Could AI Run Portfolios Someday? Maybe, but Not Now

AI is everywhere in asset management conversations right now. But when we looked across roughly 3,200 rated strategies, real adoption is still fairly limited. Our research shows that genuine AI integration is concentrated in fewer than 20 firms, most of them with long-established quantitative or systematic capabilities. 

So while the buzz is loud, the reality is more nuanced. Here’s what's actually happening, where AI is making a difference, and where it’s not (yet). 

Artificial intelligence has been part of asset management for decades. Before 2022, however, most of what managers called "AI" was traditional machine learning: statistical models trained to perform specific, predefined tasks. 

Systematic and quantitative managers have built genuine edges in this era, including, for example, better signal extraction from earnings calls, faster processing of satellite imagery, and more disciplined factor construction at scale. But the barrier to entry was significant: Firms needed data scientists, proprietary infrastructure and datasets, and years of internal development. For fundamental managers and for smaller firms without the scale to absorb that investment, AI remained largely out of reach. 

What makes the latest gen AI a genuine departure from every prior generation of investment technology is that models can now act. 

This narrowness stands in contrast to industry claims. According to the European Securities and Markets Authority's 2025 analysis, only approximately 145 of 44,000 UCITS funds formally disclose AI use in their prospectuses and regulatory filings, yet our Parent review due diligence on over 300 asset management firms suggests nearly every manager claims to employ AI in some form.  

This gap exists for three reasons:  

• There’s no clear regulatory definition of what counts as AI use.
• Managers have incentives to market themselves as AI-enabled.
• AI washing remains common (basic automation dressed up as AI).

Adoption is heavily skewed toward equity, which accounts for the majority of AI-integrated strategies in our sample. This is unsurprising. Equity markets offer the deepest and most structured data ecosystem, spanning price and volume data, company fundamentals, regulatory filings, earnings call transcripts, news sentiment, and a growing array of alternative datasets. Machine-learning techniques thrive in environments where data is abundant and varied, and the breadth of investable equity universes, often spanning hundreds or thousands of securities, creates a natural fit for systematic pattern recognition and stock selection. 

Fixed income remains comparatively underrepresented. This is consistent with the structural challenges of applying machine learning in bond markets: over-the-counter trading, less transparent pricing, lower liquidity, and fewer standardized data inputs relative to equities. 

Allocation strategies appear least frequently, likely reflecting the narrower scope for machine-learning application in asset-allocation decisions. Multi-asset funds typically make a smaller number of higher-level allocation choices across asset classes, geographies, or risk factors, offering fewer degrees of freedom for machine-learning models compared with the thousands of individual securities available in an equity universe. 

Taken together, the asset class distribution underscores that AI and machine-learning adoption in asset management remains concentrated where data is richest, investment universes are largest, and quantitative traditions are most established. While early adoption in fixed income and multi-asset strategies is evident, meaningful expansion into these areas will likely require improvements in data infrastructure and market transparency. 

In the systematic world, if a signal worked historically, you could show it with backtesting. Feed the same inputs into the same model, run it over a defined historical window, and measure the output. The results can be stress-tested and challenged. Gen AI breaks this chain at two distinct points: 

1. Nondeterminism: A large language model asked to assess the investment case for Apple today will not produce the same conclusion word-for-word if asked again tomorrow (like people would). The consequence is straightforward: You cannot construct a reliable historical track record for a gen AI-augmented investment process. 
2. Look-ahead bias: Most large language models are pretrained on internet-scale data up to a certain cutoff date—that may include financial commentary, analyst reports, news coverage, and corporate histories from the very periods a researcher might want to backtest. The model may implicitly encode information that was not publicly available at the time being simulated, which can distort results.

AI's productivity gains come with a hidden cost: Currently, the tasks AI is absorbing are roughly the same tasks through which junior investment professionals have historically learned to think. The system is increasing the demand for high-level verification while simultaneously dismantling the training ground for future verifiers. 

In investment management, AI significantly reduces the time it takes to produce research outputs: summaries, earnings analyses, thematic screenings, company profiles, academic paper testing, and so on. But because asset managers lack hindsight into how these models operate and their rapid evolution, verification is still unavoidable. They also critically need to explain how results are achieved, and that requires domain expertise. The same senior analysts whose time AI is supposed to free up are also the ones best positioned to catch its errors. 

The problem is compounded by the nature of the errors themselves. Unlike a formula error in a spreadsheet, which is typically systematic and therefore detectable once found, gen AI errors tend to be sporadic and plausible. Detecting these errors requires contextual knowledge, which isn't easily automated. 

There is another downstream effect. As AI handles increasingly complex tasks autonomously and firms scale by “better systems, not bigger teams,” this means fewer hires at the bottom, but those roles may be needed to sustain future real expertise. The risk is losing the apprenticeship pipeline that produces future experts and maintainers of judgment. 

While AI adoption by investment firms is growing and use cases are varied, there are no clear indications that regulators will introduce specific rules to govern its use. To date, no major regulator has set AI-specific portfolio management rules for mutual funds. Instead, most are extending existing conduct, governance, outsourcing, risk management, disclosure, and fiduciary frameworks to cover AI. 

Ultimately, fiduciary duty is likely to play a dominant role in how widely AI is adopted by the industry. Clients cannot be told “the AI picked this investment,” so asset managers need tools that show which variables drive decisions and what specific model inputs have been used, making results genuinely explainable to stakeholders. This means that there will likely always remain a human element across various steps in the investment decision-making process. 

The most credible near-term vision is not AI replacing analysts like-for-like. It’s the use of AI so that analysts may be able to amplify their research significantly across a much larger universe: more companies, trends, signals, and types of instruments. 

There is a hard ceiling on how many companies a single analyst can genuinely understand at any one moment. A rigorous fundamental analyst can realistically cover 20 to 30 security issuers with real depth. The rest of the investable universe is, by definition, less scrutinized. AI potentially removes that constraint. 

As AI tools become widely available and embedded in investment processes, the competitive edge they may provide will likely erode. But the data infrastructure beneath them is not commoditizing at the same pace—and that is where the durable edge is likely to sit. Quality data access is getting more expensive yet is critical for maintaining an edge. 

For most of the industry, AI may become a cost for remaining competitive rather than a source of durable outperformance. The segment most exposed is what might be called the undifferentiated active middle: firms that are neither systematic houses with deep quantitative infrastructure nor high-conviction specialists capable of directing AI to extend a genuinely distinctive process. For that tier, AI may accelerate the competitive pressure that fee compression and passive growth have already been applying for more than a decade. 

Could AI eventually run portfolios end-to-end? Technically, it’s not out of reach. Based on our research and interactions with asset managers, the components are either already in production or in active development. 

Can AI fully replace the kind of judgment that generates alpha, and can we prove it? Right now, the answer is "no.” But AI might eventually eat away at the judgment call. It could prove better at challenging consensus because it can test more hypotheses, simulate more counterfactuals, and process more contradictory evidence in parallel than any individual or team can manage unaided. 

Today, AI output still requires heavy expert review in investment management, but that doesn’t mean review must remain predominantly human forever. Over time, it’s also possible that AI models may be able to self-improve. In that world, the bottleneck does not disappear, but it may become more automated and less dependent on scarce human attention than it is today. If this is where we're headed, then the durable edge in active investment may not remain discretionary brilliance. It may shift toward better system design: proprietary data, model orchestration, feedback loops, and firms that learn faster than rivals how to turn general-purpose intelligence into repeatable investment processes. 

What would be the role of investment professionals in such a world? It's likely that they would cease to be the primary producers of research, the key architects of portfolio construction, or the designers of quantitative models, but instead become orchestrators. Their role would increasingly center on setting the direction, arbitrating between conflicting agent outputs, and exercising judgment when the system encounters ambiguity or something it was not designed for. 

Because trust will remain the last frontier for a long time, it is likely that the portfolio manager's role will therefore not simply migrate toward the orchestration of AI systems but also an even greater focus toward client-facing activities. Being the face that gives AI legitimacy with the people whose capital is at stake could become a critical part of a fund manager's role going forward.