Five Things I Learned About AI Implementation from Practitioners at Man Group, H2O, and Beyond

Five lessons on AI implementation from practitioners at Man Group, H2O, and beyond.

Neudata London Data Summit, 26 March 2026

I recently wrote about the Quantamental 2.0 thesis, the idea that generative AI is collapsing the divide between quantitative and fundamental investing. That piece was about the theory. This one is about what happens when the theory meets reality.

At the Neudata London Data Summit, I joined a panel with Timothée Consigny (CTO, H2O Asset Management), Matthew Bell (Senior Data Scientist, Man Group), and Dr Ana Armstrong (Founder, AIM Cube), moderated by Takaya Sekine. Each of us came from a very different corner of the industry. What surprised me was how quickly we converged on the same set of practical tensions that anyone implementing AI in investment workflows is actually dealing with.

Here are five lessons from that conversation.

1. If Your AI Gives Recommendations, Your PMs Will Stop Thinking

This was the sharpest insight from the panel, and it came from Timothée.

H2O built an AI tool that reviews trades before execution. The first version gave the PM a recommendation: proceed or don't. The result? PMs got lazy. They scrolled straight to the conclusion without reading the analysis. The tool was making them worse decision-makers, not better.

So they redesigned it. The new version lists only the pros and cons of a trade, with no recommendation. The PM has to decide.

This is a critical design lesson that applies far beyond H2O. If your AI workflow produces a signal or a recommendation, you are creating a shortcut that experienced professionals will take. The value of discretionary investing comes from the quality of human judgment. If you design AI that bypasses that judgment, you have undermined the very thing you are paying for.

The takeaway: design AI to challenge your team, not to think for them.

2. AI Changes the Economics of Niche Signals

Matthew Bell described a use case at Man Group that perfectly illustrates where AI creates genuinely new capability rather than just speeding up existing processes.

Someone on his team had an idea: could they measure copper production by analysing news about mine outages in Chile? Interesting hypothesis. But no data scientist or quant researcher would have built a dedicated NLP pipeline for a signal that trades one commodity in one country. The effort-to-impact ratio was too low.

With AI, the effort dropped to nearly zero. An analyst can now generate that signal on the fly, test whether it works, and move on to the next idea. Man Group's plan is to build a database of hundreds of these niche signals, run them through standard checks, and hand the pre-processed results to researchers who can validate them in hours instead of weeks.

This changes the economics of research entirely. It is not about doing the same work faster. It is about pursuing ideas that were previously not worth the effort. The long tail of investment research is opening up.

3. The Biggest Capital Allocators Are Still on the Sidelines

Ana Armstrong brought a perspective that tempered any enthusiasm about rapid adoption. Her experience managing a government pension fund revealed an uncomfortable truth: some of the largest pools of capital in the world are barely engaging with AI.

The reasons are not technological. They are institutional. Governance requirements, trust deficits, and the sheer inertia of organisations that have operated the same way for decades. These are not problems that a better model solves.

This creates an interesting dynamic. While the pension fund industry moves slowly, retail investors are embracing AI tools with enthusiasm (and often without the risk controls that institutional investors would insist on). The gap between institutional caution and retail speed is striking, and it creates both risk and opportunity.

For firms like the ones I work with, boutique hedge funds and family offices, the implication is clear. The AI adoption gap is not between firms that have the technology and firms that don't. It is between firms that have the organisational culture to adopt and those that don't.

4. AI Replaces the Team, Not the Manager

This was my contribution to the discussion, and it reframes a question I think the industry gets wrong.

The standard debate is: will AI replace fund managers? The answer from every panelist was some version of "no, it augments them." Fair enough. But that misses the more important shift.

A firm like Man Group has layers of analysts, data scientists, and risk professionals around each portfolio manager. A boutique hedge fund does not. For those firms, the question is not whether AI replaces the PM. It is whether AI can replace the team the PM never had.

Can an AI agent provide 24/7 risk monitoring that a lean firm could never staff? Can it produce the kind of factor exposure analysis that requires a dedicated quant risk team at a large firm? Can it synthesise research notes, cross-reference them against portfolio data, and surface inconsistencies?

The answer to all three is increasingly yes. And that is not about replacing human judgment. It is about giving smaller firms the support structure that larger firms take for granted.

5. Domain Knowledge Is the Moat, Not the Model

This was the thread that connected everything. Timothée talked about how the curated input from experienced PMs produces better AI output than mass data ingestion. Matthew emphasised that Man Group's discretionary PMs bring industry expertise that AI cannot replicate. Ana pointed out that frontier markets and biomedical sectors still require fundamental human judgment.

The convergent conclusion: as AI models become commoditised (and they will), the sustainable competitive advantage is not technological. It is contextual.

Your edge is the quality of your internal knowledge. How your PMs think. How your firm makes decisions. The institutional memory that your team has built over years. The firms that win will be the ones that organise this context systematically and feed it into their AI systems.

I call this "context over code." The model is infrastructure. The context is alpha.

What This Means for Boutique Firms

If you run a lean investment team and you are thinking about AI adoption, these five lessons distil into a practical framework:

Design for judgment, not automation. Build AI workflows that enhance your team's thinking, not replace it. Timothée's pros-and-cons approach is a model worth studying.

Start with the long tail. The highest-value AI use cases are often the ones that were not economically viable before. Don't just automate existing workflows. Ask: what questions could we never afford to investigate?

Adoption is cultural, not technical. The technology is available. The barrier is organisational. If your team doesn't trust the outputs or doesn't know how to ask the right questions, no amount of model sophistication will help.

AI is your team, not your replacement. For boutique firms, the right mental model is not "AI as a tool" but "AI as the analyst, risk manager, and data scientist you couldn't hire."

Invest in context, not models. Curate your data. Document your decision-making process. Build institutional memory. That is your moat.