Most Asset Managers Already 'Use AI.' Few Turn It Into Alpha

Let me be honest about something. Two of the better calls in my portfolio over the past year were not really my ideas. An AI screening model put them on my desk. Both companies had market caps above €10 billion. They were not hidden microcaps. But they were easy to miss, because the real investment story sat one layer below the obvious label.

For the past year, I have run a live equity portfolio with nine digital analysts. They do not drink coffee, they do not ask for a bonus, and they never tell me the idea was theirs. More useful than that, they widen the field, they challenge my thesis, and they force discipline. I still make the call. That is the whole difference between using AI and profiting from it.

This sounds strange for a fundamental portfolio manager. It is not. The final decision was mine. The work before the decision was not mine alone.

One more thing about me, so you know where I stand. I do not build AI models. I design the investment workflow around them. That is a very different job, and it is the job most firms are getting wrong.

Most asset managers already use AI and do not have an AI problem. They have a workflow problem. Here is one honest question to ask in your own firm. Does AI change the actual investment decision, or does it just make the old process look more modern?

In my experience, the difference comes down to four questions. Does AI:

• widen research breadth?
• shorten the time to a useful first view?
• challenge investment theses through a genuine red team analysis?
• define clear points where human judgment remains accountable?

If not, your firm may be using AI—but it is probably not yet profiting from it. That gap is what this article is about.

I recently spoke with the CIO of one of Europe's largest asset managers. His questions were sharp and specific: which metrics matter most for AI, which algorithms to use, what infrastructure to build. These are good questions. But they are the wrong starting point. Many firms jump straight into the technical layer before the process layer is clear. It was a reminder that even at the very top of the industry, the instinct is to reach for the technology before asking the simpler question: Where in my investment process would AI actually change a decision?

Mercer finds that 55% of asset managers now use AI somewhere in the investment process, yet only 8% report measurable return improvement. MIT finds similar results across industries, where roughly 95% of GenAI pilots generate no measurable profit. Meanwhile, PwC reports profit per AUM has fallen nearly 20% since 2018. Firms are investing heavily in AI, but few are capturing economic value.

First, most teams start with the tasks that are easiest to automate, which makes sense. The problem is that easy to automate and relevant to returns are often not the same thing. Summarizing an earnings call in two minutes instead of two hours is a real efficiency gain. But efficiency and alpha are different problems.

The second reason is that AI sits next to the investment process instead of inside it. Most firms have a process that works. I understand the reluctance to touch it. But using AI well does not mean rebuilding what you have. It means looking at each step and asking where can AI inform a decision that a human still owns? McKinsey found that workflow redesign was one of the strongest factors behind meaningful AI impact. They also found that only 39% of firms report any enterprise-level EBIT impact, and only around 6% qualify as AI high performers. These facts are related.

The third reason is harder to fix, because it is not technical: the people. That is what I also hear in many conversations with CIOs. Boston Consulting Group puts it at 70% people and process, 20% technology and data,10% algorithms. Portfolio managers do not trust the output, especially after the model is wrong once. Management buys the licenses, announces the strategy, and pushes it down. The team gets another tool, another login, and another password they will forget next Monday. What they often do not get is a clear place for that tool inside the actual process. The result is shelfware: expensive software that sits next to the work and never enters the decision.

So here is what I do instead.

My process runs like a research operating system. Five roles, each one handled by a different model or agent, and a human who decides at every step. That human is me.

It starts with regime analysis. Predictive models plus a structured LLM overlay tell me which sectors and factor exposures fit the current market. In early 2026, the answer pointed clearly at large cap financials and quality industrials. That alone narrowed a universe of roughly 3,000 global, liquid companies into a much smaller space.

Then the screening models go to work. They read up to 10,000 features per stock per day: technical, fundamental, and sentiment, weighted by the regime. Out comes a list of roughly 100 names.

Then, the real work starts: the fundamental deep dive. With AI support, I get to a useful first view in about 20 minutes. In my old life as a pure fundamental portfolio manager, the same work took two to three days per company. That is not a small gain. That is the gain.

Before any position is sized, a red-team model attacks the thesis on purpose. And once a stock is in the book, monitoring tells me when the facts start to drift away from the story I bought.

Here are three positions from the live portfolio which I entered more than six months ago. I will not name them, because the point is the process, not a hot tip.

The first is a mid-cap US company that, on paper, does electrical contracting. When the screening model surfaced it, my first reaction was a polite no. A Houston electrical contractor is not the kind of name that excites an investment committee. Then the 20-minute deep dive showed the real picture. Its communications segment builds the physical network infrastructure inside US data centers, the cabling and connectivity that hyperscalers need before a single server switches on. Every dollar of AI compute needs this kind of spend first. The company sits right in front of that wave, and almost nobody was looking. Backlog up more than 60% year-on-year. Sell-side coverage: one analyst, with a hold. The AI did not decide anything. It just made sure the name reached my desk.

The second is a Japanese industrial group. The first-level label was pumps. The second-level economics was semiconductor equipment. The company holds the number two global position in chemical mechanical planarization tools for chip fabrication. Every advanced chip needs this step. That business was growing fast, but the stock still did not read as an AI name to most investors. The system saw the second layer before the first layer stopped being the story.

The third one did not go so well, at least not at first, and that is exactly why I am including it. It is a European aerospace and defense company. It came in on strong fundamentals and a real tailwind: rising defense budgets and a strong position in military aircraft engines. Then the 2026 Strait of Hormuz crisis arrived. The catch is that most of this company’s money comes not from defense but from commercial aviation maintenance. An oil shock that grounds planes hits exactly that part. Short-term performance was ugly. Nobody predicted Hormuz. I certainly did not. The value was not prediction. It was discipline. The red-team layer had already mapped the revenue exposure, so when the shock came, the real question was simple: Is the thesis broken, or just slightly damaged? It was slightly damaged, not broken. Telling those two apart is the whole job, and a structured process makes it easier, not harder.

The first two names added real value over the same period the third one hurt. Since inception on June 30, 2025, to April 30, 2026, the live model portfolio has beaten the MSCI World Net Total Return Index by 14 percentage points, measured in euros, gross of fees and including dividends. This is not a backtest or a simulation. Every decision is documented, with a timestamp, under real-world conditions including trading fees. It is a model portfolio, managed daily with real market prices and real trading decisions, not a full institutional track record, and ten months are not twenty years. I will not pretend otherwise. But it is live, it is benchmarked, and it has felt real market pain. That makes it more honest than another AI pilot living in a slide deck.

Three things make the difference.

First, the cost side. I do not build models in-house. I license them on subscription, which turns a big, fixed cost into a flexible running cost. What does this mean? The fixed cost of research drops, and the output per professional goes up. Anthropic’s own research suggests AI saves around 80% of the time on typical financial analyst tasks. The profit is not in firing people. It is in how much more one person can now cover.

Second, many models beat one model. Early academic work on financial multi-agent systems in 2026 suggests that how you wire the models together can matter more for profit than which single model you pick. A workflow built on one frontier model is also fragile. Recent export-control restrictions on Anthropic models showed how fast access to a top model can disappear. Running several models is not only better analysis, it is insurance.

Third, and most important: the more AI you use, the more human judgement matters, not less. A field experiment at one of Germany’s largest savings banks found that clients trusted human-plus-AI advice clearly more than pure AI advice, especially on the bigger decisions. The same is true for the investment side, no investor wants a black box. AI widens the set of ideas and sharpens the analysis. The portfolio manager still makes the call. That is not a weakness of the setup. It is the point of it.

The tools are not the problem. The workflow around them is. I nearly scrolled past both of the first two companies I mentioned. That is the thought that stays with me. The point is not only what the system found. It is what I would have missed without it.