S03.7 · Financial Services & Capital Markets
Investment banks and broker-dealers intermediating capital markets, as agentic AI compresses trading and research cost lines.
Capital markets and broker-dealer activity — the intermediation layer between issuers, exchanges and institutional investors — carries no single consensus revenue figure; the research houses scope the business too differently. BCG estimates the global wholesale capital-markets pool (investment-banking fees plus fixed income, currencies and commodities, and equities) at $600-700B for 2024/2025E, while the global investment-banking fee pool alone runs $85-95B (2024, industry deal data). The near-term story is a cost story: agentic AI is compressing the junior-banker and research-analyst base faster than it is touching the relationship-driven origination moat that still protects complex, non-standardized transactions.
Off the 2022-2023 trough, investment-banking fees rebounded sharply — roughly 15-25% across 2024-2025 — though forward growth should settle at mid-single digits, tracking the deal and rate cycles that have always governed this business. More than half of global IB fees are generated in the US, with Europe a secondary pool and China and India growing their share of global equity capital markets activity on deeper domestic listing pipelines. The top five global banks command roughly 40% or more of the IB wallet. The long tail of boutique advisory firms and regional broker-dealers competes for the remainder on sector specialization or local relationships — never on balance-sheet scale.
Upstream sit issuers and exchanges; downstream, institutional investors and asset managers. The economics are cyclical and transactional by nature, and market-making adds a layer most fee businesses never carry: high balance-sheet capital intensity, because a dealer must hold inventory and take principal risk to provide liquidity. Broker-dealer licensing and capital-rule gating — net capital requirements in particular — is heavy. That gating has historically limited new entry to well-capitalized players, and it is a large part of why the business stays concentrated among global banks despite fee-pool volatility.
Three adjacencies frame the map: principal investing and asset management, electronic-trading technology, and private credit — the last of which now competes directly with syndicated bank lending for the same borrowers, making it a displacement threat to an existing revenue line rather than a new pool. Banks building out electronic-trading capability are responding to share already lost to non-bank market makers, who have taken meaningful volume in liquid, standardized instruments. Adding specialized sector-advisory expertise runs the other direction: origination in complex transactions still pays for deep relationships and technical knowledge that a generalist banker cannot easily replicate, and that is where incremental fee share is genuinely available.
Agentic AI is already visible in trade execution, research production and compliance surveillance — each a high-volume, rules-bound function, the standard first targets for automation. The compression lands on the junior-banker and research-analyst cost line, functions that traditionally absorbed a large share of front-office headcount in the first several years of a banking career. What AI has not yet displaced is the relationship-based origination moat for complex or bespoke transactions, where a banker's judgment, client trust and cross-institutional network still count for more than execution speed. The moat is thinnest exactly where the product is standardized: a plain-vanilla debt or equity offering is far more exposed to AI-driven disintermediation than a complex cross-border advisory mandate.
The parallel structural shift is the rise of AI-driven non-bank market makers, which blur the traditional line between regulated broker-dealer activity and proprietary trading. These firms provide liquidity through automated strategies at a scale and speed that increasingly rivals bank trading desks, without carrying the same regulatory capital burden — a structurally cheaper way to run the same risk. Trading and research automation is underway now. Displacement of relationship-based origination is the longer call, likely 5-10 years, and it turns on an unresolved question: how much of dealmaking judgment is genuinely automatable, and how much depends on trust built over time.