MUSYG · AI ADOPTION

Organisation tracks

Nonprofit or foundation track — 60 days

Adapt responsibilities and your first pilot to your situation.

Build my route

Particular considerations

Value is not limited to financial gain. Beneficiaries, donors, volunteers, partners, and vulnerable populations may experience different effects. Mission, trust, and fairness belong in the baseline.

The Local Link Foundation synthetic case shows how to bound an A2 business agent for grant-dossier administration: offline channels remain open, evidence is segmented by language and channel, and 100% of funding decisions remain human.

What published evidence supports (checked 18 November 2026)

Use is widespread but mostly unmanaged, and no measured time saving exists for a typical nonprofit task.

All of these are self-reported surveys (grade C). The only controlled study with timed tasks in philanthropy, on screening applications for MIT Solve, capped each decision at 2.5 minutes, so it cannot show a time saving. Measure your own baseline: see the measurement sheet below.

Roles to name on day 1

  • Board: approves the AI policy and reviews incidents at least every quarter. Trustees stay accountable; decisions are not delegated to AI alone.
  • AI lead or small task force: keeps the inventory of uses, including unofficial ones, the list of approved tools, and the measurement sheets.
  • Data protection lead: classifies data, completes impact assessments where required (FADP or GDPR), checks contracts.
  • Process owner and named reviewer for each use case.
  • Staff representatives, where the law requires consultation (for example a works council in France).
  • For foundations, the grant committee: keeps every funding decision.

Days 1–15 — Mission-aligned mandate

  • name a process owner and a mission steward;
  • consult people who work closely with beneficiaries;
  • inventory sensitive data, secrets, consent, and funder restrictions;
  • define intended outcomes and harms to avoid;
  • exclude any unauthorized automated decision about access to support.

Days 16–30 — Prioritize and protect

Begin with an assistive case: document research, internal summarization, controlled translation, campaign preparation, or administrative triage with validation.

  • classify populations and segments that could be disadvantaged;
  • assess accessibility, language, cultural context, and offline channels;
  • evaluate the supplier, data reuse, and dependency;
  • prepare a complaint or human-escalation mechanism.

Days 31–45 — Evaluate

  • use authorized, representative real cases;
  • measure quality by language and relevant group;
  • test fabricated facts, stigmatization, and disclosure;
  • compare time saved with the verification burden;
  • document who judges and with what expertise.

Days 46–60 — Pilot and accountability

  • use shadow mode, then copilot mode if thresholds are met;
  • inform users where necessary;
  • monitor corrections, complaints, access, and incidents;
  • make the decision understandable to the board or oversight body;
  • where appropriate, publish a proportionate description of the use and safeguards.

Mission-specific gate

Even if productivity improves, stop the pilot if mission, dignity, equitable access, confidentiality, or trust deteriorates.

Additional gates by use

Use Pass only when Why
Any use Data classes defined; approved tools listed, with no training on your data; staff given a compliant tool rather than a blanket ban Sensitive data in public tools is the most reported incident type
Content for donors A person reviews every external text; disclosure is proportionate to the risk of misleading donors; no synthetic images of beneficiaries without a dignity review Fundraising Regulator guidance (December 2025); EU AI Act Article 50 applies since 2 August 2026; experiments find disclosed AI can lower giving, depending on wording
Services to beneficiaries No crisis, health or eligibility channel run by AI alone; every AI-drafted case note is reviewed; human escalation tested A helpline chatbot was withdrawn in 2023 after harmful advice; AI transcripts in UK social work introduced false statements
Screening grant applications AI summarises or ranks but never decides; every exclusion is reviewed by a person; a retrospective test on past applications comes first; applicants are told In a randomized test, AI advice lowered pass rates by 9 points; at la Caixa, reviewers rescued 30 of 216 and 35 of 98 flagged proposals

The Swiss National Science Foundation gives reusable rules for funders: applicants remain sole authors, and reviewers form their own judgement first, use settings that prevent storage and training, disclose their use, and leave no role to AI in the funding decision (applicants, reviewers).

Measurement sheet for each use case

Record, before and after activation:

  • the task and its unit (one application, one report, one screened file);
  • time without AI on a batch of comparable tasks, and time with AI including checking and correction;
  • quality judged by a reviewer who does not know which version used AI;
  • rework rate, errors by severity, complaints and incidents;
  • for screening: share of excluded files a person reinstates, change in pass rate, agreement with human reviewers;
  • disclosure status and cost.

Self-reported estimates stay context, even when converted into money. Two school trials run by a UK charity, with timed logs and blind quality review, are a good model of protocol, not a figure to reuse.

Screening and fundraising studies: grant administration evidence note.

To print or save as PDF: Ctrl+P (⌘P on Mac).

Source and history · GitHub