Method: how the evidence was gathered and checked

1. Overview

This study is a structured rapid evidence review with qualitative content analysis, a desk review of documented shortfalls in impact finance. Its reporting is adapted from PRISMA 2020. It is not a systematic review of effects and does not estimate the size of any intervention's impact.

We asked four questions.

1. Where, according to public sources, does impact or blended finance fall short of its stated intent?
2. Which root causes do those sources identify?
3. Which groups bear the cost, and how severe is it?
4. Whose decisions do the shortfalls follow from, and which actors are visible in the evidence at all?

The period covered differs by source type and is set out in section 3. The evidence base comprises 186 source-checked records.

2. Scope and definitions

We use the Global Impact Investing Network (GIIN) definition of impact investing: investments made with the intention to generate positive, measurable social and environmental impact alongside a financial return. We use the OECD definition of blended finance: the use of development finance to mobilise additional finance towards sustainable development in developing countries.

We assess impact through the five dimensions developed by the Impact Management Project and now maintained by Impact Frontiers: what, who, how much, contribution and risk. We adapt them as the article states. Each record is assigned the one dimension in which delivery fell short. "Risk" is used more narrowly than in the source norms, for foreseeable harm that was not managed, or for a process failure with no outcome shown. Where harm occurred, the record is coded to "what", or to "who" if it fell on a group outside the intended one. "Contribution" covers an investor's claimed additionality or contribution that is unproven or absent.

The UNEP FI Sector Impact Matrix is cited as a sector reference. Nothing is reproduced from it, because its licence restricts commercial use. Reader-facing terms follow the Social Finance Academy Impact Finance Glossary.

3. Eligibility criteria

A record was eligible only if it met every inclusion criterion.

At most two records were taken from any one report, so that no source could dominate.

4. Information sources and search strategy

We drew on four tiers of source. The first comprises the public case pages and registers of the independent accountability mechanisms of development banks and development finance institutions. Complaints are a purposive sample from seven such mechanisms and make no claim to be a census. Within each mechanism, cases were taken from its public case listing and screened against the eligibility criteria until a target number for that source was reached; the selection was not random. The mechanisms are the Compliance Advisor Ombudsman for IFC and MIGA, the World Bank Inspection Panel, the IDB Group's Independent Consultation and Investigation Mechanism, the African Development Bank's Independent Recourse Mechanism, the EBRD's Independent Project Accountability Mechanism, the Asian Development Bank's Accountability Mechanism and the EIB Group Complaints Mechanism. The second comprises evaluation units, national audit offices and market surveys. The third comprises reports from aligned organisations: impact measurement bodies, evidence institutes, watchdogs, think tanks and UN analysis. The fourth is news from serious outlets and investigative desks.

Search terms for news included impact fund, impact washing, microfinance or fintech lending, carbon credits and offsets, gender-lens and impact claims.

Several sites blocked automated retrieval, among them those of GIIN, the OECD, ODI, CGD, Eurodad, the African Development Bank Independent Recourse Mechanism and the World Bank Independent Evaluation Group. These were read through a web browser, or from copies downloaded by the human lead.

We searched deliberately for failures by private impact funds. Searches were run through a news archive, news feed services, web browser searches and the websites of aligned organisations. They produced very few in-scope records. The candidates found were rejected as out of scope, mostly ordinary consumer lending or commercial private equity with no impact claim. An initial target that at least 30 per cent of news records concern private investors or intermediaries was therefore withdrawn on 28 October 2026. We report the gap as a finding and have not filled it.

5. Selection process

Counts of pages and search results screened before coding were not logged. The record flow therefore starts at coded records.

Records marked not verifiable were either inaccessible when the independent check was run or carried codes that the check did not support. They appear in no figure. By source tier, the included records come from accountability mechanisms (112), evaluators (33), aligned organisations (21) and news outlets (20).

6. Data extraction and coding

Each record was coded under a written codebook (version 0.2, revised after a 30-record pilot). The coded variables are below.

Every code rests on a verbatim passage from the source. A second passage from the same source may be quoted for root causes. Background knowledge is never sufficient to support a code, and no default group is added. Where several sources report one case, the records are linked. Findings of harm are counted by distinct case; all other frequencies count records, and three cases are reported by more than one record.

7. Quality assurance

All coding was carried out to the written codebook, and every subsequent check described below followed written instructions. A human lead set the codebook, the scope and the decisions recorded in this note, and commissioned each check, but did not code or verify records. We disclose this plainly. The safeguards were:

• Each record was checked against its source by a separate checker that had not seen the coder's reasoning. For every included record, it confirmed the quote, the date, the dimension, each root cause and each group. Records whose source it could not reach were not included.
• Records with unsupported codes were corrected and verified again.
• A consistency pass applied one rule on finding status across the evaluations, so that evaluators' findings of shortfall were not recorded as findings of fault.
• A further check, commissioned by the human lead, reopened a random sample of 20 source-checked records against their sources. Every quote was found. Twelve records passed, seven carried minor coding issues, such as a debatable actor or dimension, and one failed and was excluded as out of scope.
• Four independent reviewers, each instructed to take one perspective (evaluation, investment, affected community and legal), assessed the draft before publication. All four recommended publication with changes, which were made. This was a structured review, not a review by human experts.
• A final fact-check compared every figure in the article and in this note with the analysis dataset, and its corrections were made.

No human coded or verified the full set of records, and inter-rater reliability was not calculated.

8. Analysis

The analysis is descriptive. We report frequencies and cross-tabulations of the evidence base. They are not estimates of how often failure occurs in the market. Root causes and affected groups are counted per code, so one record can contribute up to three of each. Findings are counted by distinct case: the 16 records in which a mechanism, court, regulator or evaluator found harm represent 12 distinct cases.

9. Limitations and potential bias

Selection for failure. The sample was built from sources that document shortfalls. It describes how failure is documented.

Coverage bias by source type. News reflects what journalists covered. Complaints cover only projects with an accountability mechanism, and only where affected people know of it and can reach it. Evaluations reflect what funders chose to evaluate.

Visibility gap for private funds. Failures surface where mechanisms and independent evaluators exist, which mostly means development finance institutions. Few private impact funds publish a comparable complaint or evaluation function. Only one included record has a private investor as its primary actor. The private category in our count also includes enterprises and intermediaries that received development finance, most of which appear because a public lender's mechanism or evaluation exists. The gap concerns visibility and gives no evidence on whether private funds perform better.

Censoring of recent complaints. Of 79 complaints registered from 2023 onwards, 72 have no published finding, and many had little time to reach one. We did not record whether those cases were open, resolved through dispute resolution or closed on other grounds, and we did not measure time to finding or remedy.

Coding and checking. The coder and every checker worked to written instructions, and none was a human expert. A checker that had not seen the coder's reasoning reduces, but does not remove, the risk of shared errors, since checkers of similar design may misread a source in the same way. The only human-commissioned audit covered 20 of 186 records, and it found minor coding issues in seven.

Severity. Severity is the coder's judgement. It does not measure damage, and for allegations it reflects what is alleged. Some dimension and actor codes are debatable, and the choice between "risk" and "what" was applied unevenly in some records.

Language. Most sources are in English, so material published in other languages is under-represented.

Older evaluations. Evaluations older than five years were used only where no newer edition exists.

10. Ethics, naming and right of reply

A named case appears in the article only where a mechanism, court, regulator or evaluator has published a finding, and it is then stated factually with a link. Allegations are labelled as allegations, and no institution or project is named on the strength of an allegation alone. Compliance findings are shown with the clearance, and remain counted as allegations of harm.

Three news allegations against named private companies are withheld from the public listing but counted in the aggregates. The public charts show aggregates only.

To correct a record or exercise a right of reply, write to hello@intellisearch.org. We aim to respond within five working days.

11. Data availability

The explorer lists the included records with source links and passages, except the three withheld news records described in section 10. The full coded dataset and codebook are available on request.