We start with the business the technology enables.
A product demonstration shows a capability. An investment thesis needs to explain why a customer will adopt it, why the company can retain that customer and how the economics improve as it grows. Our interest is in AI embedded in work that matters: processing a claim, maintaining an asset, supporting a care team or operating a financial workflow.
These are the six questions we use to frame that discussion. They apply differently to a research-stage team and a company with paying customers; the evidence should fit the stage.
01. Who owns the problem and the budget?
The user, the operational owner and the purchasing decision-maker may be different people. A credible team understands all three. We want to see the task’s frequency, the cost of the current process and the consequences of getting it wrong.
Customer interviews help define the need. Repeated use, a paid pilot or a renewal can provide stronger evidence of demand when a product is ready for it. A large addressable market does not explain why the first customer will buy.
02. What remains valuable when models improve?
Access to a capable model is a starting point. We examine what the company adds: specialist knowledge, legitimate data access, a difficult integration, a reliable evaluation system or a product that fits a particular job exceptionally well.
Data is useful only when the team can use it for the intended purpose. We ask who owns it, what rights apply and whether access can persist as the business expands. A platform dependency should also be visible: what happens to cost, performance or delivery if a model provider changes its service?
03. How is performance measured in the actual workflow?
A general benchmark cannot answer every customer question. An invoice-processing system, for example, needs to handle the documents the customer receives, including exceptions, incomplete records and unfamiliar formats.
We look for evaluation cases drawn from the intended task, explicit error categories and a process for checking regressions. Accuracy should be read alongside latency, failure handling and the human effort required to accept or correct an output.
04. What does an accepted result cost?
Inference spend is only part of the equation. Human review, retries, data preparation, monitoring and customer support can change the margin materially. The useful unit may be a resolved case or an accepted document, rather than an API call.
Cost per accepted result = the total cost of running and reviewing the workflow ÷ the number of results the customer can use.
We examine which costs fall with volume and which grow with each customer. A team should be able to explain the assumptions behind its pricing and where manual delivery still sits inside the product.
05. Can the second deployment be easier than the first?
Early implementation work can teach a team what to productise. The question is whether those lessons become reusable connectors, configuration and support processes, or whether every installation remains a new custom project.
We ask about time to deployment, access controls, integration dependencies and who maintains the system. For a regional rollout, language, hosting and procurement requirements belong in that plan from the start.
06. What will the next milestone prove?
A useful capital plan connects spending to a change in evidence. That might be validating a technical assumption, demonstrating reliable performance in one workflow or showing that customers renew without intensive founder support.
We want the team to identify the hardest unresolved assumption, the work needed to test it and what would change its approach. That makes a milestone useful to both the company and its potential investors.
Our conviction grows when customer need, technical advantage and delivery economics reinforce one another.
This lens runs across our four themes: AI and digital infrastructure, health and life sciences, climate and energy, and FinTech and digital assets. In each, we look for the evidence that connects a promising technology to an enduring business.