AI in capital markets: The real test starts with trust

AI in capital markets is moving beyond the copilot model toward systems that can execute increasingly complex work. New agentic systems can carry out multi-step work on their own, from pulling data to assembling risk outputs. 

As these capabilities move closer to production, the central question shifts from what AI can do to where its output can be trusted and where stronger controls are required. That distinction becomes critical when AI transitions from assisting with workflows to influencing calculations, decisions, and outputs that firms need to validate, reproduce, and defend.

AI adoption is easier to measure than its impact

The shift toward more capable AI is already being reflected in how financial services firms are deploying these technologies. According to the Cambridge Centre for Alternative Finance's 2026 Global AI in Financial Services Report, agentic AI is in active adoption among 52% of industry respondents, and 23% have reached the more mature Scaling or Transforming stages. Looking ahead, 81% expect agentic AI to be meaningfully achieved across the industry by 2030. For comparison, adoption stands at 71% for generative AI and 75% for classical machine learning.

The pace of adoption is easier to measure than the impact AI delivers. A 2025 randomized trial from nonprofit research group METR found that 16 experienced open-source developers using AI tools on familiar codebases took about 19% longer to complete tasks, despite expecting to finish them 24% faster. The study covered 246 real tasks in mature projects the developers knew well, so its results apply to that specific setting and should not be generalized to software development as a whole.

A larger 2026 study published through the National Bureau of Economic Research, drawing on data from more than 500,000 GitHub developers, found that AI coding tools sharply increased coding activity, but those gains shrank further down the production chain. For autonomous coding agents, a 240% cumulative increase in commits translated into an 80% increase in projects and just a 30% increase in actual releases. More code written does not necessarily translate to more software shipped. 

From assistance to action 

The industry is now shifting from generative AI that assists to agentic AI that acts. An agent does not wait for a person to drive each step. It can pull positions, call an analytics engine, chain scenario runs together, and flag exceptions along the way, completing work a person once handled one task at a time.

That shift raises the stakes. When an agent acts on financial data, its output can feed a trade, a risk limit, a valuation, a client statement, or a regulatory process. The more consequential the output, the more important it becomes to know how it was produced and whether it can be checked and reproduced.

The trust gap is now the binding constraint

Inconsistent output is only part of the challenge. AI systems can be configured to give far more repeatable answers, though not always identical ones. The harder question is what happens when AI becomes part of a process that produces a number the firm has to stand behind.

Firms recognize the issue, as shown in the Cambridge study where 60% of traditional financial institutions cited loss of human oversight as a top AI risk, compared with 42% of regulators. Industry, vendors, and regulators all ranked model hallucinations and unreliable outputs among their top two AI risks.

Regulators are also shifting from offering broad principles toward practical supervision. In May 2026, the International Organization of Securities Commissions (IOSCO) published its Supervisory Toolkit for AI Use in Capital Markets, a non-binding toolkit designed to help regulators oversee AI systems used by the firms they supervise. It covers areas including governance and risk management, third-party and outsourcing risk, disclosure, and recordkeeping and reporting. The European Securities and Markets Authority (ESMA) is also monitoring AI adoption and use in EU securities markets, and its February 2026 analysis drew on a survey of 728 entities. The implication for firms is that using AI does not reduce their responsibility for the outputs they rely on.

What production-grade AI looks like

The firms getting the most value from AI are the ones separating the agentic reasoning from the calculation. AI interprets intent, orchestrates workflows, and communicates results. Its role can expand fastest in the work around a pricing engine, from market data and positions to testing. The number itself comes from a deterministic, validated analytics layer, never from the AI model.

That separation only works if the calculation layer is built for it. It has to expose instruments, curves, and sensitivities in a form AI systems can call directly, without bespoke integration code. It has to keep every calculation traceable and reproducible. And it needs governance for a user that may not be human, with agent identity, permission scoping, and audit trails.

This is the architecture we've built at Numerix. NxCore Analytics Services (NAS) delivers deterministic, production-grade pricing and risk analytics across all asset classes through a single API that people and AI tools can call directly. It gives quant teams control over their analytics stack rather than ceding it to a black box.

What comes next for AI in capital markets?

The next phase of AI in capital markets will be shaped by how well firms meet the validation, audit, and oversight demands of production use. Adoption was the first hurdle. The next is whether firms can prove that AI-enabled workflows and outputs hold up under that scrutiny.

Over the coming weeks, we'll examine what that proof requires, including where AI has earned trust at the practitioner’s desk, what it takes to prove a pricing or risk number when AI is in the workflow, and how firms are composing agents around their analytics. We're covering these core topics in Trust, Verified, a three-part white paper series. The first two papers in the series, Trust by Task, and Beyond Plausible, are available now, and the third paper will be published shortly.