On 10 June, the Financial Stability Board published its first comprehensive framework for responsible AI adoption across the global financial system. The timing is telling — not because regulators have only just noticed artificial intelligence, but because the gap between what AI can do in portfolio management and what firms can prove it did is becoming uncomfortably wide.
The report — a consultation document titled Sound Practices for Responsible Adoption of Artificial Intelligence — lays out 12 principles covering governance, explainability, performance monitoring, and third-party risk. It arrives at a moment when asset managers are racing to embed AI across investment workflows, often without the documentation infrastructure to satisfy the regulators now paying close attention.
The Gap Between Adoption and Accountability
The numbers tell a clear story. Mercer‘s 2026 AI in Asset Management Survey found that 55% of asset managers have integrated AI into at least one investment process, with a further 27% running pilot programmes. Only 18% report no integration whatsoever. Perhaps more striking: 91% plan to increase their AI usage over the next twelve months.
Yet the same survey reveals a telling asymmetry. Whilst 74% of firms describe their AI use as “operational” — automating tasks and improving efficiency — and 69% use it as a “co-pilot” for analysis, only 6% say AI plays a role in actual decision-making. The most commonly cited benefits are enhanced operational efficiency (69%) and faster insights (55%). Improved returns and reduced volatility? Just 8% each.
Most firms are using AI to process information faster, not to make better investment decisions. That distinction matters enormously when regulators start asking what, precisely, the algorithm contributed to a portfolio allocation — and whether anyone can explain why.
What the FSB Is Really Asking
The Financial Stability Board (FSB) — the Basel-based body that coordinates financial regulation across the G20 — has structured its framework around two pillars: organisation-wide governance and lifecycle management of individual AI use cases.
The governance layer is familiar territory: board oversight, clear accountability, risk appetite alignment. But the lifecycle practices are where things get pointed. Sound practice 8 addresses explainability directly, urging firms to adopt “more explainable AI or consider compensating controls” where a model’s reasoning cannot be readily understood.
This is a direct challenge to the growing use of deep learning and reinforcement learning in portfolio construction. These models can be extraordinarily effective at identifying patterns in high-dimensional data, but their internal logic is notoriously difficult to articulate. A neural network might outperform a rules-based screen across a decade of back-testing, but asking it why it favoured one stock over another often produces answers that satisfy neither compliance officers nor supervisors.
The consultation also highlights “shadow AI” — staff using AI tools without authorisation or for unapproved purposes. For asset managers, this could mean an analyst feeding client data into a public large language model (LLM) or a portfolio manager running an unsanctioned screening tool. The governance gap is real, and the FSB is signalling that firms must close it.
Deterministic Scoring: The Quieter Alternative
Amid excitement about generative AI — the technology behind tools like ChatGPT — a less fashionable approach is gaining traction among firms that need their AI to be auditable.
Deterministic scoring systems use structured, rules-based pipelines to evaluate companies against specific criteria. Unlike generative models, which produce probabilistic outputs that can vary between runs, deterministic systems follow fixed logic: the same input always produces the same output. Every decision can be traced through its reasoning chain, and every data point cited.
This matters acutely in impact investing, where investors demand evidence that a fund’s holdings genuinely align with its stated values. A deterministic pipeline screening thousands of companies against supply-chain labour standards or environmental metrics can produce a cited evidence trail for every inclusion and exclusion. A generative model might produce a convincing narrative about why a company qualifies — but that narrative could differ tomorrow.
The International Organisation of Securities Commissions (IOSCO) published its own AI guidance in 2025, and the Bank for International Settlements (BIS) has separately explored how regulators should handle AI explainability. The direction of travel is consistent: if you cannot explain it, you may not be able to defend it.
What This Means for Cross-Border Firms
The FSB framework is a consultation, with responses due by 22 July 2026. But for firms operating across jurisdictions — particularly those navigating the UK-Swiss financial corridor — the practical implications are already materialising.
The Financial Conduct Authority (FCA) has made AI governance a stated priority for 2026, and the Swiss Financial Market Supervisory Authority (FINMA) has been exploring AI’s role in both supervision and supervised firms. For an asset manager using AI-driven screening to construct portfolios distributed across both markets, the question is not whether to document the AI’s reasoning — it is how to do so in a way that satisfies multiple regulatory regimes simultaneously.
Mercer’s survey found that 57% of firms have just one to five dedicated AI specialists on their investment teams. Building the audit trail may prove harder than building the model.
The FSB’s 12 sound practices may read like guidance today. For firms embedding AI into their investment processes without proportionate governance, they look increasingly like a preview of tomorrow’s compliance requirements.
