Left to its own incentives, AI weather forecasting adds skill fastest where data are densest, for the uses existing forecasts already serve well. The populations that the first forecasting revolution never truly reached may be left behind again. Nothing about the technology itself makes that inevitable. What determines it is a set of human choices about what data these models learn from, how they are evaluated, and whose decisions they are built to serve. Policymakers, researchers, and technologists should be deliberate about the choices they make now, because they get harder to reopen as benchmarks, standards, and procurement patterns become fixed around them. (Chapter 1)

Running a leading forecast used to require a machine costing on the order of a hundred million dollars. It now takes minutes on a single processor. Rebuilding a model around one user’s question took months, and now takes days. Adapting a forecast to local conditions is within reach of a modestly resourced weather service. None of that was true five years ago. “Supply-driven” climate services—the classic approach of producing a forecast and pushing it downstream to end users, rather than building it around those users’ decisions—are now a choice, not a necessity. The time is now for weather services, health services, and funders to take advantage of what is newly possible. (Chapters 1, 3)

81% of national meteorological services report providing climate services for health, but just 23% of health ministries run a surveillance system that uses meteorological information. That gap is usually read as a dissemination problem, and some of it is. But a forecast built around what could be produced will often not fit the decision someone actually has to make, and low uptake is simply the first place that becomes visible. These ”first-mile” problems can be addressed by decision owners specifying what a forecast must do before it is built. (Chapter 2)

Some decisions need two to four weeks, enough for a health system to stage its response to a heat episode instead of reacting to it. Few forecasts exist at that lead time, so nobody asks for them. Other decisions need no forecast at all. Where a long biological lag exists between the weather and the illness, the lag supplies the warning time, and watching current conditions does more work than a forecast of what is coming. Investment has followed the forecast rather than the lead time the decision requires. Funders and service providers should start from the decision and its lead time, then invest in forecasts or monitoring accordingly. (Chapter 2)

Forecasts have been too coarse to tell one district from another, too late to act on, or built around a weather variable that does not track what makes people ill. Improving climate services needs more, because other parts of the chain between a forecast and a healthier person are still missing. Poor forecasts gave funders a reason to neglect them. Health decisions were never specified, but health ministries can specify them. Forecasts were judged on weather accuracy and not health outcomes, but meteorological services can change what they evaluate against. Nobody collected the data needed to check if a warning helped, but funders can build that evidence. Institutions that need to work together were not required to speak to each other. While the forecast was the binding constraint, fixing any one of these alone changed little. Those difficulties remain. The return on addressing them has risen. (Chapter 3)

Comparing models against each other, against a reference forecast, and against what actually happened is called benchmarking, and developers optimize whatever the benchmark rewards. Today’s AI weather benchmarks reward performance in data-rich regions by evaluating against the global atmosphere. They tell a health service little about whether a model supports decisions. A dengue alert should test models on dengue cases, and where no case record exists to run that test, it should test whether the forecast gets the weather threshold right. For a heat plan, that means the days it should activate, in the districts it covers. The same weather model can score well on one test and badly on the other. A better benchmark serves whoever specified the decision, and choosing whose decisions to specify is a separate and equally consequential step. (Chapter 4)

Demand articulation; weather data stewardship; health data stewardship; forecast creation; benchmarking; operational production; translation; delivery; and monitoring, evaluation, and feedback. Demand articulation comes first and is where the health sector drives the process: the decisions everything else is built around are stated by people who have to make them. Which institution should hold each capacity varies too much by country for a general recommendation. Cheaper AI forecasts will increase the number of warnings a health system has available. Who acts on a warning, and who decides not to, is usually unassigned. Every country can name a holder for each of the nine, and name who owns the action a warning triggers. (Chapter 6)

Someone who will be blamed for a false alarm and not credited for a quiet season is unlikely to disseminate a forecast, and added accuracy does not change that calculation. Three things change this. The ability to analyze scenarios that let a decision-maker evaluate response options against the full probability distribution, not just a single number. Plain statements from the meteorological services of what a forecast can and cannot tell you, so that health users become critical readers of model output rather than recipients of it. And accountability for forecast errors settled in advance. That means a duty of care, not liability for outcomes. Met services and health ministries can settle accountability between them before the first warning goes out, so waiting to disseminate stops being the safest option. (Chapter 4, 6)

Where an investment builds a forecast, a dataset, a benchmark, a research finding, or an institution, it has to state where decision-making, budgets, authorship, and long-term control will sit, and place them closer to the people carrying the greatest unmet burden. A model trained on African data, published by a European laboratory, and licensed back to African services satisfies every technical description of local relevance, and yet does not lead to any systematic change in forecast access or in local capacity. Without that specification, AI weather forecasting improves global forecasting capability and leaves the infrastructure concentrated exactly where it already is. Funders can require this in every proposal they approve. (Chapter 6, 7)

Absent upfront decision-making, these policies risk name-checking areas like health without actually determining which agencies are responsible for taking particular actions. National climate adaptation plans provide a cautionary tale: Every one of 59 such plans reviewed in 2025 identifies health as a sector that is vulnerable to climate change, but fewer than half even identify a lead agency responsible for managing the issue. It’s no wonder that just 0.2% of international climate adaptation finance has health as its primary focus. To prevent this, policymakers working on AI agendas should avoid naming health without any attached action or responsibility. The investment side of the same problem is set out below. (Chapter 6, 7)