Inside the DRC Ebola Response, Claude's Job Is Mostly Cleaning Up PowerPoint

Anthropic has published an account of how Claude is being used in the response to a Bundibugyo ebolavirus outbreak in the eastern Democratic Republic of the Congo, in a partnership convened by the Coalition for Epidemic Preparedness Innovations with the WHO Regional Office for Africa and the Institut National de Recherche Biomédicale. The headline number is an operational one: a daily situation report that used to consume most of a day can now be assembled in under an hour, according to the WHO AFRO team. What is striking about the deployment is how little of it involves anything resembling autonomous decision-making.

The bottleneck was never the modeling

The pipeline Anthropic describes will be familiar to anyone who has worked on data plumbing in a resource-constrained environment. Community health workers — relais communautaires — go door to door recording fevers, recoveries and deaths in notebooks. Those notes get texted to a zone manager over WhatsApp. Health facilities pass case information to districts, which assemble it into PowerPoint slides and forward it to the outbreak coordination center. Staff at the provincial health ministry work late stitching the decks together into the published situation report.

Anthropic says the Claude skill built by WHO AFRO's Tendai Muza, working with colleagues Tamayi Mlanda and Gianni Donkor Muza, extracts case and lab numbers from each health zone's deck, compares them against the prior day's report, flags trend changes with a proposed explanation, and summarizes the district reports. That is document parsing, reconciliation and summarization — a set of tasks that has been within reach of language models for some time, applied to a workflow where the cost of manual effort is measured in hours that responders do not have.

The stakes justify the attention. Anthropic says nearly 8,000 confirmed cases have been recorded in the outbreak, which began in May, and that almost half of those people have died. There is no confirmed vaccine for the Bundibugyo strain. How quickly a sick person is located, the company notes, often determines whether they survive — and the same contact data would direct vaccine deployment if one becomes available.

Dr. Jean-Jacques Muyembe, director general of INRB and a co-discoverer of ebolavirus in 1976, is quoted saying that the tools have changed completely over fifty years but the rule has not: you beat Ebola by knowing where it is today, not where it was last week.

Modeling capacity, not model quality

The second claim is about analytic throughput. Paul Ouma of the WHO AFRO team says that time constraints previously forced the team to commit to a single disease model without being able to evaluate which one was most appropriate. With Claude, he says, they can run multiple models in parallel and build forecasts that inform decisions such as where to site treatment centers.

This is a capacity argument rather than an accuracy argument, and it is worth keeping the two separate. Running more models faster does not by itself produce better forecasts; it produces more candidate forecasts, which still require epidemiological judgment to adjudicate. Anthropic does not disclose how model selection is validated, what the forecasting error looks like, or whether any of the forecasts have been retrospectively checked against outcomes. WHO AFRO's emergencies hub in Dakar handles roughly a hundred public health events a year, so the marginal value of freed analyst time is plausibly high — but the announcement offers no measurement of downstream decision quality.

CEPI draws an explicit line

The most useful quote in the release comes from Polina Brangel, R&D data innovation lead at CEPI. She says Claude does not decide which cohorts of serological samples should be analyzed to test Ervebo's potential cross-reactivity against Bundibugyo — that remains an expert scientific judgment. What the model does, she says, is organize complex multi-factor data so humans can compare across a wide range of proposals faster.

CEPI used Claude to build a dashboard tracking the work needed to advance vaccine candidates after it invited proposals in response to the outbreak. Ervebo, the existing approved vaccine, targets the Zaire strain, not Bundibugyo.

Brangel's framing — that this has been a chance to learn in real time where AI makes a practical difference for teams under pressure — is notably more hedged than the way AI deployments in health are usually described. It is the right posture for a live outbreak, and it implicitly concedes that the boundaries are still being drawn.

Genomics is the harder claim

Anthropic says INRB has produced essentially all of the Bundibugyo genomes sequenced from DRC cases in this outbreak, and that Claude Science — described as an AI workbench for scientists — lets researchers assemble genomes and build phylogenetic trees from plain-language prompts rather than terminal commands. The stated benefit is lowering the barrier to bioinformatics analysis when specialist capacity is scarce, which it usually is during a major outbreak.

This is the part of the deployment that deserves the most scrutiny. Genome assembly and phylogenetic inference involve parameter choices that materially change conclusions about transmission chains and variants. Making those steps easier to invoke is not the same as making them easier to get right, and the announcement does not describe validation procedures, reviewer oversight, or how errors would be caught. If genomic results are used to infer undetected transmission chains or flag mutations that could affect diagnostics, the cost of a silent failure is high.

What the announcement does not say

Several things are absent. There is no disclosure of how patient-identifying information is handled, whether data leaves the region, or what contractual and privacy terms govern the deployment. There is no error rate for the situation report extraction, no accuracy baseline for the manual process it replaced, and no independent evaluation. Cost, funding arrangements and the commercial terms between Anthropic and the health organizations were not disclosed. Nor is it clear how the workflow behaves when connectivity fails — a real constraint in the geography described.

WHO AFRO is already extending the approach, using Claude to clean and validate line lists for a chikungunya outbreak and planning to stitch together historical disease databases so past outbreaks can inform the next one. Anthropic says there are some promising signs of the Bundibugyo outbreak slowing in Ituri, while cautioning that containment will take time.

What to watch

• Whether any independent evaluation of the situation report pipeline is published, including extraction error rates and how discrepancies against source decks are detected.
• How the genomics workflow is validated — specifically, who reviews assemblies and phylogenies produced through natural-language prompting before they inform case linkage decisions.
• Whether data governance terms for patient-adjacent information are made public, including residency, retention and access.
• Whether the chikungunya line-list work and the historical database project produce durable tooling or remain bespoke per-outbreak builds dependent on one vendor's models.
• Whether WHO AFRO or CEPI report outcome measures — case detection speed, forecast accuracy, resource allocation quality — rather than time saved, which is the only metric offered so far.