AI Agents in Capital Markets: The New Participants

Capital markets are adding a new class of participant. By late 2025, 68% of hedge funds were already using AI for market analysis and trading strategies, and 77% of banks had launched or soft-launched generative AI applications, according to industry surveys compiled in Caspian One's 2025 AI in Financial Services report. The shift now underway is not faster analysis for human desks. It is the arrival of AI agents that screen deals, underwrite risk, allocate capital, and settle transactions with limited human intervention. That changes who acts in a market, not just how fast they act.

From Tools to Participants

For most of the past decade, AI in finance meant decision support. Models flagged anomalies, ranked opportunities, and drafted documents, but a person pulled every trigger. Agentic AI removes that constraint. An agent does not return a recommendation and wait. It takes an action — placing an order, opening a position, generating and routing a document — autonomously, with real-world consequences, on behalf of an institution or end user.

The distinction matters for market structure. A tool sits inside one firm's workflow. A participant interacts with the market itself: it consumes prices, posts orders, responds to counterparties, and reacts to other agents doing the same. When enough autonomous agents operate concurrently, the market's behavior becomes a product of their interaction, not only of the humans who deployed them. Regulators have noticed. In its May 2026 Supervisory Toolkit for AI Use in Capital Markets, IOSCO singled out agentic AI — systems that "operate with increasing autonomy and make complex decisions without direct human supervision" — as the most significant supervisory challenge on the horizon.

What Agentic Finance Actually Does

The agentic stack maps cleanly onto the capital-markets lifecycle, and each stage is already being built.

Screening. Agents ingest deal flow, normalize unstructured documents, and rank opportunities against a mandate before a human reviews anything. This is the highest-volume, lowest-risk entry point, which is why it is furthest along.

Underwriting. Goldman Sachs has built tools that generate first drafts of pitchbooks and client presentations on in-house models trained on the firm's own templates, and CNBC reported the bank worked with Anthropic to build agents for due diligence, client vetting, and onboarding. Its firmwide GS AI Assistant launched on June 23, 2025, with roughly 10,000 employees using it at launch, per reporting from BankInfoSecurity.

Allocation and transaction. This is where autonomy carries the most weight. An agent that decides position sizing and then executes is no longer assisting — it is participating. Hedge-fund adoption for trading strategy is the clearest signal that this stage is moving from pilot to production.

Servicing. Post-trade — accounting, reconciliation, corporate actions, investor reporting — is repetitive, rules-bound, and well suited to agents. Goldman's reported work on trade and transaction accounting agents sits squarely here.

Read end to end, the pattern is an asset that can be sourced, underwritten, allocated, transacted, and serviced by software that acts rather than advises. The constraint is no longer capability. It is whether the surrounding market infrastructure can verify identity, enforce permissions, and produce an audit trail for a non-human actor.

The Infrastructure Gap

An AI agent cannot participate in a market that has no way to recognize it. Today's plumbing assumes a human or a registered legal entity on each side of a transaction — KYC tied to a person, settlement tied to an account, accountability tied to a signature. An autonomous agent fits none of those assumptions cleanly. Three problems surface immediately.

First, identity and permissioning. A market needs to know what an agent is authorized to do, on whose behalf, and within what limits — and to revoke that authority instantly. Second, settlement. If an agent transacts continuously, the asset and the cash leg need to move on rails that match that cadence; multi-day settlement reintroduces the human-speed bottleneck the agent was meant to remove. Third, auditability. Every autonomous action needs a verifiable record, because supervision of an agent depends on reconstructing why it did what it did.

Tokenized real-world assets address all three at the infrastructure layer rather than bolting controls on afterward. When an asset is issued as a programmable, compliance-aware token, permissions and transfer restrictions travel with the instrument, settlement is atomic and final, and every transaction is recorded by construction. That is the substrate an agent needs to act inside the rules instead of around them. Commertize's how-it-works overview describes how compliance is embedded into the asset itself, and the marketplace shows the kind of tokenized instruments an agent would screen and transact against.

Scale, Risk, and the Pace of Adoption

The capital flowing toward this is not speculative. Capgemini's estimate, cited via Statista, puts the agentic AI market at $5.1 billion in 2024 and projects it to surpass $47 billion by 2030, a compound annual growth rate above 44%. Finance is one of the heaviest-weighted verticals in those forecasts, given its data density and the direct link between automation and margin.

Adoption will not be uniform, and it should not be. Surveys still show only a minority of AI use cases reaching full production, and institutional investors are openly split — 71% expect AI to drive further growth while 41% now flag concern about an AI-driven tech bubble. Autonomy concentrates both the upside and the failure modes. An agent that mis-screens a deal produces a bad recommendation; an agent that mis-allocates and executes produces a bad trade. That is precisely why permissioning, position limits, and circuit breakers belong in the asset and venue layer, not only in the model.

The realistic path is graduated. Agents take over screening and servicing first, where errors are cheap and reversible, then move into underwriting and allocation as identity, settlement, and audit infrastructure proves out. Markets that build for non-human participants now — programmable assets, atomic settlement, machine-readable permissions — will be the ones where agentic capital can operate safely at scale.

What This Means for Market Operators

The strategic question for any platform is no longer whether to add AI features. It is whether the venue can support a participant that never sleeps, acts on its own authority, and expects the rules to be enforced by the instrument rather than by a back office. That requires assets that carry their own compliance logic, settlement that clears at machine speed, and a record that satisfies a supervisor reconstructing an agent's decisions after the fact.

This is the direction we are building toward at Commertize — a compliance-first digital capital markets platform for real-world assets, oriented toward capital markets where AI agents are first-class participants. The work is in the rails: programmable, permissioned, auditable assets that an agent can transact against inside the rules. For the broader market context, see related coverage in our news section, and for how tokenized instruments are structured, the tokens overview.

AI agents are no longer adjacent to capital markets. They are entering them as participants, and the markets that recognize, permission, and settle for them will set the structure everyone else operates inside.

Sources: IOSCO Supervisory Toolkit for AI Use in Capital Markets (May 2026), Statista — Global agentic AI market value (Capgemini), Caspian One — AI in Financial Services Report 2025, BankInfoSecurity — How Goldman Sachs, JPMorgan and AIG Are Deploying AI.

Have an asset you're evaluating for tokenization? Send the offering memo to deals@commertize.com or start at commertize.com/tokenize, and we will return a written tokenizability and capital-structure memo within 48 hours — free, no obligation.

Confidential review. No cost, no commitment, no calls unless it is a fit.