AI can assess information faster than an investment team, but speed is useful only when the system knows what it is being asked to judge. At the point of Impact Due Diligence, investors must first define the intended outcomes, affected groups, evidence thresholds and scoring rules that will shape the decision.
A UK-based experiment by Better Society Capital provides useful evidence. Its AI assessments matched final reviewed scores more often than those produced by deal leads, but only after extensive instructions and human review.
The result should encourage impact practitioners to test AI while giving equal attention to the assessment architecture required first.
What Better Society Capital tested
Better Society Capital assessed 45 startups across three of the Impact Frontiers five dimensions of impact: who experiences the impact, the depth of change and the enterprise’s contribution. This produced 135 individual scores.
Deal leads completed their usual assessments, while Perplexity Deep Research scored the same investments using publicly available information. Better Society Capital used prompts exceeding 200 lines, then asked its Venture Impact Lead and deal teams to review disagreements and obvious errors.
The AI scores matched the final reviewed scores in 81% of cases, compared with 64% for deal leads. The team estimated that an optimised AI assessment could take 15 to 20 minutes, while human assessments ranged from 15 minutes to several hours.
These results suggest that AI can improve consistency and reduce research time, not that a general-purpose model can independently determine impact potential.
The framework did much of the work
The AI did not begin with a blank instruction to assess impact. Better Society Capital had already defined its dimensions, established a four-point scoring scale and developed an Impact Measurement and Management process through years of practice.
Even then, the prompt expanded from 10 lines to more than 200 through repeated testing. The team had to clarify how the model should follow a theory of change, distinguish business customers from ultimate beneficiaries and interpret concepts such as affordability.
One example exposed the difference between absolute and relative affordability. A service could appear affordable because it cost less than market alternatives while remaining beyond the means of the intended lower-income group. Until the definition was made explicit, the AI applied both interpretations.
A model will apply the framework it receives, including its ambiguities. Investors therefore need to state what each criterion means, what evidence qualifies and how trade-offs should be treated before asking AI to score an opportunity.
Due diligence is the point to become explicit
Pre-investment assessment sets the impact terms of the relationship. It establishes the intended outcomes, identifies who should benefit, tests the pathway from product or service to change, and examines whether commercial incentives support that pathway.
Making these elements explicit improves the investment process before AI is introduced. It clarifies approval, enterprise expectations and later portfolio management.
AI can then locate evidence, compare claims with external sources and apply defined criteria consistently. It can also expose assumptions that practitioners understand informally but have never documented, which was one of the most useful effects reported by Better Society Capital.
Applying the method in African markets
The experiment used UK venture investments and public information available online. Applying the same method across African markets would introduce different evidence conditions.
Many promising enterprises have limited public documentation, particularly at an early stage. Material evidence may sit in internal records, customer data, field reports, local-language interviews or conversations with distribution partners rather than in searchable websites and media coverage.
Better Society Capital found that AI made more mistakes where less public information was available and estimated that 10% to 20% of assessments contained hallucinations. Relying heavily on public sources could therefore favour enterprises with stronger communications capacity or greater international visibility rather than stronger impact performance.
A dependable African workflow should combine public research with due diligence documents, management interviews and evidence from intended beneficiaries or affected stakeholders. The assessment framework should also account for local affordability, informal markets, language and uneven data quality without lowering the required standard of evidence.
Building a dependable assessment process
Impact specialists should treat the framework as the first deliverable. The investment team needs to agree on the impact thesis, intended beneficiaries, evidence required for each score, treatment of missing information and distinction between enterprise impact, investor contribution and wider system effects.
Testing should include different business models, sectors and evidence conditions. Where AI and experienced practitioners disagree, the review may identify a model error, a gap in the framework or inconsistency in the human process.
The final assessment should remain a documented human decision. Source citations, confidence levels and unresolved evidence gaps need to remain visible so an investment committee can understand how the conclusion was reached.
Better Society Capital’s experiment provides credible evidence that AI can support more consistent and efficient impact assessment. For investors in Africa, the larger opportunity is to use this moment to make Impact Due Diligence more explicit, comparable and connected to later impact management.
The best starting question is whether the investment team has defined impact clearly enough for a model, an enterprise and an investment committee to apply the same logic. Find out more.
