When Capria decided to become an AI-native fund in 2023, we made one commitment above everything else: AI would adapt to us, not the other way around. We would never outsource our thinking to it, let it reshape our thesis, or allow it to substitute our judgment. We would use it as a force that helps us multiply ourselves, something that lets us get more done and think more sharply.
Three years in, AI runs across every corner of the fund. Investing, marketing, investor relations, compliance, reporting. Most of those were quick to build and quick to pay off. But the place that mattered most, and the hardest to get right, was investment analysis itself. Our thesis, our priorities, and our judgment all come into play when evaluating an opportunity. How do you embed all of that into AI in a way that holds up across every deal and adds value to our team’s thinking without replacing it?
For us, the answer came down to two things. The first was alignment, making sure the AI models we used understood our investment thesis well enough to work within it. Capria invests across the Global South, which meant the model needed to understand not just how we evaluate companies but the emerging market context every opportunity sits within.
That part was straightforward. We have a carefully documented thesis and clear investment mandates, which gave us a strong foundation for building the system prompts, context, and guardrails that keep AI in our lane. The second one was dealing with non-determinism, and that turned out to matter a great deal.
Sorting through randomness
Large language models, or LLMs, are probabilistic machines. Even when given the exact same input, they do not produce the exact same output. They sample from a range of likely responses, which means running the same prompt twice can give you two completely different answers. For finding the best Lebanese restaurant in town, this might not be a problem, but for investment analysis it can be a nightmare.
Just imagine the following. You upload the same investment deck twice into a generalist AI tool and ask it to score the company’s go-to-market strategy. The first run gives a 7 out of 10, highlighting the founders’ clear ideal customer profile and channel strategy. The second run, with the exact same deck, returns a 5 out of 10, raising concerns about customer acquisition cost assumptions the model conveniently ignored the first time. Run thousands of those a year and the randomness piles up. You cannot reproduce decisions, you cannot clearly defend them, and over time you cannot tell whether your judgment is improving or whether the model is quietly steering it.
Quantifying our way to consistency
To solve this, we built a quantitative evaluation framework. Instead of asking the model for a holistic assessment based on broad criteria, we make separate, targeted LLM calls, each one focused on a single evaluation area and tasked with extracting specific quantifiable metrics from the company’s materials. These can be anything from founders’ years of domain experience to unit economics indicators or impact metrics like lives impacted or CO2 avoided. Those metrics then feed into predefined scoring rules that produce a numerical score for each area. The rules stay the same across every deal, and the benchmarks behind them were built from more than 12 years of investment memos, which means the framework reflects how we actually evaluate companies, not generic best practices.
In this way, you are combining the ability of LLMs to read, interpret, and extract meaning from qualitative materials with the consistency of mathematical calculations. What we get back is not a long AI-generated report you have to decipher. It is a transparent set of scores grounded in quantifiable data, with sources attached to every metric, that give you a clean read on a company in seconds. No verdicts, no subjective explanations, just numbers you can actually trust to steer your attention in the right direction.
Consistency builds leverage
Because the score comes from predefined mathematical rules rather than open-ended AI reasoning, the same materials produce much more consistent results. Variance is not eliminated, but it is reduced to decimals in scores.
In practice, that changes the rhythm of the whole process. Every deal starts from the same baseline, which means our team can get to the interesting questions faster. Deal screening that used to take a full day now takes under an hour. Sourcing also evens out, because every deal we look at gets the same thorough treatment regardless of how it came in. And we now build a structured dataset of every company we evaluate, which lets us study our own pattern recognition over time.
The quantitative layer does not replace qualitative analysis, but gives it something concrete to build on. We still generate a brief qualitative summary based on the results, which helps us understand where to dig deeper and what questions to bring into diligence first. But because that analysis is grounded in numbers we trust, the model is no longer improvising its way to a conclusion. It is interpreting data we produced, using rules we designed, and applying them the same way from one deal to the next.
Designing the judgment layer
Building this kind of framework is not mainly a technical exercise. The hard part is setting the rules and making the decisions behind them. You need to define the objective of your analysis, the inputs you will use, the areas you will evaluate, the metrics relevant to each area, how to handle qualitative factors that are hard to quantify, and how metrics translate into scores. None of these choices are neutral. Each one of them is a key component of what matters to your fund and the outcomes you are after.
This approach is only as good as the thinking behind it. The rules, benchmarks, and metrics we built into it reflect more than a decade of investment experience across the Global South, which means they capture how we think today, not necessarily how we will think tomorrow. A rubric that works well for the companies we have backed over the years may quietly underweight the twenty-something dropout from Bangalore rebuilding credit infrastructure for the unbanked with AI. That is not a flaw in the system. It is just a reminder to stay actively involved with it, review the outputs critically, update the rubric when reality teaches you something new, and remain open to the fact that the best opportunities sometimes arrive in unfamiliar shapes, with pieces still missing and questions still unanswered.
Describing, not declaring
We are also responsible for shaping how the LLM communicates its findings, because the wording determines how your team will act on what they see. What the AI produces can never be treated as a verdict. The model does not make decisions. It does not determine what is good or bad. It surfaces intel, flags patterns, and provides context, but it does not declare outcomes.
Instead of “Proceed with investment,” we aim for “this company is in the top quartile of our pipeline for projected growth”. Instead of “no relevant impact potential,” we want “projected beneficiaries reached fall below our benchmark for this sector”. The model describes, compares, and flags. It does not declare. This keeps the analysis useful without letting it become a directive.
Drawing the (decision) line
AI gives us speed, consistency, and structure. The call is still ours and will remain ours until we see the model making better decisions than we can. We don’t expect this soon given the complexity of the judgment calls required for early stage investing, but we don’t rule it out. The framework is the line between the two, and what we do on our side of it is the work that actually matters. AI will keep getting better, and we will keep handing it more of our work.
Eventually, the line between what the model contributed and what we contributed will only get blurrier, but the mechanisms we build now are what keep us at the wheel, even when the line becomes difficult to see. The more AI can help us be smarter and faster, the more time we can spend on uniquely human evaluation areas.
Francis Perelman is AI program manager at Capria Ventures.
Guest posts on ImpactAlpha represent the opinions of their authors and do not necessarily reflect the views of ImpactAlpha.
