The questions we used to ask and the ones that matter now in AI investing

Most of us carry our investing habits forward longer than we should. 

We reuse questions that worked well in previous cycles, even when the underlying dynamics have shifted. In AI, that habit is particularly costly. Many of the questions investors instinctively reach for now obscure more than they reveal. 

This is not because those questions were wrong. It is because the environment they were designed for no longer exists. 

Below are some of the questions we used to ask as investors, and the questions I now find far more useful when assessing early-stage AI companies.

This question still matters. But in AI, it requires more discipline. 

I want to understand how the world works today without the product. What is manual, fragmented, slow, or expensive? How much human time is being wasted? Where does risk sit? What outcomes suffer as a result? 

If I cannot clearly describe the before state, I cannot assess whether the solution is meaningful. AI makes it easy to improve things marginally. I am interested in whether it improves something that genuinely matters to the people paying for it. 

When a startup pitch ends, the first question I ask myself is whether this meaningfully moves the needle for the users it is built for.

A quick poll before we jump in:

In most cases, the answer comes down to workflow. This is one of the most important questions in AI diligence. 

Durability in AI often comes from workflow ownership. Products that embed themselves into daily operations are far harder to remove than products that remain optional, regardless of how impressive the technology appears. 

This is a question I now ask in every AI diligence. 

Early traction alone tells me very little. What I want to understand is whether usage makes the product meaningfully better over time.

If nothing compounds with usage, durability is fragile. 

Product market fit used to be treated as a milestone. In AI, it is a moving target. 

Customer expectations shift quickly as capabilities evolve. What feels differentiated today can become baseline tomorrow. As a result, I am less interested in whether a company has hit product market fit and more interested in whether it can maintain relevance as the environment changes. 

The real question is whether something becomes harder for competitors as the company grows, or whether advantage resets every cycle.  

Speed is often mistaken for execution. 

The strongest AI teams do move fast, but what matters is not how much they ship. It is how quickly they turn real usage into better decisions. 

Learning velocity compounds. Motion does not. 

This is visible in companies like Lovable and ElevenLabs, two of the fastest-growing AI companies to emerge from Europe. Both companies were valued at $6.6 billion most recently and are a few years old.  

Lovable allows users to design, build, and iterate on software products using natural language, dramatically compressing the time from idea to shipped product. Launched publicly in November 2024, the company ships constantly, often daily, with engineers announcing frequent updates and customers describing the product as always improving. 

ElevenLabs followed a similar pattern. Launched in December 2023, it scaled far faster than traditional SaaS benchmarks, becoming the default infrastructure layer for high-quality AI voice across media, gaming, publishing, and enterprise. The company reached over $100m in ARR in under two years, before traditional moats were fully formed. I was fortunate to have Eleven Labs in my portfolio via an investment I made into another fund.  

In both cases, speed was not cosmetic. It matters because learning has kept pace with execution, and each release has sharpened the product in response to real usage. 

Every AI startup today has access to the same core models. 

For that reason, I spend very little time underwriting technical novelty at the application layer. Instead, I focus on where defensibility might emerge over time. 

Thin wrappers are easy to build. Compounding systems are not. 

Sanius Health is a useful illustration of how these questions change the analysis. We invested in Sanius Health back in April 2025 and again in August 2025.  

Sanius focuses on rare and chronic diseases, where promising therapies are often delayed not by science but by a lack of access to patients and high-quality, deeply scientific, longitudinal, real-world data. Small, dispersed populations and fragmented datasets make it difficult for pharma teams to recruit patients, generate evidence, and understand disease progression outside controlled trials. 

Sanius addresses this by building a continuous, longitudinal view of patients in the real world. Rather than relying on occasional hospital visits, the platform integrates medical records, patient-reported outcomes, remote monitoring, diagnostics, and genomics into a single dataset that reflects how patients actually live and respond to treatment. 

Because this data is generated as a byproduct of ongoing care, it materially shortens time to value for pharma. Trial recruitment improves, evidence gaps close faster, and development timelines compress. In practice, this can accelerate important therapies by many months. The company is also converting the power of its data platform through the use of AI to predict disease deterioration and rare disease patient find to reduce the time to diagnosis from 10 years to 3 years, a real value for patients, pharma and healthcare providers 

The insight comes from asking where the product sits in the workflow. Sanius is embedded into how rare disease data is created and accessed, not layered on top of it. Over time, ownership of patient relationships and longitudinal data compounds into an advantage that is difficult to replicate. This makes it increasingly valuable over time.  

Distribution is often treated as a go-to-market problem. In AI, it is often the moat. 

I want to understand how users discover the product today, and why that channel works for this company in particular. More importantly, I want to understand what happens as the company scales. 

Does distribution get cheaper or stronger over time? Is there trust, access, or integration that competitors would struggle to replicate? 

In many AI businesses, distribution advantage explains outcomes better than product quality alone. 

Finally, I pay close attention to how founders reason when answers are not yet clear. 

AI markets move too fast for static plans. I am less interested in perfect certainty and more interested in how teams think out loud, how they handle trade-offs, and how quickly they adapt when assumptions break. 

These signals appear early, long before they show up in metrics. 

AI has not removed the need for judgment. It has raised the bar for it. 

The questions that matter today are not more technical. They are more grounded, more behavioural, and more forward-looking. 

That is where I now spend my time. 

I’ll be teaching a full-day, in-person Angel Investing Course at Regent’s University, London, on February 27th, 2026. I’d love to extend a £50 discount to our signal subscribers who’d like to join us. Just use ARAYASIGNAL50 when purchasing.

The programme will focus on portfolio design, deal analysis, risk management, and live case studies covering the mechanics of investing well at the early stage, along with a new module on investing in the age of AI.

Warmly,

Rupa Popat

with Team Arāya