Agência Mural is a nonprofit newsroom in Brazil with its reporting focused on underrepresented communities in the metropolitan areas of São Paulo. With Local Climate Alert, the team is applying AI to service journalism, sending concise, location-aware alerts about climate risks directly to people’s phones and inboxes. As Brazil prepares to host a major climate summit, the project positions local audiences to make safer, faster decisions.

The problem: Being usefully fast

The impetus is rooted in lived experience. “All of us are from the peripheries,” says Vagner de Alencar, director of journalism at Agência Mural. “We know what is happening in these places.” Mural’s network of more than 60 local correspondents consistently reports that climate impacts are felt first and hardest in the margins. The team wanted to move beyond one-off reports to provide actionable, real-time information in the channels people actually use. “Our audience wants information directly through WhatsApp,” notes de Alencar. “They want to know about air quality or floods right now.”

The project also builds on prior WhatsApp experiments, including a pandemic-era micro-podcast sent as short audio messages and Papo Reto no Zap, a community initiative across five neighbourhoods. Those experiences validated WhatsApp as a primary interface for news and services.

Building the solution: Roadmap to prototyping

The team followed a pragmatic path from concept to pilot:

• User research and scoping: They ran surveys and interviews to learn what people wanted, how often, and in which formats. The priority was clear, short alerts about immediate risks, delivered where people already converse.

• Data sourcing and cleaning: Rather than rely on third-party APIs, the team identified open public databases relevant to weather, floods and air quality. Early tests with generative text “had too much hallucination,” so they designed a pipeline focused on extraction, cleaning and concise templated outputs.

• Technical partnership: Recognising a gap in-house, Mural partnered with Agência Tatu, a newsroom in Brazil’s northeast with data and automation experience. “We put together our strengths. Without them it would be almost impossible,” says de Alencar.

• Minimum viable alerting: They launched a WhatsApp Community and an email list for five diverse communities across the capital region. People register once and receive periodic alerts, plus links to a project web page that hosts related service content.

• Live community launch: True to Mural’s ethos, the pilot launched with an in-person event featuring exhibitions, conversations with residents and participation from local correspondents. “AI is online, but our essence is being connected in person,” de Alencar reflects. “We exist for these communities. AI is a means to serve them better.”

The tools and technologies involved in Local Climate Alert

The team intentionally deprioritised free-form LLM writing in favour of structured templates to avoid factual drift. “We now have a process that extracts and cleans the data,” says de Alencar. To build and make Local Climate Alert work, they relied on different tools:

• Open public datasets for air quality, rainfall and flood risk

• A data pipeline to extract, normalise and validate fields before publishing

• Lightweight templating to generate clear, non-hallucinatory alert copy

• WhatsApp Communities and email for distribution

• A web hub to host evergreen service information and updates

• Visual and UX support to clearly label what is AI-assisted versus human-authored

The team was assembled to place community needs at the centre of the development process, with technical expertise provided primarily by Agência Tatu, Mural’s partner in the initiative. Within Mural, an editorial lead ensures the service is framed in ways that remain relevant to local audiences, while a tech and business lead coordinates the build and partner integration. Audience engagement is strengthened by a community coordinator who works with more than 60 correspondents to root the alerts in local realities. Research is carried out by a dedicated role focused on surveys and interviews with residents and specialists. A visual and web designer contributes to clarity, accessibility, and proper labelling across channels. Additionally, data and automation support is provided via Agência Tatu, bringing Python and interface development expertise to the project.  

Challenges they encountered

• Limits of WhatsApp Communities: Communities are discoverable but less direct than one-to-one messaging. “Our dream is to send as a pop-up directly to people’s numbers,” says de Alencar, but business messaging at scale costs money and is priced in dollars. That makes full automation difficult for a nonprofit.

• Avoiding hallucinations: Early experiments with LLM-written alerts produced inaccuracies. The team switched to a constrained, template-driven approach backed by rigorous data cleaning and editorial checks.

• Data fragmentation and scarcity: Useful public data exists but is scattered and uneven. Normalising formats, ensuring freshness and matching signals to specific neighbourhoods require ongoing maintenance.

• Resourcing specialised roles: It took time to hire for community management and audience development, roles that proved essential to adoption and trust.

The opportunities: To multiply the impact and drive revenue

Replicability is a central insight. “We created something completely replicable,” says de Alencar. The alerting “robot structure” can be pointed to other useful datasets, such as job opportunities or public services. The model also opens pathways to sustainability by deepening local reach and potentially attracting place-based advertising on Mural’s site. Above all, the project has unlocked organisational capacity to produce service journalism at scale, not only as special projects.

Lessons for newsrooms

• Meet people where they are: Design for the channels your audience already uses. For Mural’s readers, WhatsApp is the front door to timely, trusted service information.

• Constrain AI to protect accuracy: Templated outputs on top of clean, validated data reduce errors and build trust.

• Pair tech with presence: Digital alerts travel fast, but legitimacy is earned face-to-face. In-person launches and ongoing community contact anchor the service in real needs and local knowledge.