Alexandra Mihailescu Cichon is global head of market development at RepRisk
AI is not just a tool for processing information faster. For asset owners and asset managers, it is increasingly a direct input into investment screening, risk assessment, stewardship priorities and portfolio oversight – with costly consequences when it fails.
Two recent studies point to the same concern from different angles. McKinsey’s latest state of AI survey identifies inaccuracy as the top AI-related risk organisations have already experienced.
In research conducted by RepRisk and Oxford Economics this year, more than 500 C-suite risk leaders at financial institutions identified AI-related conduct risks – including biased outputs, model misuse and insufficient auditability – as among the most material risks they expect to face within three years.
What is striking is not that these leaders expect AI risk to grow, but where they locate the problem. Their concern is less about whether AI can generate outputs, and more about whether those outputs can be governed, explained and defended once embedded in real business decisions.
For investment committees, risk teams and GPs, AI‑generated signals must do more than arrive fast – they must be reliable, explainable and auditable. This is not primarily a technology issue. It is a governance and accountability issue that investors cannot outsource or ignore.
Much of the public conversation around AI focuses on engineering breakthroughs: larger models, better performance benchmarks and impressive demonstrations. But investors need a different lens – one grounded in how AI is actually used inside organisations, and what happens when it informs a decision that needs to be defended.
In practice, effective and trustworthy AI systems share a small number of core building blocks, forming a practical framework for evaluating any AI-enabled provider.
From raw data to defensible decisions
AI models are only as good as the data they ingest. Investors should understand whether a provider relies on a curated and controlled source universe, or on broad, largely unfiltered data collection – effectively open web scraping.
Coverage breadth matters, but so does source quality. While web scraping can deliver results quickly, each run pulls from a different and shifting set of sources, meaning outputs change over time and cannot be reliably reproduced or defended.
Equally important is traceability. Can every insight be linked back to an original source? Can the output be reproduced? If a system cannot show where information originated, it introduces “black box” risk – making both internal review and external scrutiny difficult.
Finally, lawful access matters. As regulatory and contractual expectations tighten, providers must be able to demonstrate that their data sources are accessed and used in a compliant manner, including respect for paywalls, copyright, content restrictions and fair-use requirements.
Consistency, comparability and entity matching
Raw information does not become decision ready on its own. It must be evaluated using a clear, consistent methodology that defines what constitutes a signal, how it is classified, and how edge cases are handled.
“As AI becomes embedded in investment processes, trust cannot be retrofitted”
For investors, consistency over time is critical. Outputs that change unpredictably with shifting models or criteria undermine comparability and erode trust. A consistent – and ideally publicly accessible – methodology enables historical benchmarks, back-testing and robust time-series analysis.
Consistency is how unstructured information is transformed into stable, comparable, decision-ready and audit-ready data, rather than one-off model outputs. It’s also important for investors to ask about historical data: is there an intact historical record, or are data points reverse engineered (perhaps after sources change or disappear)?
Crucially, investors should ask whether they can see the underlying evidence behind an output. If results cannot be explained in plain terms, they cannot be responsibly relied upon.
Another point of failure is accurate and scalable entity matching. Often overlooked, it is where many AI systems fail in practice. Data only becomes actionable when it connects accurately and reliably to the relevant investment universe – at scale and across complex ownership structures, funds, securities and counterparties.
Misattributing river pollution to the wrong company isn’t just an error – it’s a liability. In the same way, investors cannot afford AI that flags minor labour issues at a hotel while turning a blind eye to its money laundering.
Why trust in AI only scales when humans lead
The building blocks above are only the starting point for an accurate and responsible AI solution. What makes it trustworthy at scale is the disciplined foundation beneath it – anchored in three pillars.
The first is ground truth, which refers to human-verified examples of reality against which models are trained and tested. Whether the domain is environmental incidents, cyber events or performance signals, models must learn from data that has been confirmed by human expertise, not inferred after the fact or generated synthetically at scale. Without ground truth, accuracy claims are largely theoretical.
Second is model specialisation. Best practice in AI system design increasingly favours multiple task-specific models rather than a single, monolithic system. Different tasks require different optimisations, with human validation playing a critical role at decision points. Specialisation also reduces hallucinations and increases auditability compared with one black-box model.
Third is domain expertise. AI does not operate in a vacuum. Providers need a deep understanding of the domain they serve, otherwise they struggle to identify meaningful signals, interpret nuance and manage edge cases – precisely where risk accumulates.
Five questions every investor should ask
For investors assessing AI-enabled data or analytics providers, the following questions offer a practical starting point. They separate enterprise-grade systems from impressive demonstrations:
- Where does your training data come from – and what proportion is verified ground truth versus synthetic or generated data? If you do not know what a model learned from, you do not know its blind spots.
- How are you sourcing your data – what is your source universe? Is it curated and controlled, or effectively open web scraping? Can lawful access be demonstrated?
- Can you explain how every data point is generated and assessed? Is the methodology consistent over time, and can the original evidence be shown every single time?
- Can this data be reliably connected to our investment universe at scale – across complex structures, funds and thousands of entities?
- Can outputs be audited, traced, and reproduced? If a result cannot be reproduced, the decision cannot be defended.
These are not technical questions. They are fiduciary questions.
As AI becomes embedded in investment processes, trust cannot be retrofitted. It must be designed in from the outset through data discipline, methodological consistency and governance-ready outputs.
For asset managers and asset owners alike, the goal is not to slow AI adoption but to ensure AI strengthens decision-making rather than obscuring it.
Trustworthy AI is fast – but also defensible.
