Over the last decade I’ve deployed capital across 30 businesses in 19 countries, from fintech in Europe to electric mobility in the UAE, specialty coffee in Dubai and real estate in Central Europe. In that time, I’ve used AI tools more than most of my peers. I run them against deal flow, I stress-test financial models with them, and I rely on them to flag patterns I would otherwise miss.
I’ve also watched AI tools consistently fail at the one thing investors actually get paid to do, which is deciding which founders are worth backing.
This is a practical argument grounded in four structural gaps I see every time I compare what the algorithms recommend with what actually plays out in the market.
1. The survivor bias problem
AI investment screening tools are trained on outcomes that already exist. They learn from founders who succeeded, founders who failed and the patterns in between. What they cannot see is the entire population of founders who would have succeeded but quit six months before their breakthrough.
The training data is survivor-biased by definition. Every persistence signal the model learns is calibrated on people who made it to the end, which tells you almost nothing about how to evaluate someone still in the middle of the arc.
In practice, this means AI systematically underweights resilience and overweights early traction. I’ve seen screening tools flag founders as low-probability precisely because they were going through the hard phase that every eventual winner goes through. An investor who understands the pattern backs them anyway. An algorithm optimising on historical data passes.
The real question AI cannot yet answer is the only one that matters early in a deal, which is whether this person has the psychological reserves to keep going when the numbers look terrible. Character doesn’t have an API.
2. The unstructured data AI can’t see
Most of my deals are cross-border. In 19 countries, I’ve learned that the decisive moments in an investment rarely happen in the data room. They happen at dinner, in a hallway conversation, in the way a founder responds to a hard question about a co-founder conflict they thought I didn’t know about.
AI can analyse the legal cleanliness of a deal. What it cannot do is read the micro-signals that tell you whether a founder is lying, hiding something, or genuinely doesn’t understand their own business. Those signals exist in tone, in eye contact, in how someone handles a moment of pressure.
This is a data availability problem more than a capability problem. The most predictive information in cross-border investing gets generated in rooms no algorithm will ever have access to. That’s why quantitative screening underperforms in emerging markets where trust networks do more work than due diligence reports.
You cannot digitise a handshake. In my experience, the handshake is often where the deal is actually decided.
3. Why optimisation models miss the biggest returns
AI is trained to minimise risk and optimise around existing patterns. That’s exactly what you want for a screening tool operating at scale. It’s also the reason AI alone would never have funded the investments that defined the last twenty years.
Every breakthrough deal I’ve been part of looked irrational on paper at the moment of decision. The highest-return investments in history are the ones that the pattern-matchers passed on. If the industry relied solely on AI screening, venture capital would fund incremental improvements forever and miss the disruptive bets entirely.
This doesn’t mean AI is useless here. It means the division of labour has to be honest. AI is good at pattern recognition; humans are good at pattern-breaking. The moment an investor outsources the anti-logic bet to an algorithm, they’ve stopped being an investor and started being a spreadsheet.
4. The intangible I still look for myself
Across 30 businesses, one factor has been more predictive of success than any financial metric. That factor is whether the founder can transmit conviction. Can they make a twenty-year-old engineer work sixteen-hour days? Can they make a customer believe in a product that doesn’t fully exist yet? Can they walk into a room of skeptics and leave with believers?
I call it fire. Some people have it. Most don’t. The ones who do can carry a business through the years when the data says they should quit.
AI can write a perfect pitch script. What it cannot do is transfer a state of conviction. No amount of natural language processing will change the fact that fundraising, hiring and customer acquisition are all fundamentally about emotional transmission from one human being to another.
This is the single hardest factor to measure, which is why founders who have it are systematically undervalued by screening tools. Until AI can capture behavioural signals at this level, and the data to train it doesn’t yet exist, identifying fire will remain a human judgment.
How I actually use AI
None of this means I distrust AI. I use it every day. It works as a powerful telescope, helping me see details across markets I couldn’t cover alone, surface deals I’d otherwise miss and stress-test assumptions at a speed no human can match.
When I was scaling EcoWay, our UAE business supplying high-power electric bikes and battery-swap infrastructure to last-mile delivery operators, I ran AI models against fuel-price scenarios, delivery unit economics and fleet utilisation across eighteen months of forward projections. The models were accurate. They showed me what the risks looked like based on historical data. What they couldn’t tell me was whether to commit capital into a market where electric two-wheeler penetration was still below 5%, or whether the delivery operators themselves had enough motivation to actually make the switch from combustion to electric. Those calls I made myself.
The decision to land on a particular planet, to put capital behind a specific founder, I make myself. Investing is a contest of will, intuition and the ability to back a human being when every graph points down.
AI makes me a better investor. It will never be one.
