Digital public health AI has moved rapidly from technical promise to implementation planning. Systems have been proposed for outbreak detection, syndromic surveillance, vaccination outreach, environmental risk warning, chronic disease registries, service triage, public communication, and administrative prioritization [1-4]. International guidance appropriately emphasizes transparency, accountability, safety, privacy, and equity [1,2].

These systems produce different kinds of outputs. A thresholded alert may directly create an action opportunity, a prioritization score or recommendation usually requires professional judgment before it does, and generated content or chatbot advice must first be linked to a defined need for communication, escalation, or other follow-up. A deployed dashboard, connected registry, message, or chatbot therefore demonstrates system availability, not an action change. Threshold drift, automation bias, and plausible but unsafe generated reassurance further widen this gap [5-7]. The practical question is narrower: after a defined AI output, what action opportunity existed, who was accountable, and what happened next?

We use public health action end point to describe a structured, AI output–specific specification of an existing proximal process or implementation end point, not a wholly new family of outcomes. It is prespecified, auditable, actor bound, time bound, and denominator based. Each analytic unit begins with a defined AI output and records whether a predefined action opportunity was completed, appropriately deferred, missed, unresolved, or lost from the denominator. Its distinctive contribution is the combination of output provenance, a named accountable actor, a defined action window, and a traceable action status. It is not model discrimination, calibration, user satisfaction, alert volume, dashboard use, message delivery, or chatbot fluency. Nor is it, by itself, proof of lower morbidity, mortality, transmission, hospitalization, or inequity.

Claim discipline means restricting the strength of a conclusion to the most distal link that has actually been measured. The end point is not a causal model; an AI output can precede an action without initiating, accelerating, changing, or causing it. Its role is to make that question observable and to prevent availability or output generation from being read as an implementation value or a population health benefit.

Justified nonaction should not become a convenient label for failure. Acceptable reasons can include an action already completed, a verified duplicate signal, a contraindication, an eligibility error, a person unreachable after a prespecified number of attempts, resource unavailability with a documented escalation pathway, or a human override with a recorded reason. Ambiguous or consequential classifications should be independently adjudicated using prespecified rules, with the adjudicator and disagreement resolution process reported.

Repeated outputs require a rule before data collection. When multiple outputs for 1 person or event fall within a prespecified episode or lockout window, they should usually be linked to 1 index action opportunity unless each output is expected to trigger a separate action. If 1 output can prompt several actions, investigators should define the primary action and report secondary actions separately. If multiple outputs contribute to 1 action, the action should not be counted repeatedly; the audit trail should preserve the outputs and the documented AI role. Where more than 1 window is clinically or operationally defensible, sensitivity analyses using alternative windows can show whether the interpretation is stable.

Textbox 1 summarizes the core end point specification and standard reporting categories.

Figure 1 summarizes the action pathway, attribution gate, and claim ceiling that follow from the last measured link.

A defined AI output is linked to an eligible action opportunity and an accountable workflow, after which the action status is classified as completed action, predefined justified nonaction, missed action, unresolved status, or denominator loss. Time stamps and workflow records can document temporal proximity and the reported role of AI. Attributable action change requires an appropriate comparator or causal design, and a population health benefit requires outcome-linked causal evidence. The allowable claim should stop at the last measured link. Cross-cutting safeguards for harm, workload, privacy, equity, and model or threshold drift should be monitored across all links. The conceptual layout was developed with assistance from OpenAI Codex; all scientific content and the final design were reviewed and approved by the authors.

The contribution of public health action end points is not the proposition that implementation matters. Existing implementation science, digital health reporting, quality, and AI evaluation frameworks already establish that proposition [5,8-22]. Nor do we claim that these frameworks cannot accommodate an action measure. Public health action end points instead supply a repeatable event-level specification when authors intend to claim that an AI-supported workflow changed accountable action. They can be nested within these frameworks and make explicit the provenance and denominator of the action opportunity (Table 1).

This positioning limits the framework. It is not a reporting guideline, checklist, trial design, or implementation theory. It is a structured specification of a proximal implementation or process end point. The action pathway is also not assumed to be linear; comprehension, usability, capacity, and action completion can be parallel or interacting mechanisms. The framework is useful because it requires authors to name the pathway they intend to measure instead of assuming that deployment or a downstream observation proves it.

An action end point should be developed as a pragmatic implementation measure, not selected only because it is easy to log. Measurement quality matters in implementation research [9,24]. Before use, investigators should specify why the action is meaningful for the relevant public health program, whether the needed records can be captured reliably, and which claim the end point could support (Table 2). The table is a proposed appraisal framework, not a consensus-derived validation instrument; empirical testing is needed to establish its feasibility, reliability, and usefulness across settings.

Three inferences should be kept separate. First, time-stamped logs can show temporal proximity, indicating that an AI output preceded an action. Second, a workflow record or actor documentation can show a documented role, meaning the output initiated, accelerated, modified, supported, or had no documented role in that action. Third, a causal design can estimate attributable action change, that is, whether an action changed because the AI implementation was introduced. The first two are necessary for an auditable pathway but are not substitutes for the third. For causal attribution, studies need an appropriate comparator or counterfactual strategy, such as randomization, a controlled before-and-after design, interrupted time series, or another design that addresses secular trends, case mix, concurrent interventions, and selective follow-up.

The selected observation window should reflect clinical urgency, public health relevance, workflow capacity, and the consequences of delay. A descriptive postdeployment action rate can support an implementation description. A claim that AI improved action requires an attribution strategy; a claim of clinical or population health benefit requires outcome-linked causal evidence. For patient-facing communication, message delivery, clicks, and conversation volume are not evidence of understanding. Comprehension should use language-appropriate, literacy-sensitive measures, teach-back, or other prespecified measures [25]. Usability may be assessed separately with an instrument such as the System Usability Scale [26]. Workload can use task time, queue or message volume, or a prespecified instrument such as the NASA Task Load Index [27].

The practical test is whether a manuscript can translate a weak end point into a measurable action end point. The examples in Table 3 are author-proposed illustrations, not a systematic review–derived taxonomy or a claim that every published implementation used the listed weak end point. Their purpose is to show how the denominator, action opportunity, accountable actor, window, attribution record, safeguards, and claim boundary can be made visible.

Public health action end points apply to both research implementations and production systems, but they do not collapse the governance requirements of those settings. For research, investigators should state whether human research ethics review, trial registration, data governance approval, and funder due diligence apply to the planned evaluation. For production systems, the end point complements rather than substitutes for jurisdiction-specific regulatory requirements, institutional oversight, quality management, and postdeployment monitoring. It does not determine whether a system is regulated as a medical device or the level of evidence required for authorization.

Research ethics committees or institutional review boards, public health and clinical leaders, data governance teams, and funders where relevant should be involved before implementation. The proposed workflow should include contingency actions for unsafe outputs, capacity failures, threshold or model changes, human overrides, and fallback or rollback. This is especially important where automated recommendations or generated content can affect triage, escalation, access to services, or allocation of limited resources.

Action end points should not be imposed on every early-stage AI study. Model development, external validation, silent evaluation, and usability testing can be valuable without proving action change. Nor does the framework require every system to demonstrate population health benefit immediately. It asks that the claim stop at the stage tested (ie, technical performance, feasibility, descriptive action completion, attributable action change, or outcome-linked population health benefit). A completed-action rate alone supports only a descriptive implementation statement, a causal action-change claim requires a comparator or counterfactual strategy, and outcome claims require a design that links action changes to relevant outcomes.

The end point should not reward action volume for its own sake. Low thresholds can increase unnecessary field work, testing, anxiety, stigma, overtreatment, or resource diversion, while high thresholds can miss people or communities most likely to benefit. Unequal data quality can also make a model appear accurate while excluding digitally marginalized populations [28]. Prespecification of denominators, thresholds, action categories, nonaction reasons, duplicate rules, observation windows, equity strata, and audit logs helps prevent gaming and keeps missed action, unresolved status, and denominator loss visible.

For authors and reviewers, the key reporting question is not whether an action end point is universally required. It is whether an AI-supported action-change claim is proportionate to the action pathway that has been measured. Manuscripts making such a claim should state the output, denominator, action opportunity, accountable actor, observation window, action status categories, attribution strategy, safeguards, and intended level of inference. Without these elements, an article may be reporting a model or system rather than an evaluated AI-supported public health intervention.

Digital public health AI should not be judged only by whether models classify risk, generate messages, or automate surveillance. A public health action end point makes a proximal action pathway observable by detailing what opportunity followed which AI output, for whom, by whom, within what window, with what action status and documented AI role. It is a structured process or implementation end point, not a causal outcome or a surrogate for population health benefit. The allowable claim should stop at the last link that has been measured with an appropriate attribution strategy.