For fifteen years, funders have asked us essentially the same question: What did this money actually change?
For most of that time, the honest answer was that we could respond rigorously for only a fraction of a portfolio. A funder with several hundred active investments in a country might account deeply for twenty or thirty. The remainder would appear in annual reports, grantee narratives and indicator tables, but not in anything a finance director would recognize as a complete and continuously maintained account of performance.
That was a limit of economics, not of a method.
The work required to build a credible account of one investment is substantial. Documents must be reconciled, results of frameworks interpreted, sector data tested, grantees interviewed, conflicting sources surfaced, and contribution claims reviewed. Repeat that process across hundreds of grants, and the cost becomes prohibitive.
AI is beginning to remove that constraint.
The consequence is more significant than faster reporting. When the marginal cost of maintaining a credible grant record falls, institutions don't need to choose between depth and coverage. They can begin to build an evidence system in which every investment is visible, proportionately assessed, and connected to the decisions it was intended to influence.
Impact accounting is moving from a periodic evaluation activity to management infrastructure.
Two Forces Are Colliding
The first force is financial. Official development assistance fell in 2024 and then contracted sharply again in 2025 — down 23% to US$174 billion, the largest annual drop in its history. Further pressure is projected for 2026: the OECD expects another 7% fall, with bilateral aid to sub-Saharan Africa down around 12% and health financing heading as much as 63% below its 2022 peak. Several major philanthropies are operating against defined time horizons, while governments are increasingly expected to assume responsibilities previously financed or managed by donors.
In that environment, aggregate success stories are no longer enough. Every remaining dollar faces a sharper question: What changed, how confident are we that the investment contributed, and what did it take to produce that change?
The second force is technological. Capable AI has become dramatically less expensive, while newer models can examine long documents, structure dispersed evidence, and support multi-step analytical workflows that were impractical only a short time ago — a more than 280-fold fall in the cost of GPT-3.5-level inference in two years, by Stanford's count.
But cheaper intelligence does not automatically produce better accountability. AI can make an undisciplined reporting process faster without making it more useful. The opportunity appears only when the technology is applied to a well-designed evidence architecture: defined indicators, agreed source hierarchies, explicit contribution methods, documented validation rules and clear human ownership.
These two forces are colliding. Funding is becoming harder to justify at exactly the moment when a much more complete account can be produced.
That resets what funders, public institutions and implementing partners should be expected to know about their portfolios.
Reporting Is Not Accounting
Most institutions already collect enormous volumes of information. Grantees submit reports. Program officers maintain trackers. Evaluations are commissioned. Dashboards aggregate indicators.
Yet the existence of data does not mean that an institution can explain its impact - The Center for Effective Philanthropy found that nearly two-thirds of foundations and nonprofits now use AI, mostly for writing and administration, while 90% of foundations do not support their grantees' use of it.
Reporting asks whether activities occurred and targets were met. Evaluation typically examines selected interventions at defined moments. Impact accounting asks a broader management question:
Across the portfolio, what changed, what is the plausible contribution of each investment, how strong is the supporting evidence, what resources were consumed, and what should leadership do differently as a result?
The distinction matters. A program may deliver its outputs while the sector it is intended to influence moves in the opposite direction. Two investments may report comparable results while relying on evidence of very different strengths. A grant may appear successful in isolation but duplicate an existing government activity or create a reporting system that cannot be sustained locally.
Impact accounting does not eliminate those tensions. It makes them visible.
The unit of value is therefore not the report. It is the decision-ready grant record: a living, reviewable account connecting the investment’s intent, activities, contribution, resources, evidence of strength and sector context.
What “being ahead” actually means
Nearly two-thirds of foundations and nonprofits now report using AI, but much of that use remains concentrated in writing and administrative work. At the same time, most foundations are not supporting grantees to develop equivalent capabilities. AI is helping the sector write about impact more efficiently before it has transformed how impact is accounted for.
Being ahead means something more specific.
Every active investment should be able to answer three questions, at a depth proportionate to its significance:
- What is changing in the surrounding sector or system?
- What is the funder’s plausible contribution, and how strong is the evidence supporting it?
- What financial, organizational and public-system resources are required?
“Contribution” is deliberate. Complex social outcomes rarely belong to one institution. AI should not manufacture attribution where the evidence supports only a plausible contribution. Its role is to assemble, compare, and test the evidence so that experts can make the limitations of each claim as visible as the claim itself.
Those answers should also operate as living records, refreshed when meaningful evidence changes, rather than reconstructed retrospectively for the next annual review.
Finally, the evidence base should belong to the institution using it. A funder, ministry or country office should not become dependent on the consultant, platform or individual who initially assembled the system. Institutional ownership is part of the methodology, not a handover activity at the end.
Built at Country Scale, Now Running on AI
Impact accounting is where more than a decade of designing results frameworks, monitoring systems and evaluation approaches for funders and governments was always heading. Its binding constraint was not a method. It was the expert time required to apply the method consistently across an entire portfolio.
A rigorous cycle might involve investment documents, results-framework mapping, sector-data analysis, grantee interviews, source reconciliation and executive summaries capable of surviving leadership scrutiny. That work could take around eight months, which is why comprehensive treatment was normally reserved for priority investments.
Coverage was limited by cost, not by the absence of a method.
In a live East African country program for a major philanthropic funder, impact accounting has moved from a supporting exercise to the backbone of the health-strategy narrative. Close to thirty priority investments across maternal, newborn and child health, family planning, nutrition and immunization carry contribution maps. Sector dashboards incorporate new survey rounds. First-hand grantee interviews anchor executive summaries. Indicators are aligned with national health information systems so that the portfolio does not create unnecessary parallel collection.
Where data sources disagree, as they often do in immunization and nutrition, the system does not quietly select the most convenient number. The competing sources are presented side by side.
That principle is crucial. A credible impact-accounting system should not make uncertainty disappear. It should make uncertainty manageable.
Leadership direction has remained consistent: fewer indicators, a clearer account of change, and insights that lead to decisions.
AI Changes the Production Model, Not the Standard of Proof
On this live portfolio, AI production runs are already supporting two parts of the operating model.
First, sector pages can be refreshed from established sources with repeat effort reduced from approximately 114 consultant-hours to 26 hours per cycle.
Second, AI-supported document analysis, interview structuring and results-framework mapping are compressing a grantee assessment cycle from approximately eight months toward three or four, at roughly half the cost. Annual-report cross-checks are applied after programme-officer sign-off.
Every AI-supported run is validated against the established manual method before it becomes part of the operating model.
The control model remains clear:
- Rules-based automation handles stable, repeatable steps.
- AI prepares first-pass structures, classifications and evidence summaries.
- Named experts review anything causal, interpretive or leadership-facing.
AI does not independently author a contribution claim, counterfactual, target or recommendation.
This division is not a temporary safeguard until models become more capable. It is the design of a responsible system.
Social-impact evidence contains ambiguity, unequal data quality, political context and competing interpretations. Greater model capability does not remove the need for accountable judgement. In many cases, it increases the need to specify where that judgement sits.
From Portfolio Reporting to Public Decision Infrastructure
The most important effect of this model is not simply that funders can review more grants.
When grant evidence is aligned with national indicators, government reporting routines and leadership decisions, impact accounting can become part of the country’s own decision infrastructure. The same record that helps a foundation examine its contribution can help a ministry compare implementation across regions, identify a delivery bottleneck or decide where capability needs to be transferred.
That changes the intended endpoint.
The goal is not a more sophisticated donor dashboard that disappears when an engagement ends. It is a durable evidence architecture that governments, country offices and implementation partners can continue to use. Recent Evalueserve delivery experience in Nigeria reflects this broader direction through state-level delivery and insights units, governance routines, leadership reporting, decision tracking, analytics, and capability transfer to government and local ecosystem stakeholders.
A system that remains externally owned may produce excellent analysis and still fail institutionally. Sustainability must therefore be measured not only by whether the technology continues running, but by whether local organizations can question the evidence, maintain the process, and act on what it shows.
Why the Economics Change the Operating Model
When a validated grant record costs materially less to produce and maintain, coverage stops being an unavoidable budget limitation. It becomes an operating-model decision.
A tiered portfolio can combine:
- deep, consultant-led assessments for strategically important or high-risk investments
- AI-supported, human-validated records for the majority of active grants
- proportionate registry entries for small, early-stage or closing investments.
This does not mean applying one standard of analysis indiscriminately across every investment. It means ensuring that nothing disappears from institutional view.
Evaluation resources can then move away from repeatedly reconstructing basic evidence and toward the work that remains intrinsically human: interviews, contextual judgement, relationship-building, challenge and decision facilitation.
A common methodological core also becomes possible. Indicator libraries, source rules, results-framework schemas, interview tools and contribution methods can be established once, then configured for different countries and programs. That creates comparability without pretending that local contexts are interchangeable.
Three moves institutions should make now
First, define proportional accountability across the full portfolio.
Assign every investment an evidence tier. Do not confuse receipt of an annual report with a complete account of contribution, evidence strength and resource use.
Second, build evidence discipline before automating it.
Agree indicator definitions, source hierarchies, validation rules, contribution standards and ownership. AI applied to an undisciplined process produces more output, not more confidence.
Third, test the operating model on live work.
Run an AI-supported approach alongside the established manual method. Compare the outputs, document disagreements and specify what would justify scaling, adjusting or stopping. Include institutional ownership and capability transfer in the design from the beginning.
The institutions that make these moves will not simply report faster. They will see more of their portfolios, identify weak evidence earlier, compare investments more consistently and redirect resources while action is still possible.
The future of impact accounting is not a perfectly automated verdict on whether a grant “worked.” Social change is too complex, and causal evidence too contextual, for that.
The more credible ambition is also the more powerful one: every grant visible, every claim contestable, every uncertainty explicit and every major decision connected to evidence.
For the first time, that ambition is becoming economically achievable.
The question is no longer whether a funder can account for every grant.
It is whether it will choose to.
