Every major asset manager now claims AI capability. Most of those claims describe something real but narrow — better tools, faster processes, improved analytics. The question for allocators is what AI in the investment process actually produces in terms of returns, and how to evaluate it honestly.
The Spectrum of AI Deployment
The financial industry’s use of AI in asset management spans a wide range, and most of it is not what the marketing language implies.
Process AI — the majority of deployments — covers portfolio screening, research organisation, reporting automation, and compliance monitoring. These are genuine improvements in operational efficiency. They are not alpha generation.
Signal-support AI covers applications where machine learning tools process inputs that human portfolio managers then act on. Sentiment scores, NLP-derived signals, alternative data feeds. The human still decides. AI narrows the research aperture and improves the quality of inputs. Evidence suggests modest performance improvement here.
Decision-architecture AI — the rarest category — describes systems where AI is not supporting a human decision but replacing the decision layer entirely. Signal discovery, testing, validation, portfolio construction, and execution all run through systematic frameworks without discretionary override. This is where the evidence on outperformance is most compelling, and where the structural logic is hardest to argue with.
Why the Human Layer Is the Problem
The conventional understanding is that AI in asset management augments human intelligence. But the evidence on active management performance suggests the opposite framing: the human decision layer is precisely where returns are destroyed.
89.5% of active equity managers underperform their benchmark over 15 years (SPIVA Year-End 2024). This has been documented consistently across decades, geographies, and market regimes. It is not a talent deficit — the same data shows that buying decisions often demonstrate skill. The destruction happens elsewhere.
Akbas et al. (Journal of Finance, 2023) studied 783 institutional portfolios and found that selling decisions — not buying — lost 50 basis points per month relative to random selection. Experience made no difference. The failure was stable across seniority levels. Human judgment applied to portfolio management produces systematically negative value in aggregate.
AI tools that improve the inputs to that judgment improve a process that still ends in a structurally disadvantaged step. The higher-order intervention is removing that step.
What Genuine Alpha Generation Requires
Generating alpha with AI — rather than merely using AI more efficiently — requires a specific architecture.
Scale of hypothesis evaluation. A human research team can seriously evaluate a few dozen signals per year. A systematic research engine can evaluate orders of magnitude more, test them rigorously, and reject the vast majority that do not survive. The valuable signals emerge from the volume and rigour of the process, not from a senior analyst’s conviction.
Point-in-time data discipline. Most backtests are contaminated by hindsight. Financial statement data revised after the fact, index constituents that do not reflect historical composition, corporate events that were unknowable at the time. Genuine AI-driven alpha requires the same discipline: signals must be validated on data that was actually available at the decision point.
Out-of-sample validation before deployment. The most common failure mode in quantitative investing is in-sample overfitting: optimising a model on historical data until it fits perfectly, then discovering it has no predictive power on new data. A rigorous research process maintains strict separation between the data used to develop a signal and the data used to validate it.
Deterministic execution. The final failure mode is at the portfolio level: a validated signal, discretionarily interpreted by a portfolio manager who decides it does not apply this quarter. Structural AI in asset management encodes the execution rules before the fact and does not allow retrospective override.
The Meaningful Distinction
For allocators evaluating AI claims in the asset management industry, the operative question is not “does this firm use AI?” but “where in the decision process does AI end and human judgment begin?”
Firms where the answer is “AI supports research, humans decide” are describing process improvement. Firms where the answer is “the system decides, humans govern the risk framework” are describing a genuinely different investment architecture — with the structural advantages that implies.
AvB Capital is an AI-native systematic manager based in London. Track record and strategy documentation are available to qualified investors on request.
Frequently Asked Questions
How is AI currently used in asset management?
Applications range across the full spectrum: portfolio screening and idea generation, sentiment analysis of earnings calls, risk factor modelling, order execution optimisation, client reporting automation, and regulatory compliance monitoring. The majority of applications improve process efficiency. Fewer are integrated directly into the alpha-generating research loop.
Does AI actually improve investment returns in asset management?
The evidence is mixed, and highly sensitive to how AI is deployed. Using AI to accelerate research on signals that humans then judge discretionarily produces modest improvements. Using AI to discover, test, and validate signals inside a systematic framework with deterministic execution produces more durable results — because it removes the human judgment layer where most underperformance originates.
What is the difference between AI asset management and quantitative investing?
Traditional quantitative investing uses statistical models on structured data — price, volume, financial ratios. AI-native asset management extends this to unstructured data and uses machine learning to identify non-linear relationships that rule-based quant models cannot express. The deeper difference is in the research process: AI systems can evaluate a far larger hypothesis space than human analysts.
Are AI asset managers regulated the same as traditional managers?
Yes. AI-native managers operating in major jurisdictions are regulated identically to traditional active managers — FCA authorisation in the UK, SEC registration in the US. The regulatory framework applies to the entity and its activities, not the methodology.
What should allocators look for when evaluating AI in asset management?
Focus on three things: the research process (how are signals discovered and validated?), the decision architecture (where exactly does AI end and human judgment begin?), and the risk controls (are limits structural constraints or policy documents?). Firms that are vague on any of these three are likely using AI for marketing rather than for alpha.