AI for Good: Maximizing Social Impact
How to use AI to expand social impact: AI as a block of the flow, with human review, organized data, LGPD and an implementation roadmap in six steps.
Published on By João Pedro Megid Carrilho
AI expands social impact when it is used as a block with a defined function inside a flow, such as transcribing, classifying or summarizing, with the team reviewing what it writes. It works best with organized data, consent at collection and decisions kept in the hands of people.
What AI does well in a social operation
In CORRE projects, AI appears in tasks like these:
- Transcribe audio from interviews and WhatsApp messages.
- Classify answers and messages into categories defined by the team.
- Read documents, such as the notice of a call for proposals, and fill in a record.
- Draft reports, diagnoses or e-mails.
One example is the 2026 Mapping by Lab Nova Longevidade, with Ashoka: interviews over WhatsApp with transcribed audio, diagnoses and recommendations with more than 50 rules and AI, and more than 1,000 hours automated (Lab Nova case).
In the TIC OSFIL 2025 survey, by Cetic.br, 29% of the nonprofit organizations interviewed said they had used AI technologies in the previous 12 months. It is the first time the indicator appears in the survey.
What AI should not do alone
AI does not approve a benefit, does not decide who receives support and does not respond to a risk situation. These cases call for a person, and the flow should recognize them and route them to the team. An AI that "acts on its own" in a social service creates a problem that the time saved does not pay for.
Data before AI
AI on top of data scattered across spreadsheets repeats the mess faster. Before it, three things come first:
- A single database of beneficiaries and services.
- Standardized questions, so the answers are comparable.
- Written rules for what each piece of data means and how it is used.
With that, reports and dashboards become a consequence of the routine.
Ethics, governance and LGPD
- Consent at collection. The LGPD asks, in art. 6, for a legitimate, specific purpose informed to the data subject and the minimum necessary data.
- Record. Art. 37 requires the controller and the operator to keep a record of processing operations.
- Human review. The text that AI writes is a draft, and a person reviews it before it goes out.
- Transparency. Whoever talks to an automated flow should know it and have a way to talk to a person.
At CORRE, LGPD and security run through the whole flow: we ask for consent at collection, each action is recorded and we sign a data processing agreement before starting.
Implementation roadmap
- Diagnosis. Where the data is, who does each step and what rules exist.
- Use cases. Choose one or two processes with volume, a clear rule and pain.
- Tool. Compare generic tools and a platform with a team, by the process and not by fashion.
- Training. Train the team and involve those who will operate it in the design.
- Pilot. Start small, measure and adjust.
- Measurement. Hours given back, response time and quality of the reports.
If your network or program needs to listen to many people, see the Networks and programs that listen page.
Questions on this topic
What is the role of AI in an NGO?
To do well-defined parts of the work: transcribe audio, classify answers and draft report texts. It is a block inside a flow, and the team reviews what the AI writes. Decisions remain with people.
Will AI replace social workers and technicians?
No. It takes over repetitive tasks, such as copying data, putting reports together and sending reminders, and frees the professional for listening, follow-up and decisions that require judgment.
How do I reduce the risk of AI bias and errors?
Define the function of the AI precisely, use categories and rules that people check, keep human review at decision points and record what was done. When the AI and the review diverge a lot, adjust the flow design.