A Trillion-Dollar Industry Deserves a Better Engine
With a new wave of philanthropic capital on the horizon, the most pressing binding constraint on that capital won't be the will to give it away; it will be the operational capacity to do so well.
The Field Today
The world does not lack problems worth solving, and it does not lack evidence about how to solve them. Thousands of peer-reviewed papers are published every week. Development economists run trials in dozens of countries. Public health researchers across the world quantify the burden of neglected conditions in great detail. Scientists across dozens of domains produce findings that accumulate far faster than can be absorbed, let alone acted upon.
Philanthropy’s gift is its ability to catalyze high-impact work that markets won’t fund. And yet, philanthropists can only act on what they can parse. The global knowledge base is vast, multilingual, and growing faster than any team can track. The best program officers in the world are still limited by finite time and attention.
In the field of artificial intelligence, a ‘capability overhang’ refers to capabilities that exist but haven’t yet been fully translated into deployed value. Philanthropy has its own overhang: an enormous body of evidence and intervention ideas have accumulated, but little has been synthesized into actionable form.
This structural gap is about to get more expensive. A new wave of philanthropic capital is on the horizon. The most pressing binding constraint on that capital won't be the will to give it away; it will be the operational capacity to do so well: the people, institutions, and analysis needed to turn funding into impact. Globally, philanthropy is already a multi-trillion-dollar enterprise, but a significant portion of it sits idle in foundations, family offices, and donor-advised funds (DAFs): funding deployed routinely falls short of funding pledged. Raising the efficacy of this capital is one of the transformative levers of our time.
Where Analysis Stops
At the core of every funding decision are two complicated but interrelated questions: how big is the impact of an intervention at scale, and how cost-effective is that intervention? Answering them rigorously, across the full range of what could be funded, is a significant undertaking.
Consider what a serious systematic search actually requires. Is potassium-enriched salt a high-value philanthropic bet for preventing cardiovascular disease? To answer this responsibly, someone has to (at bare minimum):
Synthesize dozens of randomized controlled trials across multiple countries,
Estimate what share of the global sodium-attributable disease burden is realistically addressable,
Model cost-per-life-saved under different deployment scenarios, and
Benchmark all of it against the current best alternatives.
In a best-case-scenario, with no other follow-up work, this will take weeks of research. Multiply that across thousands of potential opportunities that cross the proverbial philanthropic desk in a given month, and the scale of the problem becomes clear.
Yet at the end of the day, a funding decision has to be made, so analysis must stop somewhere – and it often stops short.
Breadth, then Depth
We’re building Overhang, a platform designed to address this analysis gap and augment program officers, field leaders, and domain experts to scope and launch funds at unprecedented scale. Overhang is an agentic research and discovery system that runs structured, iterative search and synthesis loops across defined problem spaces. In a typical workflow, Overhang first conducts broad, breadth-first, parallel search across the landscape of a given problem area, generating taxonomies of thousands of possible interventions. With expert input, it then conducts deep vertical analysis on the top-ranked candidates from that search (as per a target metric), resulting in highly technical reports for domain experts. Effectively, for any given topic, it can, within the span of days:
Map the intervention landscape,
Extract supporting evidence from source literature,
Estimate impact and cost-effectiveness,
Synthesize technical reports for each select intervention
Produce decision-relevant artifacts to guide funding and deployment.
The outputs are analytical: structured documents with cited evidence, explicit assumptions, quantitative estimates, confidence levels, and visible reasoning chains. The following examples provide some texture:
Rodent crop losses: Asked to canvass global agricultural losses from rodent pests, the model synthesized more than 80 peer-reviewed sources, produced economic loss estimates by species and region, estimated rodenticide-attributable animal suffering in terms of comparable lives saved, and identified the highest-uncertainty nodes in the analysis for further investigation.
Potassium salt fortification: The model produced three distinct philanthropic pathways, computed risk-adjusted cost-per-lives-saved estimates for each, and flagged the key behavioral and political barriers that would determine whether any of them could succeed.
Arctic sea ice loss: The model scanned the academic literature and produced a ranked list of intervention candidates across aerosol injections and brightening ocean clouds, complete with causal mechanisms, backing citations, and Fermi estimates of impact in terms of annual saved km². These outputs and reasoning had significant overlap with a report built by a team of climate experts over the course of months.
Overhang is built to be driven by the field leader or domain expert. Its modular architecture allows for customization for a given domain in terms of topic context, problem formulation, cost calculus, and target metrics. One such customization is to tune its criteria to hunt for interventions that are tractable yet neglected. Agent-driven discovery modules scan literature across domains, cultures, and geographies to systematically consider under-explored interventions: agricultural practices with 30-year evidence bases that never reached smallholder farmers; public health interventions validated in wealthy countries but never deployed in poorer ones; governance reforms with strong quasi-experimental evidence that remain politically ignored. These are exactly the analyses that are nearly impossible to produce, permute, and scale by hand, and they are exactly the kind we are designing Overhang to make routine.
Limits and Design
Despite field progress and fast-improving model capabilities, there’s still a lot of work to be done. Some of the limits we’re currently grappling with:
Poor judgment. Models are strong at mechanical work (Fermi estimates, unit conversions, summarizing literature), but weak at discernment that requires taste. Agents are invaluable for breadth and narrowing the search space, but humans ultimately drive the last mile of review, judgment, and decision-making.
Varied output. Generated interventions are noisy, but there’s nearly always a small set worth exploring. Different parameters and model families exacerbate this variation, so domain-specific evaluation is necessary to anchor the experimentation.
Long tails. Models drift towards the center of their training data, which often ends up being the conventional and the already-funded. Fighting this drift is an open problem, given that the valuable interventions often live in the neglected tail alongside noise.
False precision. Estimates remain, at best, estimates. Figures with many decimal places imply a level of confidence rarely supported by evidence.
Even so, relative ranking, narrowed focus and directional accuracy turn out to be very useful signals for guiding discovery, particularly in a philanthropic setting. We’re actively designing in this spirit, shaping Overhang as a recall instrument, and not a precision oracle.
Augmenting the Program Officer
We aim to keep developing Overhang further along the philanthropic workflow – moving from discovery and intervention mapping into policy scanning, talent identification, program design, portfolio management and beyond. We’ll keep testing its capabilities across diverse problem spaces and evaluate it alongside domain experts, whose feedback is the experimental data we’ll use to tune the system.
The goal is not to automate away the judgment of program officers or field leaders. In philanthropy, as in many domains, human expertise, context and domain taste remain irreplaceable. What agentic tooling does is dramatically widen the canvas. The program officer's job doesn't shrink in this model – it expands across a larger surface area of what gets canvassed, analyzed and prioritized, all of which only raises the odds of catalyzing field-changing work.
The Intelligence Layer
Our larger ambition is to build a coherent philanthropic intelligence layer that functions as shared infrastructure for the sector. A collection of tools is useful to the builder; an infrastructure layer is useful to everyone. Over time, the value of this infrastructure compounds across philanthropy. Funders refine opportunities, source teams, and launch funds at unprecedented speed. As this workflow grows more seamless, the flywheel accelerates: more funders find high-impact opportunities across larger scales that they would otherwise have missed, and more grantees find the right funding anchors for their ideas – all faster. All of this is now within reach.
Philanthropic capital today is underused. What stands in the way is not the funds or the will, but the efficiency and efficacy of the infrastructure around it. We’re building the engine to change that.
Overhang is a project of Renaissance Philanthropy, focused on applying AI to improve the quality and reach of philanthropic decision-making. We are in active development and plan to publish insights and outputs as we go. If you want to learn more, share lessons from your own work, or explore collaborations, we would love to hear from you. Please reach out at info@renphil.org