AI in philanthropy is beginning to influence how foundations, donors, and nonprofit organizations research social needs, manage resources, communicate with communities, and evaluate results.
Its value, however, will not be determined by how many tools an organization adopts, but by whether those tools improve decisions without weakening the trust on which philanthropy depends.
Executive summary
Foundations and nonprofits are already using artificial intelligence for communication, document analysis, fundraising support, and internal productivity. Available research also shows that adoption is moving faster than governance, staff training, and meaningful measurement.
A responsible strategy must distinguish between automating routine work, improving organizational decisions, and funding AI-based solutions for the public good. It must also protect donor and beneficiary data, examine bias, document how systems are used, and preserve human oversight for consequential decisions.
The goal should not be to turn philanthropy into an automated process. The more valuable opportunity is to free staff capacity, identify unmet needs, and direct resources more effectively without reducing people to predictions, profiles, or algorithmic scores.
What is AI in philanthropy?
In this article, the term refers to the responsible use or funding of artificial intelligence systems that support charitable giving, grantmaking, nonprofit operations, impact measurement, and public-interest solutions.
It includes much more than using a generative assistant. It also covers data governance, equity, community involvement, risk evaluation, transparency, and responsibility for outcomes.
Five key takeaways
- Adoption should begin with low-risk tasks whose outputs can be verified.
- Beneficiary information requires stronger safeguards than ordinary marketing data.
- Affected communities should participate before a system is funded or deployed.
- Every investment needs measures for mission, risk, equity, and sustainability.
- Executive summary
- What is AI in philanthropy?
- A large giving sector with uneven technological capacity
- A decision map for responsible adoption
- Three layers philanthropic strategies should separate
- 1. AI used inside an organization
- 2. AI embedded in social programs
- 3. AI funded as public-interest infrastructure
- The central gap is governance, not access
- The dual-impact test: mission and trust
- Moving equity from awareness into practice
- Mistakes that turn innovation into an additional burden
- Purchasing a tool before defining the problem
- Entering sensitive data without reviewing the terms
- Automating decisions that require context
- Measuring volume instead of outcomes
- Funding pilots without a sustainability plan
- Ten questions before funding or deploying a system
- What this shift means for U.S. philanthropy
- Frequently asked questions
- The judgment that must remain human
A large giving sector with uneven technological capacity
Charitable giving to U.S. organizations reached an estimated $617.20 billion in 2025, representing nominal growth of 5.7 percent. Foundations contributed approximately $117.15 billion of that amount. The scale of these resources makes disciplined technology governance a material issue for the U.S. philanthropic sector.
AI use is already widespread, although adoption estimates differ according to survey population, organizational type, and the definition of “use.” The Center for Effective Philanthropy found that almost two-thirds of the participating foundations and nonprofits used AI, primarily for internal productivity and communication. The same research found that 90 percent of foundation leaders said their institutions provided neither financial nor nonfinancial support for grantees’ AI implementation.
A separate survey of 346 nonprofits conducted in December 2025 reported that 92 percent were using AI in some capacity, while only 7 percent reported major improvements in their ability to advance their missions. Because this commercial study focused heavily on fundraising practices, its findings should not be generalized to the entire sector, but they help illustrate the gap between experimentation and measurable organizational transformation.
A decision map for responsible adoption
The first question should not be which platform to purchase. It should be which problem warrants a technological intervention and what evidence would demonstrate that the intervention worked.
| Scenario | Recommended decision | Evidence required | Boundary to preserve |
|---|---|---|---|
| Internal drafting and summarization | Begin with a supervised pilot | Net time saved, correction rate, and final quality | Do not enter confidential data into unauthorized tools |
| Grant research | Use AI to organize and explore information | Traceable sources and review of every recommendation | Do not present automated suggestions as verified facts |
| Donor segmentation | Proceed only with data governance | Consent, data quality, and nondiscrimination testing | Avoid inferring health, vulnerability, or other sensitive traits |
| Proposal review | Limit AI to administrative support | Published criteria, auditability, and human review | Do not delegate the final funding decision |
| Beneficiary-facing programs | Co-design with affected communities | Utility, safety, accessibility, and outcome testing | Do not experiment on vulnerable groups without safeguards |
| Funding public-interest AI | Evaluate the problem, team, data, and sustainability | Theory of change, technical evidence, and maintenance plan | Do not confuse a compelling demonstration with proven impact |
Three layers philanthropic strategies should separate
1. AI used inside an organization
The first layer includes administrative support, communication, translation, document search, preliminary analysis, and fundraising assistance. Its central promise is to reduce operational workload, but that benefit can only be established after accounting for review time, corrections, training, implementation, and error management.
TechSoup and Tapp Network reported that 85.6 percent of participants were exploring AI tools, while only 24 percent had a formal strategy. Forty-three percent depended on one or two staff members for technology or AI-related decisions, a particularly relevant capacity constraint for smaller organizations.
2. AI embedded in social programs
At this level, systems may help expand access, adapt educational materials, identify patterns, improve accessibility, or support field decisions. The risks also increase because automated outputs may affect people who did not choose to use the technology and may lack an effective way to challenge its conclusions.
A technically functional model is therefore not sufficient. Organizations must examine who is represented in the data, which groups could be burdened by errors, how mistakes will be corrected, and what human alternative will remain available.
3. AI funded as public-interest infrastructure
The third layer emerges when foundations finance research, open tools, training, data infrastructure, or AI solutions developed by mission-driven organizations. In April 2026, Candid reported that 84 percent of the AI-powered nonprofits surveyed by Fast Forward needed funding to develop and scale their solutions. The analysis also emphasized the importance of flexible, early-stage capital before projects can demonstrate results at scale.
This approach prevents AI in philanthropy from becoming limited to purchasing commercial products. Philanthropic capital can also strengthen shared infrastructure, independent research, accessible tools, and the participation of communities that are frequently excluded from technology design.
The central gap is governance, not access
Generative systems may produce inaccurate information in a persuasive form, combine reliable facts with unsupported inferences, reproduce existing bias, or expose information to external providers. These limitations become especially consequential when systems handle beneficiary narratives, health records, immigration information, financial data, or information concerning children.
The National Institute of Standards and Technology organizes its voluntary AI Risk Management Framework around four functions: Govern, Map, Measure, and Manage. NIST states that the framework is designed to incorporate trustworthiness considerations into the design, development, use, and evaluation of AI systems. NIST also notes that AI RMF 1.0 is currently being revised.
For a foundation or nonprofit, the four functions can be applied as follows:
Govern
Assign accountability, approve an acceptable-use policy, classify data, define prohibited activities, and establish which decisions require human authorization.
Map
Identify affected people, the operating context, potential harms, external providers, and the consequences of incorrect outputs.
Measure
Evaluate accuracy, bias, security, accessibility, usefulness, cost, complaints, and mission-related performance.
Manage
Correct failures, restrict use, suspend unsafe tools, respond to incidents, and maintain alternatives that allow services to continue.
The OECD AI Principles complement this approach by promoting human-centered systems that support fairness, transparency, explainability, robustness, safety, security, and accountability. The principles were updated in May 2024 to reflect new technological and policy developments.
The dual-impact test: mission and trust
Before describing an initiative as successful, Roberto Isaias recommends examining two outcomes together: whether the system improves the organization’s ability to fulfill its mission and whether it preserves the trust of donors, employees, beneficiaries, and partners.
A tool might analyze hundreds of documents more quickly than a program team and still fail if inaccurate information enters a recommendation. It might reduce administrative costs but lose legitimacy if communities later discover that their information was reused without a meaningful explanation.
The dual-impact test asks:
- Does the system improve an outcome the organization defined in advance?
- Do affected people understand when an automated system is involved?
- Is there a practical way to review, challenge, or correct an outcome?
- Can the organization explain which data are used and why they are necessary?
- Does the benefit remain positive after risk, cost, and oversight are included?
A negative answer does not always require abandoning the project. It may indicate that the use case should be narrowed, that a smaller pilot is necessary, or that the underlying problem requires organizational reform rather than a technological product.
Moving equity from awareness into practice
Candid’s survey of 850 nonprofits found that 80 percent were familiar with AI and 64 percent were familiar with the possibility of AI bias. Yet only 36 percent reported implementing equity practices, and approximately 15 percent had an organizational responsible-use policy. The findings suggest that familiarity with ethical concepts does not automatically create the time, budgets, procedures, or accountability required to apply them.
The Center for Effective Philanthropy also found that more than 80 percent of participating nonprofits were not involved in activities intended to advance equitable AI. Almost two-thirds of foundation and nonprofit leaders said that none or only a few members of their staff had a solid understanding of artificial intelligence and its applications.
For Roberto Isaias, an equity commitment has limited value when it remains a general statement. It should shape data selection, budgets, vendor procurement, testing, community participation, and grievance mechanisms.
Mistakes that turn innovation into an additional burden
Purchasing a tool before defining the problem
What happens: An organization buys licenses because the platform appears innovative.
Warning sign: There is no baseline, prioritized need, or person accountable for results.
Better approach: Define the problem, document the current process, and conduct a time-limited pilot.
Entering sensitive data without reviewing the terms
What happens: Staff upload case files, donor lists, or beneficiary narratives to external services.
Warning sign: No one can explain whether the data are stored, reused, or used to train a model.
Better approach: Classify information, restrict access, review contracts, and use approved environments.
Automating decisions that require context
What happens: A summary or scoring system effectively determines which proposal receives support.
Warning sign: The organization cannot reconstruct why an applicant was prioritized or excluded.
Better approach: Preserve published criteria, human judgment, and a reconsideration process.
Measuring volume instead of outcomes
What happens: Generated documents, processed applications, or delivered emails are reported as impact.
Warning sign: The activity cannot be connected to a mission-related result.
Better approach: Measure quality, access, program outcomes, equity, unintended effects, and trust.
Funding pilots without a sustainability plan
What happens: A project operates during the grant period but collapses when a contract ends or provider costs rise.
Warning sign: There is no maintenance budget, data ownership agreement, or exit strategy.
Better approach: Evaluate total cost, vendor dependence, interoperability, and local capacity.
Ten questions before funding or deploying a system
The editorial framework used by Roberto Isaias recommends asking:
- What specific problem is the organization trying to solve?
- Is there a simpler, less expensive, or more human alternative?
- Who will benefit, and who may carry the risk?
- What data does the system require, and how were they obtained?
- What mistakes can it make, and what would their consequences be?
- How will affected communities participate?
- Which decisions will remain under human responsibility?
- How will the provider, model, and full process be audited?
This filter allows organizations to evaluate quality and accountability without requiring every nonprofit to become a technology laboratory.
What this shift means for U.S. philanthropy
The United States combines substantial charitable resources, an advanced technology ecosystem, and an extremely diverse nonprofit sector. That combination can accelerate innovation, but it can also increase the gap between institutions with specialized teams and community organizations with limited technical capacity.
Candid reported that 90 percent of participating nonprofit leaders were interested in using more AI, while only 17 percent said their funders had engaged them in conversations about AI-related needs. This suggests that support should not be limited to software licenses. It should also cover training, governance, research, security, staff time, and evaluation.
AI in philanthropy will create more sustainable value when foundations treat technological capacity as organizational infrastructure, alongside leadership development, evaluation, administration, and institutional strengthening. Funding a tool without funding the conditions required to use it responsibly transfers the implementation risk to the grantee.
Frequently asked questions
The judgment that must remain human
Technology can process more information than a team could review manually, but it does not possess the legitimacy to define social priorities on its own. That legitimacy comes from mission, evidence, community participation, and the accountability of those entrusted with charitable resources.
The perspective associated with Roberto Isaias treats AI in philanthropy as a capacity that should remain in service of people, rather than a substitute for the human relationships that make generosity possible. The most valuable advance will not be the system that automates the greatest number of decisions, but the one that helps organizations act with clarity, learn from evidence, and expand impact while preserving trust.
