This article first appeared in the Financial Times. Reproduced with permission. © FT Limited 2025.

By Emre Kazim, co-founder and co-CEO of Holistic AI

Volatile markets in 2025 are exposing a dull spot in AI’s shiny veneer: model drift. Macroeconomic shocks like sky-high tariffs can upend AI models that were previously operating flawlessly, leaving banks at risk of jeopardising profits and customer trust.

Financial institutions are grappling with economic uncertainty amid rapid technological advancement. US President Donald Trump’s near-constant threat of tariffs, which began in April 2025, have triggered a chain reaction: plummeting stock markets, surging government bond yields and billions wiped off bank stocks in days.

Beneath this chaos lies a quieter crisis: AI model drift, a gradual degradation in the performance of machine learning models as they encounter data that diverges wildly from their training sets. When AI systems trained on historical data fail to adapt to new patterns, their predictions become less accurate.

While not a new phenomenon, model drift is acutely relevant today. The S&P 500’s 12 per cent drop in April, followed by a 9.5 per cent rebound after tariff suspensions, highlighted how quickly market dynamics can shift. AI models calibrated to stable trade conditions became obsolete overnight.

For financial institutions, the stakes are existential.

Trade wars

The most obvious risk is miscalculations in credit risk. Pre-tariff models underestimated default risks for industries hit by higher import costs, such as manufacturing and retail. Banks like Citigroup and Bank of America saw their loan portfolios degrade as borrowers struggled with rising prices.​

New fraud patterns are also emerging as consumers shift spending amid inflationary pressures. Legacy AI systems, trained on pre-crisis behaviour, are missing sophisticated phishing schemes targeting stressed households.​

Treasury yield spikes are also disrupting liquidity management models. Banks that rely on AI tools to optimise cash reserves have faced shortfalls as deposit withdrawals accelerate.

Tariffs have reshaped consumer and corporate behaviour.

JPMorgan Chase has stated that it expects an increase in small-business loan defaults in sectors exposed to Chinese tariffs. Most models trained on data from as recently as 2024 have not been trained to price in supply-chain disruptions.

Algorithmic trading systems are misreading bond market volatility, triggering erroneous sell-offs during yield spikes. Some robo-advisors that overweighted tech stocks missed the tariff-driven rally in defensive sectors like utilities.

Without continuous model updates, these errors compound. Unchecked model drift could materially impact bank profitability, with some estimates suggesting 3-5 per cent losses in annual profits for institutions lacking robust AI governance.

Regulatory scrutiny intensifies

The EU AI Act and US Model Risk Management (SR 11-7) now mandate rigorous drift monitoring.

The Federal Reserve’s 2025 stress tests incorporated tariff scenarios designed to expose weaknesses in banks’ AI-driven capital adequacy models.

Post-tariff job losses, which disproportionately affect low-income regions, could skew credit scoring models toward higher default rates for marginalised groups.

Ignorance is not an excuse. Regulators are penalising institutions which lack transparent retraining protocols. Earlier this year, the European Central Bank fined three banks €1.24mn for using outdated anti-money laundering models.

Reputational fallout

Research published earlier this year and in 2024 confirms that algorithmic errors, whether due to bias, drift or lack of explainability, inflict “significant reputational damage and financial losses on banks”, and may even destabilise the broader financial system if left unchecked.

When AI models produce biased or erroneous outcomes, such as unfair loan denials or inaccurate risk assessments, banks face public backlash. This erodes customer trust and can drive clients to competitors.

How to stave off model drift

Continuous monitoring Establish real-time monitoring systems to detect deviations in model performance promptly

Regular model updates Retrain models with recent data to ensure they reflect current trends and behaviours

Robust validation processes Implement rigorous validation protocols to assess model accuracy and fairness continuously

Cross-functional collaboration Encourage data scientists, compliance officers and business leaders to align AI models with organisational objectives and regulatory requirements

Invest in explainability Adopt AI systems that provide transparent decision-making processes and help to identify and correct drift-related issues.

The cost of inaction

The Penn Wharton Budget Model projects Trump’s tariffs will reduce US GDP by 6 per cent in the long term, with middle-income households losing $22,000 in lifetime earnings. For banks, outdated models will exacerbate these losses.

Unchecked drift could increase defaults by between 8 and 20 per cent in exposed sectors. Fines for inadequate model governance may exceed $500mn annually for top banks. M&A models that fail to price in tariff risks could lead to overpaying in an acquisition, a strategic misstep.

The tariff crisis is a wake-up call.

Financial institutions must treat AI not as a static tool but as a living system requiring continuous investment. This technology can provide significant benefits and a competitive advantage, but only if paired with robust governance.

In 2025, a bank’s ability to manage model drift could define its survival.