A foundation approves a three-year grant to build an AI tutoring tool for under-resourced classrooms. The diligence is thorough, the theory of change is sound, the board is satisfied. By the time the second tranche is released, the underlying models have turned over twice, a commercial product does most of what the tool was designed to do, and the grantee, funded to execute a fixed plan rather than adapt as the technology evolves, is a year into executing a plan that the technology has already overtaken.
Nobody made a mistake. The process worked exactly as designed. And that is the problem.
Stories like this are becoming routine, and they expose an uncomfortable friction. The processes and disciplines that make philanthropy trustworthy, including patience, proof before capital, and deference to grantee autonomy, also make it slow. In most domains where philanthropy operates, slow is a virtue given the high cost of failure. But in the age of AI, slow can mean arriving late and funding yesterday’s solution.
In moments like these, philanthropy reaches for the venture philanthropy playbook, urging foundations to think in portfolios, actively partner, and orient to results. But simply borrowing from venture capital comes with tradeoffs. Philanthropy cannot gain speed and adaptability at the expense of the independence, accountability, and stewardship that distinguish it from private capital.
This article identifies five ways philanthropy must adapt for the AI era: what to fund, whom to partner with, how to make decisions, how to engage after the check clears, and what capabilities are required. We also address the difficult tradeoff between preserving philanthropy’s traditional approach and achieving the speed this moment demands. Change is overdue. But moving faster can introduce new risks, and the same practices that make a funder faster can also make it less accountable. Philanthropy must therefore be deliberate about what to change, what to preserve, and what new risks it is willing to take on.
Five Shifts for Philanthropy in the Age of AI
1. What to fund: establishing the floor where the market won’t invest
The most important contribution philanthropy can make starts with a key question: what will markets systematically fail to address? The answer defines philanthropy’s most valuable and least replicable role—and it is increasingly ambiguous in the AI era.
There are crucial problems that the private market is unlikely to solve on its own because the commercial opportunity is relatively modest, even though solving these issues could create large benefits for society. Here are some examples:
- Markets that are too small or insufficiently profitable to attract sustained AI investment. Education, for example, has a declining K-12 and postsecondary market representing ~$20 billion to $25 billion, compared with a growing ~$45 billion to $50 billion health care market. The same dynamic applies to other sectors, including global health, social services, and legal aid, which are areas with limited commercial incentives but significant public value, particularly for low-income populations.
- Populations and languages that are commercially underserved. Populations and less widely spoken languages that are poorly represented in training data or offer limited commercial opportunity can receive less investment in AI tools tailored to their needs.
- Workforces that lack AI capabilities. The AI technology may exist, but the people responsible for deploying it for public benefit lack the training and resources to use it effectively. The AI fluency of educators, community health workers, public defenders, and other mission-specific professionals will determine whether these tools equitably reach and benefit the populations they serve.
Across these gaps, philanthropy is uniquely positioned to absorb early risk, steward shared assets, and build capacity where the market won’t. Key investments include the following:
- Public goods. Philanthropic organizations may invest in verified real-world datasets, open evaluation infrastructure, open-source model architectures, compute access, and shared technical standards for high-stakes contexts. These foundational resources need to be openly accessible and responsibly governed, yet few commercial actors have an incentive to build and maintain them. Philanthropy is uniquely positioned to fill that gap.
- Capability building. Philanthropic organizations may invest in field-level AI fluency programs for mission-sector workers to address skills gaps that commercial providers have little incentive to serve. This includes building the training infrastructure the philanthropic sector itself needs to develop and sustain AI capabilities.
The risk: Infrastructure funding is a step removed from the people it serves, making it harder to justify to boards and communities seeking to measure direct impact. It also raises a governance question that is critical to get right: who controls what the funder builds?
2. Whom to partner with: engaging with AI labs and hyperscalers
Deciding what to fund is only half the question. The other half is whom to partner with to achieve maximum impact. VC firms have always understood that portfolio success depends as much on relationships as on capital. In the AI era, VC firms have cultivated direct relationships with frontier labs and major cloud providers, such as OpenAI, Anthropic, Google, and Microsoft, to secure early intelligence, compute access, and insight into what’s coming. Philanthropy hasn’t yet built this at scale, but doing so will increasingly be necessary if it wants to influence the most consequential AI decisions being made.
Frontier labs are not grantees and they don’t need philanthropic capital. They are the most powerful direction-setting actors in the field philanthropy is trying to shape. Partnering with them may also be genuinely uncomfortable. Philanthropy typically avoids picking “winners,” and the foundational models these labs have built weren’t designed to represent the populations philanthropy exists to serve. Given these tensions, the relationship needs to be peer-to-peer rather than funder-to-grantee, grounded in a clear-eyed understanding of what philanthropy brings to the table and what it needs in return on behalf of the populations it serves.
What philanthropy brings: Philanthropic organizations offer public trust, accountability to communities under-represented in training data and deployment decisions, and a network of grantees and experts whose ground-level insights can help labs understand and demonstrate real-world impact beyond synthetic testing.
What it needs in return: Philanthropic organizations need visibility into how model capabilities are expected to evolve, compute access for mission-driven users who can’t negotiate it alone, and a voice in shaping frontier model investments, including training data and open data connectors that determine whether those models work for mission-driven use cases.
In practice, this will require a deep understanding of the priorities and technological trajectories of frontier AI labs, structured relationships with their deployment and safety teams, and the ability to negotiate compute and model access on behalf of grantees who cannot secure it themselves. All these changes impact the talent and capabilities that philanthropic organizations require.
The risk: Philanthropy must guard against becoming too embedded in a lab’s ecosystem or it risks compromised independence and impact. It is critical to actively foster its distinct point of view on where commercial and public interests diverge.
3. How to decide: adaptive evaluation for an uncertain environment
The standard philanthropic decision framework often privileges certainty. If the organization can demonstrate that an intervention is likely to work, then it will be funded. If not, then funding is withheld. AI complicates that standard because funders must ask not only whether an intervention will work but also whether the underlying technological assumptions will still hold by the time it is deployed.
In an environment where the field can change almost daily, excessive caution carries its own cost. Backing only proven approaches risks investing in capabilities LLMs (Large Language Models) already provide and leaving philanthropy in a reactive stance of responding to problems after they emerge rather than helping prevent them in the first place.
To counter this risk, VC-style adaptive evaluation starts with a different question: not “Is this safe enough to fund?” but “What risks are we taking, and does the expected return justify it?” That means assessing risk across four dimensions:
- Solution risk. Will the approach—whether an AI technology, training model, or organizational change process—work as intended for the populations it is designed to serve?
- Implementation risk. Can this grantee execute as needed and adapt as conditions change, particularly given AI’s rapid evolution?
- Market risk. Is the market expected to address this need? If not, is there sufficient demand for meaningful uptake, and is there a credible path to sustainability?
- Mission risk. Does the intervention advance our theory of change without causing harm?
High technical or market risk should not automatically disqualify a high-impact investment if mission risk is low, expected social impact is high, and the investment can be designed to surface and resolve uncertainty quickly.
In practice, adaptive evaluation requires three things: milestone-gated funding sequenced to the investment type (for example, solution-oriented grants with check-ins every 6–12 months and infrastructure investments with longer cycles and explicit governance triggers); a small allocation for exploratory bets, sized so that failure generates learnings rather than crisis (for example, 5% to 10% of a program area); and embedded leading and lagging metrics tracked throughout development and implementation.
The risk: “Embrace failure” is easy to say but expensive in practice. Boards face genuine credibility costs when bets fail publicly. With that in mind, this shift requires redefining what accountable risk taking and “success” look like when not taking risks is itself a form of failure.
4. Post-grant engagement: evolving what “trust-based” philanthropy looks like
VC investors stay close to founders after the check clears. They maintain regular contact, offer strategic guidance, challenge pivots, and make valuable connections with people and resources in their network. Evidence suggests that this matters: startups receiving meaningful post-investment support experience 45% lower failure rates and recover 60% faster from setbacks than startups that don’t receive such funding.
By contrast, most philanthropic grantees encounter their funder primarily through reporting cycles. This restraint has developed for good reason: to preserve grantee autonomy, resist a funder’s temptation to substitute its judgment for the grantee’s, and avoid creating dependency. Those concerns remain valid, but in a rapidly evolving field, disengagement carries growing costs. When tools turn over in months, a grantee without access to the right expertise at the right time risks becoming obsolete before their project has even gotten off the ground.
A meaningful tension here is that the very things that make post-grant engagement valuable are the same things that can make it dangerous. A funder can help a grantee navigate a pivot while also, consciously or not, nudging it toward the foundation’s preferred direction. The asymmetry of dollars and authority means grantees are rarely in a position to police the difference. Engagement that feels like partnership from the funder’s side of the table can feel like direction from the other side, particularly when the grantee’s next tranche depends on the relationship remaining warm.
To address this risk, the question is how to structure engagement so that it functions as an offer rather than a mandate—one that grantees can decline without consequence. In practice, that means offering the following:
- Predictable, low-friction access: office hours and cross-portfolio convenings that grantees can access when useful
- Practical expertise on demand: AI and domain knowledge when grantees face specific obstacles, sized to the grantee’s need
- Cross-grantee learning: connecting grantees who are facing similar challenges so insight travels without the funder serving as the conduit
- Real-world feedback loops: introductions to users and domain experts who can rapidly pressure-test work before it scales
- Leadership coaching: support for navigating organizational change and succession planning, shaped by the grantee rather than the funder
The risk: The line between active partnership and control is thin, and only grantees can see whether it is being crossed. To deliver on the promise of genuine partnership, grantees must be able to decline engagement without fear of penalty.
5. Building the teams that AI-era philanthropy requires
In the AI era, grant makers must be able to evaluate infrastructure investments, engage frontier labs credibly, interrogate grantee technical claims, make decisions amid uncertainty, design milestone-gated evaluation frameworks, and provide meaningful support to grantees. To deliver on these requirements, philanthropic organizations need a new mix of talent, organizational structures, and a clear model for how technical expertise and philanthropic judgment interact. Two talent shifts are particularly important:
- “T-shaped” talent that pairs AI fluency with domain expertise. While every role will benefit from a common floor of AI literacy, the question for philanthropy is no longer “Do our people understand AI?” It is “Do the right people have the right level of technical depth for the decisions they are making?” This will require grant makers to pair AI fluency with the domain expertise and philanthropic judgment they have traditionally prioritized.
- Dedicated technical talent. In addition to ensuring deeper technical expertise across all employees, building genuine AI capability will require philanthropic organizations to recruit roles they have not traditionally employed, including AI engineers and data scientists who can engage frontier labs, evaluate technical claims, and translate model capabilities into mission-relevant terms.
Achieving this shift requires the following:
- Continuous learning as an organizational norm. Staying current on the latest AI technologies should be treated as a professional expectation, with dedicated time for learning rather than leaving it to individual motivation.
- Role-specific training with a consistent baseline. Establish a foundational curriculum paired with the domain depth and technical specificity each function requires rather than delivering a single program uniformly across the organization.
- Structured career pathways. Use rotations, shadowing, and live grant reviews to build judgment through applied experience, not classroom instruction alone.
- Targeted hiring. Recruit deliberately for roles such as data architecture, technical program management, and frontier lab engagement, with the same strategic intent as senior program officer hiring.
- Cross-functional team design. Integrate technical and domain expertise at the point of decision rather than keeping them in separate tracks. Program officers working alongside data architects and technical staff improves the quality of investment decisions.
When it comes to capability building, the same logic that supports shared AI infrastructure for grantees applies to philanthropy itself. Cross-foundation learning networks and pooled training infrastructure are precisely the kind of sub-exponential investment that no single foundation should build alone—and that could democratize AI capability across funders of all sizes rather than concentrating it among the largest organizations.
Reckoning with a core tension: philanthropy + influence
While each of these shifts is individually about effectiveness, taken together, they also give philanthropy greater influence over what gets built, the relationships that set the sector’s direction, and how grantees navigate pivots. At a moment when influence is already concentrated among a small number of labs and investors, philanthropy has an important role to play in addressing the gaps. In doing so, it must avoid replicating the same concentration of power it seeks to counterbalance.
The VC model offers little guidance here. VC investment decisions are typically concentrated among a relatively small group of investors, with limited participation from those ultimately affected by them. Importing its disciplines without correcting for its asymmetries means importing those asymmetries too. The key is to build safeguards into the operating model: governing shared infrastructure collectively rather than in-house, keeping frontier lab relationships independent from internal decision-making, ensuring that grantees can decline post-grant engagement without consequence, and giving the people most affected by these decisions a meaningful role in making them.
What AI-first philanthropy looks like: an early vision for the near future
The philanthropic organization built for this moment looks different from the ones that currently exist, starting with who is in the room, how quickly decisions are made, and how AI itself changes the work of philanthropy.
Fund the floor. The AI-first philanthropic organization invests in the public goods, shared infrastructure, and capability building that markets will not provide, governed through structures that survive its own exit. It cultivates relationships with a frontier lab from a position of independence, using its network as leverage, and engages grantees as genuine partners.
Move fast. Grants are structured in tranches rather than multi-year commitments, with initial bets sized to generate specific signals and pre-negotiated milestones guiding go/no-go decisions. Decision authority is distributed; program officers with sufficient technical depth can advance exploratory bets within defined parameters, while the board’s role shifts to setting portfolio strategy and approving the framework. The standard for a good decision shifts accordingly—not “Is this certain enough?” but “Are our assumptions explicit, do we understand the risks, and do we know what evidence would change our view?”
Leverage scale. Philanthropy partners with a new class of organizations, including hyperscalers and frontier labs that are shaping what AI can do and for whom. It also co-invests deliberately, identifying co-funders whose geographic reach, sector depth, policy relationships, and financial resources can extend the impact of investments beyond what any single funder can achieve.
Operate AI-first. Program teams are cross-functional by design, with program officers working alongside data architects, technical program managers, and AI practitioners at the point of decision rather than consulting them sequentially. Human-to-agent collaboration becomes a core way of working, with AI actively used to synthesize grantee data, identify early signals of progress or drift, and support faster, better-informed decisions across the portfolio. The same case philanthropy makes for AI adoption among its grantees applies to itself.
All these shifts are a set of deliberate tradeoffs: speed against safeguards, influence against accountability, and effectiveness against the question of whose interests the organization serves. The foundations that navigate those tensions deliberately, and design mechanisms to mitigate potential risks, will drive disproportionate impact going forward. The ones that don’t risk retreating into an increasingly narrow version of what philanthropy can accomplish.
