Artificial intelligence is attracting growing attention across health systems. It can help teams identify patterns, prioritize work, summarize information and support faster decisions. Yet many promising projects never become useful tools in routine public health practice.
The problem is not always the model. A system may perform well during development and still fail when introduced into the setting where people are expected to use it. The data may arrive late. Staff may not understand the output. The recommendation may not fit an existing responsibility. The tool may identify a problem without showing what should happen next.
These are not minor implementation details. They determine whether artificial intelligence becomes part of better public health practice or remains an impressive demonstration.
Technical success is not the same as public health value
Data science teams often evaluate an AI system through measures such as accuracy, sensitivity, specificity or prediction error. These measures matter, but they do not answer the full implementation question.
A public health programme needs to know whether the system improves a real decision. Does it help a district team identify a service gap earlier? Does it help a programme manager direct support to the right facilities? Does it reduce the time required to review reports? Does it help staff communicate risk more clearly?
If the output does not change an appropriate action, strong technical performance may produce little operational value.
An AI project becomes useful when its output reaches the right person, at the right time, in a form that supports a responsible decision.
Projects begin with technology instead of a decision
One common mistake is starting with a model and searching for somewhere to use it. A team may decide to apply machine learning because the technology is available, because a dataset exists or because innovation funding has become accessible.
A stronger starting point is a specific public health decision. The project team should ask who makes the decision, what information is currently used, where delays occur and what an improved process would look like.
Consider a system designed to predict facilities at risk of missing an immunization target. A prediction alone is not enough. The district team must understand the factors influencing the result, confirm whether the underlying data are complete and decide whether the response should involve supervision, community engagement, logistics or another form of support.
When the decision is clear, the team can judge whether AI is necessary. In some situations, a simple threshold, trend chart or reporting alert may solve the problem more reliably.
Poor data quality becomes hidden inside the model
Public health data are produced through real service delivery. Registers may be incomplete. Definitions may vary. Reports may arrive late. Staff may face heavy workloads, weak connectivity or limited access to equipment. A model trained on these records can reproduce their weaknesses while presenting its output with a level of confidence that users may misunderstand.
Data quality assessment should therefore happen before modelling and continue after deployment. Teams need to examine completeness, timeliness, consistency, representativeness and changes in reporting practice.
The system should also communicate uncertainty. Users should be able to distinguish between a genuine warning and a result that may be influenced by missing or unusual data. This is especially important when resources are limited and one recommendation may redirect attention away from another priority.
The tool does not fit the working environment
A solution designed in a well connected development environment may perform differently in a district office or health facility. Internet access may be inconsistent. Staff may share devices. Some decisions may happen during meetings where information must be understood quickly. Others may depend on printed summaries, telephone follow up or field visits.
Workflow first design studies these realities before choosing the final form of the technology. It asks how information moves, where responsibility sits and what constraints users experience.
This may lead to practical design choices such as a simple mobile view, a downloadable report, clear alert categories, limited data entry or an option that remains useful when connectivity is weak. The best interface is not necessarily the most advanced one. It is the one that supports the work without adding unnecessary burden.
Public health professionals are consulted too late
People who understand programme delivery should not be invited only when the product is ready for testing. Their knowledge is essential during problem definition, data interpretation, design and evaluation.
A nutrition officer may understand why an indicator changed after a community activity. A health information officer may recognize a reporting pattern that appears unusual but has an operational explanation. A disease control officer may know which alerts require immediate escalation and which require confirmation first.
This knowledge helps the project team avoid false assumptions. It also increases trust because users can see that the system reflects their responsibilities and experience.
Responsibility becomes unclear
An AI recommendation should never create confusion about who is responsible for the final decision. Public health work involves professional judgement, programme guidance, ethical obligations and accountability to communities.
Every project should define who reviews the output, who can act on it, what evidence must be checked and how decisions are recorded. There should also be a clear process for reporting errors, questioning recommendations and correcting the system.
The World Health Organization emphasizes that ethics, human rights, transparency and accountability should remain central to the design and use of artificial intelligence for health. Human oversight is therefore not an obstacle to innovation. It is part of responsible implementation.
Success is measured at launch instead of over time
A successful demonstration does not prove that a tool will remain useful. Data patterns change. Programmes change. Staff members move into new roles. New reporting requirements emerge. A model that worked during development may become less reliable as its environment changes.
Evaluation should continue after deployment. Teams should monitor technical performance, user experience, unintended consequences and whether the tool is improving the intended decision.
Useful questions include:
- Are the intended users consulting the output?
- Do they understand what it means?
- Does it influence an appropriate action?
- Are some facilities or communities affected unfairly?
- Has the tool reduced work or created another reporting burden?
- Can the organization maintain it with available skills and resources?
A practical workflow first framework
Public health organizations can improve the chances of success by following a simple sequence.
- Define the decision. Describe the exact decision that needs to improve and why it matters.
- Map the workflow. Identify the people, information, timing, approvals and actions involved.
- Assess the data. Examine quality, access, privacy, representativeness and limitations.
- Choose the simplest useful method. Compare AI with rules, summaries, visualizations and process changes.
- Design with users. Involve the professionals and communities affected by the system.
- Define oversight. Clarify responsibility, review, escalation and correction procedures.
- Test in the real setting. Observe how the tool performs within actual constraints.
- Measure public health value. Evaluate decisions, actions, equity, workload and outcomes.
What will the responsible person do differently when this information appears?
AI should strengthen professional judgement
The future of AI in public health should not be framed as a competition between people and machines. The more useful question is how technology can support professionals who already understand communities, programmes and health system realities.
AI can help review large volumes of information, detect unusual patterns and reduce repetitive work. Public health professionals bring context, ethical judgement, communication and responsibility. Strong implementation combines these capabilities.
This balance is particularly important in settings where resources are limited. Technology should help teams focus their time and attention. It should not create another system that requires constant support while offering little value in return.
Conclusion
Many AI projects in public health fail because implementation is treated as the final stage rather than the foundation. The model is developed first, while workflows, users, accountability and sustainability receive attention later.
Workflow first design changes that order. It begins with a real decision, studies the environment in which that decision occurs and selects technology only after the need is understood.
Artificial intelligence has meaningful potential in public health. Real progress, however, will not be measured by the number of models created. It will be measured by whether those models help people make timely, fair and responsible decisions that improve health.