Superintelligence for Nonprofits in Brazil: Donors and Reviewed Reports

Superintelligence for Nonprofits in Brazil: Donors and Reviewed Reports

Learn how a Brazilian nonprofit can use AI for donor questions and draft financial reports, with sourced figures, minimal personal data and human review.

•

Want to apply this in your company?

Book a call with Juan →

A nonprofit can use AI to answer donor questions and draft accountability reports, provided that every number comes from an approved source, donor personal data stays at the minimum necessary, and a member of the team reviews what is sent. The word superintelligence comes up often in this topic, but what works in practice is an agent with a closed scope, a record of conversations, and handoff to humans in sensitive cases. This article describes that design for leaders who need to decide and act.

Key takeaways

  • Answering donors is work of lookup and routing. It does not decide how funds are allocated or promise what the organization has not yet approved.
  • Every published figure needs an official source, a cutoff date and a named person responsible for checking it.
  • Donor data enters the agent's context only when the task requires it, and only after the donor is verified.
  • A draft accountability report speeds up the team, but it does not replace financial review or board approval.
  • A pilot with a small set of real questions and a success criterion agreed in advance is worth more than a broad launch.

What superintelligence means for a nonprofit

The term is used very loosely in software marketing and in general conversation. For an organization that receives donations and reports on them, the useful question is different: which task does the system perform, with which sources, and who answers for the result. If the board is still separating the concept from the product it intends to buy, the article Is ChatGPT Superintelligence? What It Is and What It Is Not helps align the vocabulary before any decision.

In practice, AI enters as a tool for lookup and writing. It finds information in approved documents, organizes the answer, and records what happened. It does not decide how funds are spent, does not approve expenses, and does not replace the accountant or the board.

Which donor questions the agent can answer

Start with what is repetitive and public. In general, questions fall into three groups.

  • Public information: mission, ongoing projects, where to find the annual report, and how to make a new donation.
  • Operational: how to cancel a recurring donation, how to update the payment method, and how to request a receipt, following the internal policy as published.
  • Sensitive: complaints, questions about how a specific campaign's funds were used, requests to delete or access data, and any legal or tax matter concerning an individual donor.

The agent answers the first group based on approved texts, walks through the second group using the written procedure, and routes the third group to the team with a summary of the conversation. In that last case, it should not invent a substantive answer. A similar triage, with clear limits between what the system answers and what the team decides, is described in AI Customer Support: What to Automate and What to Keep Human.

How to build the accountability summary step by step

The summary for donors and partners usually comes from spreadsheets, bank statements and project reports that are scattered across folders. A safe workflow separates the calculation, which is deterministic, from the narrative, which can be assisted by AI.

  • Define the period and the cutoff date. Transactions after the cutoff stay out of this summary and move to the next period.
  • Gather the official sources: the bank statement, the income and expense spreadsheet already reviewed, and project reports approved by the board.
  • Calculate the totals in the spreadsheet or the financial system. The model can describe the result, but it should not add up amounts on its own.
  • Ask the agent for a draft in which each figure carries a reference to its source, such as a tab, document or report.
  • Send the draft to the financial officer for line-by-line checking and then to the board for approval.
  • Publish only the approved version, with the date and responsible person recorded.

This chain has a practical benefit. Months later, when a donor or a board member asks about a figure, the team can go back to the published version and the source behind it, without rebuilding the reasoning from memory.

How to protect donor data

Donor records often include name, contact details, national ID number, address, donation amount and frequency. Almost none of these fields need to be in the agent's context to explain how to cancel a donation. The guiding rule is to limit access to what the task requires.

  • Restrict lookups of personal data: the agent checks the status of a donation only after the donor identifies themselves through a verified channel.
  • Do not paste full donor databases into personal chat tools or into accounts the organization does not control.
  • Log what was asked, what was answered, and when a person took over the conversation.
  • Check where the vendor processes data and whether it is transferred outside the country, which requires attention under the LGPD. The ANPD publishes guidance on this topic that is worth following.
  • Treat donor messages as untrusted input. Someone may try to instruct the agent to reveal other people's data or to ignore its own rules. OWASP documents this type of risk among the main concerns for applications built on language models.

Where human review is mandatory

A sound design defines in advance the points where a person must decide. As a starting guide, consider the following.

  • Any figure that is published or sent to a donor passes through review by an identified responsible person.
  • Requests to delete, correct or access data go to the organization's data protection officer.
  • Complaints about the use of funds go to the coordination team, with a summary of the conversation and no automatic substantive reply.
  • Tax questions or individual receipts go through the finance team or the accountant.
  • Any decision that materially affects a donor, such as blocking a record, requires human review.

The LGPD provides the right to request a review of decisions made solely on the basis of automated processing. It is worth considering this right from the design stage of the workflow, as explained in LGPD Automated Decision-Making: The Right to Review (Art. 20).

Illustrative scenario: a question about a campaign

Illustrative scenario, with a fictional organization: a food assistance nonprofit receives, through WhatsApp, a donor's question about where the money from a winter campaign was spent. The agent finds the published report for that campaign, answers with what the document says, and sends the link. The donor then asks why a specific amount does not appear in the report. The agent does not risk an explanation. It states that the question will be forwarded to the finance team, opens a case, and records the summary. The team checks the transactions and replies through the same channel. The record shows clearly what was automated and what was decided by a person.

How to pilot without promising what does not exist yet

Juan Carlo's approach follows four steps: map the current process and its baseline, limit data and actions, pilot a real workflow with logging and a success metric agreed beforehand, and then hand over the operation and decide whether the workflow expands, is adjusted, or stops. In a nonprofit, the pilot can begin with the accountability question that reaches the team most often in the recent period. Before that, it helps to organize the sources, as suggested in Data Ready for Superintelligence: What to Standardize First.

Before putting the agent in contact with donors, test its answers against real questions and expected answers. OpenAI describes this kind of evaluation in its evals guide, and the same principle applies to a small team: if the expected answer is not written down, there is no way to know whether the agent got it right.

Frequently asked questions

Can AI write the accountability report on its own? It is not advisable. It can draft text from verified sources, but the totals should come from the spreadsheet or financial system, and publication depends on approval by a responsible person and the board. AI speeds up the writing, but accountability stays with the organization.

Do donors need to know they are talking to an automated system? The safest approach is to state at the start of the conversation that it is automated and offer a clear way to reach a person. This transparency makes it easier to exercise the right to review automated decisions provided by the LGPD. The notice text should be validated by the organization and, if in doubt, by its legal advisor.

Should a nonprofit run a model inside its own infrastructure for donor data? It can be an option to reduce data leaving the organization, but it does not remove the need for access control, operation logs and review. A local installation that is poorly governed is still a risk. The decision should start from mapping which data enters the workflow.

Where should a small nonprofit begin? Choose one frequent question, one official source and one person responsible for review. With that trio defined, you can measure the result within a few weeks and decide based on records, not impressions.

Sources

AI for nonprofitsdonor communicationfinancial reportingLGPD compliancehuman review