Governance implications
Define who can authorize agent actions, approve higher-risk steps, pause execution, investigate incidents, and retire a workflow.
Security and privacy implications
Threat-model prompt injection, tool abuse, credential scope, cross-system data movement, unsafe chaining, and audit-log integrity before autonomous execution.
Limits of the evidence
This is emerging association guidance, not measured outcome evidence; treat the recommendations as an operating hypothesis to test.
How to read this briefing
Methodology and definitions
Selection and freshness
This edition prioritizes primary government material, public audits, independent research, and relevant public-sector association guidance available for theAugust 27, 2026 run. Every surfaced item remains in the All view and keeps its original source.
Evidence classes
- Government evaluation
- A public body’s measured evaluation or documented pilot.
- Government audit
- An oversight review of performance, controls, or operations.
- Academic research
- Research produced through an academic institution or peer-reviewed venue.
- Independent research
- Research conducted outside the implementing organization.
- Public-sector association guidance
- Practitioner guidance or an association-supplied case; not independent outcome evidence.
- Independent reporting
- Independent reporting with attributable sources but without a formal evaluation design.
- Standards or public-body guidance
- Normative or advisory guidance from a standards body or public institution.
- Vendor claim
- A supplier-provided assertion that has not been upgraded to independent evidence.
Outcome labels
- Effective
- Evidence supports a useful result within the tested scope.
- Mixed
- Benefits and material limitations appear together.
- Cautionary
- The record surfaces failure, risk, or a control gap.
- Emerging
- A developing practice or claim without measured outcomes.
Claims discipline
Vendor, operator, and association claims are attributed and are not upgraded to independent evidence. Caveats identify self-reporting, bounded pilots, contested findings, and missing outcome measures.