Capital markets AI — inside investment banks, hedge funds, asset managers — is simultaneously one of the earliest-adopting AI verticals and one of the most hype-resistant. The use cases that have shipped are narrower and less glamorous than the headlines suggest. This post is what we see at mid-tier banks and funds: what actually ships, what stays in the research lab, and the regulatory context that shapes both.
Research desks
Earnings transcript analysis. Every quarter, analysts read 50+ transcripts. AI summarizes, extracts KPIs, flags management commentary changes vs last quarter, identifies topics getting more/less airtime. Analyst time on first-pass reading drops 60-80%.
10-K/10-Q summarization. Same pattern on SEC filings. AI pulls out material changes, new risk factors, footnote oddities. Analyst reviews AI output instead of full document.
Thematic trade ideas. Cross-reference company data, macro signals, news flow. AI surfaces candidates for human judgment. Not autonomous trade generation — the human decides.
See investment research AI post for deeper dive on analyst copilots.
Trading floor
News sentiment scoring. Real-time news flow tagged for sentiment, entities, event types. Pre-existing use case (pre-LLM) enhanced by LLMs. Not revolutionary; incrementally better.
Trade execution copilots. For voice-traded markets (FX, some fixed income), AI summarizes recent quotes, suggests pricing, flags unusual patterns. Trader decides; AI informs.
Post-trade summaries. Daily, weekly summaries of activity for management, compliance, client reporting. Automated drafts; analysts review.
Algo trading. Still dominated by specialized ML and statistical models, not LLMs. LLMs have poor properties for high-frequency trading (latency, determinism). Where LLMs are applied, it's in longer-horizon strategies.
Risk and operations
Scenario analysis. Given a macro shock, how does the book behave? AI generates narrative descriptions of results, flagging risks humans might miss. Supplements quantitative models.
Regulatory reporting. Narrative portions of regulatory filings (CCAR, DFAST, MiFID transaction reporting commentary). AI drafts; regulatory team reviews. Submission-ready quality after human review.
Surveillance and market abuse. Trader chat surveillance has used NLP for years; LLMs improve detection of coded language and context-dependent patterns. Compliance wins.
See fraud detection post for the broader compliance AI landscape.
What doesn't ship
Fully autonomous trading agents. Regulatory and risk constraints too high. Every major firm has experimented; none have deployed without heavy human oversight.
LLM-driven portfolio management. Specialized quantitative models still win. LLMs are useful for narrative and exploration; portfolio construction remains quantitative.
Client-facing AI investment recommendations. Fiduciary risk. All client communications human-reviewed.
Real-time pricing from LLMs. Latency too high; determinism insufficient. Rule-based and model-based pricing wins.
Regulatory context
SEC and FINRA on both sides of the ocean (SEC in US, FCA in UK, BaFin in Germany, ESMA for EU). All are increasing scrutiny on AI use in capital markets.
Specific rules emerging: Model Risk Management (SR 11-7 in US) extended to AI/ML models. Firms document validation, monitoring, ongoing review of AI systems.
Documentation requirements substantial. Firms that treat AI as another tool without governance face regulatory questions. Firms with AI governance programs pass smoothly.
Build vs buy
Vendors: Bloomberg Terminal integrations, Symphony AI, AlphaSense, Kensho, many others. Mature category with established players.
Internal build common for proprietary analytics and trade ideas. Firms want the intellectual property. Build takes 6-24 months depending on scope.
Hybrid common. Buy for commodity capabilities (news ingestion, transcript analysis); build for differentiated analytics.
Where investment is landing next
Deeper integration between AI and workflow. Analysts don't want a new AI tool; they want AI inside Bloomberg, inside the OMS, inside email. Integration is the 2026-2027 story.
Agent workflows for research and trading. Multi-step tasks (find companies matching criteria, pull financials, construct comparables, generate thesis). Human-supervised but increasingly agentic. See agents post.
Private markets and alternative assets. More data becoming structured; AI adoption lags public markets but accelerating.