From the Health & Human Services edition of November 6, 2026
JMIR scoping review finds only eight practice reports of AI in high-income public health outside outbreaks, with mixed effectiveness and a large gap versus speculative use cases
University of Exeter Medical School et al. (NIHR-funded) · Public health practice research · High-income countries (UK-focused relevance; included US WIC chatbot and other international practice reports)
- Publisher
- Journal of Medical Internet Research
- Original publication
- 2026-10-29
- Source retrieved
- 2026-11-07
- Event date
- 2026-10-29
What happened
After screening 5,400 records and 155 full texts, Harrison et al. included only eight reports of AI tools actually used in public health practice in high-income contexts excluding outbreaks. Uses spanned chatbots, demand forecasting, symptom checking and evidence synthesis; effectiveness evidence was mixed (helpful WIC chatbot vs WHO S.A.R.A.H judged potentially trust-eroding). The authors conclude the thin practice evidence base makes it hard to know which AI uses to adopt and urge local evaluation and accessible dissemination—highlighted again in Nov 5 public-health AI roundups.
Why it matters
State and local public health AI procurements often cite promise more than practice evidence. This review is a cautionary control for SLED buyers comparing vendor pipelines to evaluated deployments, including U.S. WIC chatbot experience among the few practice reports.
Evidence and measured results
J Med Internet Res 2026;28:e108250, published Oct 29, 2026. Searches to July 24, 2025 across MEDLINE, Embase, Web of Science, Scopus. Eight included reports: chatbots (3), ML demand forecasting (1), symptom checking (2), evidence synthesis (2). Mixed outcomes explicitly noted; CareCall expanded reach but increased worker callbacks; S.A.R.A.H criticized; Maya WIC chatbot viewed as helpful. Authors contrast 8 practice reports with 44 potential uses in a linked review. NIHR funded; AI not used in the review process.
Limitations and uncertainty
UK-led relevance filter; outbreaks excluded by design so surveillance AI is out of scope. Publication Oct 29 precedes this edition's last HHS run but was not archived; included for scrutiny pass and fresh Nov 5 secondary attention. Effectiveness assessment was out of scope for the review itself.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-11-06; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Public health leaders face vendor and political pressure to adopt AI without transferable practice evidence outside outbreaks.
- Stakeholders
- State/local public health directors
- evidence/evaluation units
- procurement officers
- Discovery questions
- What practice evidence exists for this tool in a comparable agency?
- What outcome will you stop the pilot on?
- How will results be shared beyond the vendor?
- Value hypothesis
- Structured local evaluation against operational outcomes reduces costly scale-ups of tools that look strong only in benchmarks.
- Potential engagement
- Evidence-gap assessment and pilot evaluation design for one candidate use case.
- Unsupported claims
- That eight reports prove AI fails in public health
- That outbreak AI evidence transfers automatically
- Vendor performance guarantees from this paper
Pre-sales engineering
Role takeaway
- Fit
- Evaluation instrumentation and narrow-scope pilots rather than enterprise AI suites justified only by potential-use catalogs.
- Architecture and integration
- Logging of AI outputs vs human decisions, sampling for quality review, dashboards for workload and escalation.
- Prerequisites
- Baseline metrics
- Ethics/privacy review
- Dissemination plan
- Constraints
- Outbreak tools out of scope for this evidence base
- High-income focus may miss LMI lessons
- Publication lag vs fast model change
- Security
- Evaluation datasets and chatbot logs need the same protections as production PHI/PII.
- Proof of value
- Pre-registered pilot outcomes
- Trust/workload measures alongside accuracy
- Public or shared after-action
Delivery
Role takeaway
- Work
- Select one use case, run time-boxed pilot with comparison, publish accessible findings for peer agencies.
- Dependencies
- Academic or epi evaluation partners
- Vendor cooperation on logs
- Leadership air cover to stop bad pilots
- Ownership
- Public health science/evaluation lead with program ops co-owner.
- Skills and adoption
- Staff need appraisal skills for AI-specific harms and equity effects.
- Governance checkpoints
- Protocol registration
- Interim futility review
- Scale/stop decision gate
- Acceptance criteria
- Documented effectiveness and cost signals
- Harm/trust monitoring
- Reusable methods note
- Risks
- Pilot theater without power to stop
- Overgeneralizing UK findings
- Ignoring outbreak-excluded domains incorrectly
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Prefer narrowly scoped tools with evaluation hooks (logging, human comparison, stop criteria) over broad 'AI-ready' platforms lacking practice evidence outside outbreaks.
Governance
Who approves, reviews and stays accountable for outcomes?
Treat 'AI-ready' as an evidence question: require prospective evaluation plans, comparable-population performance and operational outcome metrics before scaling.
Security and privacy
What data, permissions and controls need testing?
Practice reports include chatbots handling benefits and health concerns; the review does not audit security but underscores trust risk when tools fail publicly.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Thin evidence and mixed chatbot results imply workforce training and non-digital alternatives remain essential; CareCall showed reach gains can raise callback load.
Procurement
What should contracts, pricing and exit terms secure?
RFPs should demand practice evidence or funded evaluation, not only technical benchmarks or 'potential use' lists.
Operating model
Which teams own the service once it runs?
Build evaluate-before-scale into AI operating cadence for public health, especially outside outbreak tools.
Publication history
- 2026-11-06Health & Human Services · Issue 094 resources
Stable resource ID: jmir-harrison-ai-public-health-practice-scoping-review-20261029