Primer · 2026-05-19 · Companion to the 2026-05-15 Agentic AI Primer

Agentic AI in Financial Services — What's Actually Live

Companion: docs/briefs/2026-05-15-agentic-ai-primer.md (read for the underlying mechanics — ReAct loops, tool use, MCP, the latency-budget argument). This piece is the FS application layer: who is shipping what, where the dollars are moving, and what's still science fiction.

Anchor question: "Are they building agents for trading? Are AI use cases piling up in hedge funds and Citadels of the world? What are the use cases of a bank?"

One-line answer. Yes, the use cases are piling up — but almost all the production deployment so far is in knowledge-worker augmentation (research, coding, document summarization, advisor and banker copilots) and back-office workflows (KYC/AML, claims, customer service). Agents on the actual trade execution path remain rare and constrained, and HFT is structurally off-limits for LLM-style agents on latency grounds. The biggest disclosed AI spend in the sector is JPMorgan at ~$2B/year specifically on AI inside a $19.8B 2026 technology budget — and JPM has reclassified AI from "discretionary innovation" to "core infrastructure."

Terms defined on first use: HFT (high-frequency trading — strategies that execute in microseconds to milliseconds, where any added wall-clock latency destroys the edge); KYC (Know Your Customer — the identity-verification and risk-rating workflow banks run on every new account); AML (Anti-Money Laundering — the transaction-monitoring and suspicious-activity-reporting workflow run continuously on existing accounts); SR 11-7 (the 2011 US bank-regulator framework that governs how financial institutions validate, document, and govern any quantitative model used in decisions — including AI models — superseded April 17, 2026 by revised guidance that explicitly does not yet cover generative or agentic AI); RAG (retrieval-augmented generation — the standard technique of fetching relevant documents from a vector database and including them in an LLM prompt, rather than relying on what the model memorized in training); CIB (corporate and investment banking — the sell-side trading, M&A advisory, and capital-markets businesses inside a universal bank); CCAR (Comprehensive Capital Analysis and Review — the annual Fed stress test that requires banks above $100B in assets to model losses under macro-scenario shocks).

1. Map of FS AI use cases by subsector

The seven subsectors below organize the rest of this piece. The pattern, read top-to-bottom: as you move from HFT at the top (microsecond decisions, no LLMs anywhere on the hot path) toward insurance back office at the bottom (multi-day workflows, agents already shipping), the latency budget relaxes and the agentic surface area expands.

Subsector Where AI is live today Latency-sensitive on hot path? Agentic depth
HFT / quant trading (Citadel Securities, Jane Street, Virtu, Hudson River Trading) ML for signal generation and execution; classical/deep ML, not LLMs Microsecond budget — no LLM on hot path None on execution; offline training only
Discretionary hedge funds (Citadel, Bridgewater, multi-managers) Research assistants, document summarization, machine-learning macro funds Decision horizon: minutes to days Read-only research agents; one full-AI macro fund (Bridgewater AIA)
Long-only asset managers (BlackRock, Vanguard, Fidelity) Aladdin/eFront Copilot, client-summary generation, advisor commentary, internal RAG Decision horizon: days to weeks Workflow agents inside the platform, not on trade execution
Investment banking (CIB) (Goldman, Morgan Stanley, JPM CIB) Coding agents (Devin, Copilot), banker copilots (GS AI Assistant), research assistants (AskResearchGPT) Days-to-weeks workflows for advisory; sub-second for trading Heavy on coding side; banker copilots are RAG, not autonomous
Retail brokerage (Robinhood Cortex, IBKR Ask IBKR, Schwab) Conversational research assistants; trade-idea generation with human confirmation Order entry has human in the loop Limited agency — agent surfaces an order, customer clicks
Retail / commercial banking (JPM, BofA, Citi, Wells, HSBC) Customer-service bots (BofA Erica — 200M Q4 2025 interactions), document review (JPM COIN), KYC/AML, fraud, internal LLM platforms (JPM LLM Suite — 200K users) Minutes to days for the workflows AI touches Agentic deployment beginning in 2025-26 for KYC/AML, advisor support
Insurance (Lemonade, Progressive, Allstate, large P&C carriers) First-notice-of-loss intake, photo-based damage estimation, claims triage, agent-assist Hours to days for claims Highest agentic depth in FS — Lemonade processes simple claims end-to-end

Two cuts cross the table: front-office (revenue-touching) versus back-office (cost-cutting), and on-trade-path versus off-trade-path. The honest read is that almost every shipped deployment is back-office and off-trade-path — augmentation of human decisions, automation of paperwork, deflection of customer-service load. Front-office, on-trade-path agentic AI exists in two narrow forms: Bridgewater's AIA Macro Fund (machine-learning-led, but not LLM-agent-led on execution) and a thin layer of retail-brokerage assistants that require human click-confirmation before any order goes in.

2. What's actually being done today — per public disclosures

This section walks through the named deployments under review. Each subsection cites primary sources or first-rate trade press; vendor self-disclosure is flagged as such.

2.1 JPMorgan — LLM Suite, IndexGPT, COIN

JPM is the cleanest single-company case study in FS, because of the volume of public disclosure. Three named systems:

Spend disclosure: per third-party reporting on JPMorgan disclosures, "JPMorgan Chase spends $2 billion per year on AI and finds an equal amount of cost savings as a result" (eMarketer — In the AI arms race, JPMorgan is unstoppable, Oct 9 2025 · local: …/snapshots/2026-05-19-emarketer-jpmorgan-ai.html). Crypto.news framing of the same disclosures: JPMorgan "has moved its AI spending out of the discretionary innovation category and placed it alongside data centers, payment systems, and core risk controls inside its $19.8bn total technology budget for 2026" (crypto.news — JPMorgan makes AI core infrastructure spending · local: …/snapshots/2026-05-19-crypto-news-jpmorgan-ai.html). The reclassification matters: it removes AI spend from the line that gets cut first when markets turn.

2.2 Goldman Sachs — GS AI Assistant + Devin

Two layers, one for knowledge workers, one for engineers.

2.3 Morgan Stanley — AI @ Morgan Stanley Assistant, Debrief, AskResearchGPT

Three deployments stack in wealth and research.

2.4 BlackRock — Aladdin Copilot, eFront Copilot, Auto Commentary

Aladdin is the dominant institutional asset-management platform — approximately $25 trillion in assets are managed on Aladdin and over 1,000 organizations use it for portfolio management, risk analytics, trading, operations, and compliance (BusinessStats — $25T Assets BlackRock Aladdin Platform Statistics 2026, updated March 2026 · local: …/snapshots/2026-05-19-businesstats-aladdin.html). Three AI surfaces sit on top of it:

2.5 Bridgewater — AIA Labs and the AIA Macro Fund

Bridgewater under Co-CIO Greg Jensen and Chief Scientist Jasjeet Sekhon launched the AIA Macro Fund in July 2024 with approximately $2B initial capital (Fortune, Bloomberg). Bridgewater's own AIA Labs page describes the systems as managing "billions of dollars" and frames AIA as "an artificial investor designed to perform rigorous, explainable, fundamental research" with the AIA Forecaster described as "the first publicly documented AI system to verifiably match the performance of expert human forecasters at scale." Bridgewater emphasizes causal reasoning and explainability rather than labeling the architecture as ML- or LLM-agent-driven. Secondary press reports describe the fund as using proprietary ML models alongside LLMs from OpenAI, Anthropic, and Perplexity, with a regime-classification approach (growth acceleration / stagflation / deflation) positioning across FX, commodities, government bonds (Bridgewater — AIA Labs · Fortune — Bridgewater $2B fund · Bloomberg Professional Services · Institutional Investor — Behind Bridgewater's Surge). Specific 2025 return figures (an 11.9% figure has circulated in secondary trade press) are not confirmed on Bridgewater's own disclosure and should be treated as secondary-attribution only.

The Bridgewater frame matters because it's the only large-fund public commitment to AI-driven decision-making at the fund-product level. Caveat: per Bridgewater's framing, AIA is closer to an "artificial investor" doing fundamental research than to an LLM agent calling trade execution — the LLMs feed signal generation and reasoning; the trade decision is a separate model output operating under Bridgewater's standard risk-control architecture.

2.6 Citadel — AI Assistant for equities investors

Hedge fund Citadel rolled out an internal AI Assistant for equities investors, disclosed at Reuters Next in December 2025 by CTO Umesh Subramanian. Trained on licensed transcripts, regulatory filings, brokerage research, and Citadel's proprietary strategies. Surfaces risks, generates tailored research, and creates reading lists aligned to investor portfolios. Use spans "almost all of the firm's equities investors" (Hedgeweek — Citadel rolls out AI research assistant · local: research/2026-05-19-financial-services-ai-primer/snapshots/2026-05-19-citadel-hedgeweek.html · Investing.com — Citadel debuts AI tool for equities investors (Reuters wire)). Ken Griffin himself has been explicit that investment decisions remain with humans. Speaking at the JPMorgan Robin Hood Investors Conference in October 2025, Griffin said that while GenAI can enhance productivity, it "just falls short" for uncovering alpha — and noted that the technology has not replaced the human work of investing (Hedgeweek — Generative AI fails to deliver alpha for hedge funds, says Griffin · local: …/snapshots/2026-05-19-hedgeweek-griffin-genai.html · Bloomberg — Ken Griffin Says GenAI Fails to Help Hedge Funds Produce Alpha, Oct 15 2025 (primary source — 403 to crawler + Wayback; verbatim partial quote sourced via Hedgeweek)).

The contrast with Bridgewater is important. Bridgewater is willing to let ML drive the trade; Citadel is explicit that AI is a research-augmentation tool only.

2.7 BlackRock peers — Vanguard, Fidelity

2.8 Bloomberg, FactSet, S&P Capital IQ, Moody's, MSCI — the data-vendor layer

The financial-data vendors are racing to embed an LLM-driven interface over their corpora before clients build their own.

2.9 Buy-side AI startups — Hebbia, Rogo, AlphaSense

The startup layer under review. Three names worth tracking:

Treat startup vendor self-disclosure with the same skepticism due any vendor pitch — the customer counts are real (Rogo's named-client list is verifiable through bank disclosures), but the workflow-completion claims are not yet independently audited.

3. Latency-sensitive vs not — the structural constraint

This is the single most important framing for the question of whether banks are building agents for trading. The agentic-primer §3 covered the hardware physics in detail; here is the FS-application version.

Workflow Latency budget LLM agent viable today?
HFT execution (Citadel Securities, Jane Street market-making, Virtu) Microseconds (1µs–10µs) No. A single LLM forward pass takes 50ms+. Agents add tool calls on top. Three to five orders of magnitude too slow.
Mid-frequency quant (Bridgewater AIA Macro, Renaissance, Two Sigma signal generation) Minutes to hours per trade decision Partial. ML models (not LLM agents) drive trade decisions; LLMs feed signal extraction and alt-data parsing.
Discretionary research / fundamental (Citadel L/S, hedge funds, asset managers) Hours to days Yes. Research agents are the most-shipped front-office use case.
M&A advisory, ECM/DCM, structured products Days to weeks Yes. Banker copilots (GS AI Assistant, Rogo) compress junior-banker work substantially.
Back office — KYC, AML, claims, customer service Minutes to days Yes — and this is where agents are most fully deployed.
Compliance and surveillance Real-time monitoring with human review on flag Yes for the alert-triage layer; not for trade-blocking.

The latency story has a second wrinkle: even where LLM-agent execution would be too slow, the model training and feature engineering on the back end runs in agent-like loops over historical data. Arc Compute's HFT-focused write-up describes the resulting split: "FPGAs handle the critical nanosecond loop. GPUs handle signal generation, simulation, research workloads, and real time risk." The same piece names a concrete HFT example: "Lynx, a proprietary trading firm, migrated from the public cloud to Arc Compute's on premise NVIDIA HGX B200 systems" — eliminating cloud jitter and enabling larger model training without sacrificing the FPGA-anchored execution path (Arc Compute — Becoming AI Native in High Frequency Trading: Why GPUs Are Now Essential · local: …/snapshots/2026-05-19-arccompute-hft-lynx.html). The training infra is agentic-shaped; the production hot path is not.

4. Compliance + regulatory friction — what's actually constraining deployment

The latency physics is one constraint. The regulatory regime is the other — and it is loosening at exactly the moment large banks are rolling out generative AI broadly.

SR 11-7 and its April 2026 successor. SR 11-7 is the Fed/OCC/FDIC framework for model risk management — the rules banks above $30B in assets must follow when validating, documenting, and governing any model used in business decisions. Until April 17, 2026, SR 11-7 (issued 2011) was the binding standard, and regulators applied it to AI/ML models with the same rigor as traditional statistical models. On April 17, 2026 the three agencies issued revised guidance that supersedes SR 11-7 — and notably, the revised guidance explicitly does not cover generative or agentic AI. The agencies have said they will issue a separate request-for-information on generative and agentic AI in the near future (Sullivan & Cromwell — Federal Banking Agencies Issue Revised Guidance on Model Risk Management · OCC Bulletin 2026-13).

The carve-out is the load-bearing fact. It signals that US bank regulators do not currently have a settled framework for governing generative AI in regulated decisions — which is why deployment has clustered in augmentation roles where a human is on the decision and the AI's output is reviewable, rather than in decision-substitution roles where the model is making the call.

FINRA Regulatory Notice 24-09 (June 2024). Confirms that FINRA Rule 2210 (broker-dealer communications) governs all public-facing communications, regardless of whether they were produced by a human, a third-party vendor, or an AI model. Concretely: AI-generated marketing copy, AI-drafted advisor communications, and AI-generated research summaries are all in scope, must meet accuracy and fair-dealing standards, and must carry the right disclosures (FINRA — Artificial Intelligence key topic page · A-Team Insight summary). FINRA has also begun enforcement on algorithm-related supervisory failures — Interactive Brokers was fined $475,000 and censured on 24 October 2024 for securities-lending segregation deficits stemming from a faulty algorithm and inadequate supervision (including allowing an unregistered person to oversee changes to the securities-lending algorithm) (Securities Finance Times — FINRA issues US$475,000 fine to Interactive Brokers · local: …/snapshots/2026-05-19-finra-interactive-brokers-fine.html).

EU AI Act. Entered into force 2024 with phased application. Per Hogan Lovells, Annex III classifies as high-risk AI systems used for evaluating the creditworthiness of natural persons or establishing their credit score (with an explicit exception for financial fraud detection), as well as systems used for risk assessment and pricing in relation to natural persons in the case of life and health insurance — these trigger the full conformity-assessment, risk-management, data-governance, and human-oversight regime. Fraud-detection AI is carved out of the high-risk regime and treated separately; AML risk-profiling is governed by parallel financial-services rules (DORA, CRR/CRD) rather than being explicitly enumerated in Annex III. Full enforceability for high-risk systems is August 2026, though proposed Digital Omnibus amendments would push that to late 2027 (Hogan Lovells — AI regulation in financial services · local: …/snapshots/2026-05-19-hogan-lovells-eu-ai-act-wayback.html · EBA — AI Act implications for the EU banking sector).

Net pattern across regulators: the binding constraint right now is process — documentation, validation, human oversight — not output prohibition. That means generative AI deployments are economically feasible if banks are willing to absorb the governance overhead, which the largest banks are.

5. Where the buy-side and sell-side are actually moving dollars

This is the central question. Three observable data points from public disclosures:

  1. JPMorgan — $2B/year on AI specifically, inside $19.8B 2026 technology budget. AI moved out of discretionary innovation and placed alongside data centers, payment systems, and core risk controls per October 2025 disclosures (eMarketer — JPMorgan unstoppable, Oct 9 2025 · local: …/snapshots/2026-05-19-emarketer-jpmorgan-ai.html · crypto.news — JPMorgan makes AI core infrastructure spending · local: …/snapshots/2026-05-19-crypto-news-jpmorgan-ai.html).
  2. Bank of America — $13B 2025 total technology spend, of which >$4B on new technology initiatives. Per Moynihan's letter to shareholders (BofA 2025 Annual Report, p. 7): "in 2025 we spent over $4 billion on new technology initiatives alone — representing just a part of our total $13 billion technology expenditure for the year." Erica (BofA's customer-facing AI agent, in market since 2018) hit ~200 million interactions in Q4 2025 across 20 million users (BofA 2025 Annual Report, p. 7): "We deployed an AI agent (Erica®) in 2018; 20 million people used it in the fourth quarter of 2025 nearly 200 million times" (Bank of America 2025 Annual Report, SEC EDGAR PDF · local: …/snapshots/2026-05-19-bofa-annual-report-2025.pdf · web letter version: BofA Newsroom — Moynihan 2026 shareholder letter · local: …/snapshots/2026-05-19-bofa-moynihan-letter-2026.html).
  3. Goldman Sachs — scaling Devin from "hundreds of instances, potentially growing to thousands" with human supervision on every instance, alongside ~12,000 human developers (TechCrunch — Goldman Sachs is testing viral AI agent Devin · local: …/snapshots/2026-05-19-goldman-devin-techcrunch.html).

A useful frame: the named systems above tell you who is spending; the regulatory carve-out from §4 tells you why the spending is currently aimed at augmentation rather than decision-substitution; and the latency table in §3 tells you where in the workflow the dollars are landing. The largest banks are spending tens of billions of dollars per year on technology, of which AI is a growing but still single-digit-billions slice — and almost none of it goes to on-trade-path agents.

McKinsey's own framing on agentic AI in KYC/AML is more granular than a single cost-reduction number: traditional analytical AI offers "15 to 20 percent productivity uplifts" on investigation handling, while agentic AI — where one human supervises ~20 AI workers — yields "productivity gain[s that] can be significant—anywhere from 200 to 2,000 percent, our experience shows" (McKinsey — How agentic AI in banking drives KYC/AML transformation · local: …/snapshots/2026-05-19-mckinsey-kyc-wayback2.html). Those are forward-looking experience-based ranges from a consultancy, not realized run-rate savings disclosed by a named bank.

6. Investment angle

Four buckets. None are clean public pure-plays except in the infrastructure layer.

Pure-play AI-for-FS vendors (mostly private). Hebbia, Rogo, AlphaSense, Paradigm. Rogo's $300M+ total raised across a four-round arc culminating in the April 2026 Series D tells you the venture market is pricing this as a winner-take-most category. The competitive question: does a Rogo-shaped horizontal-banker-copilot survive when (a) FactSet Mercury and S&P Capital IQ both ship the same workflow inside their data subscriptions, and (b) JPMorgan, Goldman, and Morgan Stanley each build internal versions on the foundation models directly? The honest answer is that this is an unsettled market and the equilibrium is not in. Public-market exposure to this layer is near-zero today — Hebbia/Rogo/Paradigm are all private.

Incumbent FS-data vendors. Bloomberg (private), FactSet (FDS), S&P Global (SPGI), Moody's (MCO), MSCI (MSCI), London Stock Exchange Group (LSE.L). The thesis here is defensive — the existing data moats let the incumbents ship LLM-driven interfaces over their corpora before the startup layer can dislodge them. FactSet Mercury, Moody's Research Assistant, and S&P's ChatIQ are the public examples. The risk: agentic interfaces commoditize the data-vendor relationship, because what the customer pays for becomes the workflow rather than the raw data feed. Open question whether MSCI, MCO, SPGI are net winners or net at-risk from this transition.

Infrastructure plays — the cross-thesis bridge. This is the bucket that connects this FS-vertical primer to John's main AI-infra thesis. Every named FS deployment above runs on NVIDIA GPUs for both training and inference (HFT shops are the exception that proves the rule: even Lynx, an HFT shop, is on B200s for training; production inference moves to FPGAs/ASICs). The cluster-level economics — fiber, power, CPU pull-through, switching, KV-cache offload — all map back to the BOM-archetype lens carried in the broader thesis. Net: AI-in-FS adds incremental token volume to the demand side of the bottom-up token-economics model; it does not require a separate sub-thesis. The investable names are the same NVIDIA / network / fiber / CPU / cooling names already on John's screen.

At-risk incumbents. Three categories carry the most credible disruption risk:

7. What's still science fiction vs near-term reality

A clear list, mapped to time horizons. Anchor each bucket to a primary-source signal where available.

Happening now (in production, 2025-26): - Knowledge-worker copilots at every major bank (JPM LLM Suite, GS AI Assistant, Morgan Stanley Debrief) - AI coding agents in software-engineering org charts (Goldman + Devin; JPM + Copilot) - Banker copilots / pitch creators (FactSet Mercury, Rogo) - Retail customer-service chatbots (BofA Erica, Klarna's case) - Insurance claims triage and FNOL intake (Lemonade end-to-end; Progressive / Allstate assist-mode) - KYC and AML alert triage with machine-learning false-positive reduction (HSBC, Citi) - ML-driven trade signal generation at quant and macro funds (Bridgewater AIA Macro, Two Sigma, Renaissance — not LLM agents, but the pipeline that feeds them increasingly is)

12-24 months out (visible technical path, regulatory/operational work ongoing): - End-to-end agentic KYC and AML at scale (McKinsey's order-of-magnitude productivity-gain framing — 200-2,000% per supervised-agent unit; first production deployments at HSBC and Citi disclosed but not yet at full run-rate) - Wealth-management AI providing personalized investment guidance directly to clients (Vanguard 2026 chatbot; expect Schwab, Fidelity to follow) - M&A advisory workflow agents (Rogo Felix, FactSet Pitch Creator) handling first-draft CIMs, deal screens, buyer lists end-to-end with banker review - Retail-brokerage AI handing off completed trade tickets to human-confirmation step (Robinhood Cortex, IBKR Ask IBKR; both already shipped in 2025-26 but adoption still early) - Generative-AI-specific regulator guidance in the US (Fed/OCC/FDIC RFI pending post April 2026); EU AI Act high-risk enforcement (Aug 2026 or late 2027 depending on the Omnibus)

Not happening anytime soon (5+ years, structurally constrained): - LLM agents on the HFT execution hot path. The latency budget is three to five orders of magnitude wrong. The only way this changes is a structural shift in model architecture that collapses an LLM forward pass into the microsecond regime — not visible on any roadmap. - AI-only investment decisions at large discretionary funds with fiduciary obligations. Even Bridgewater's AIA, which is the most aggressive disclosed deployment, is ML-driven rather than LLM-agent-driven on execution, and operates inside a clearly disclosed wrapper for LP consent. The fiduciary-duty surface area for an LLM-agent-only decision remains untested in court. - Fully autonomous credit decisions on consumer or commercial lending in regulated jurisdictions. EU AI Act explicitly classifies these as high-risk; SR 11-7's revised guidance carves out generative AI pending separate rulemaking. Augmentation yes; substitution no. - Sell-side trading-desk pricing run by LLM agents without human dealer approval. Same regulatory pattern — Rule 15c3-5 (SEC market-access rule) puts hard limits on automated trading without pre-trade risk checks, and FINRA Notice 24-09 confirms AI doesn't get to bypass that.

8. What this means for John's broader thesis

Three connections back to the AI-infrastructure capex thesis.

(a) FS is a real but small token-demand bucket — and structurally back-loaded. The bulk of FS AI workload is text-heavy back-office and research workflows that consume tokens but not at hyperscaler volumes. The latency-sensitive trade execution path stays off LLMs. So the demand-side contribution from FS to the bottom-up token model is meaningful, but the right-tail comes from agentic knowledge work (Rogo, GS AI Assistant, FactSet Mercury, Moody's Research Assistant) and back-office automation (KYC, claims), not from FS becoming a trading-side LLM consumer.

(b) The data-vendor disruption story is the most asymmetric public-equity bet in this vertical. Specifically: does Bloomberg's quiet productization of BloombergGPT (i.e., bundling AI into the Terminal at no price increase) defend the Terminal moat, or does Rogo/Hebbia/AlphaSense compress the data-vendor wallet share at the margin? The public names exposed are FDS, SPGI, MCO, MSCI, LSE.L. Worth a deeper dive in a separate brief — the variance in outcomes here is high.

(c) The HFT-shop GPU spend is real and rising — but it's training spend, not inference spend. Quant shops moving training workloads from cloud to on-prem B200 / GB200 systems is consistent with the cluster-buildout demand picture, but does not generate per-token inference volume the way hyperscaler API consumption does. Watch for the disclosed quant-fund GPU-cluster builds as a complementary signal alongside the hyperscaler capex line.

Bottom line. The question — "are AI use cases piling up in hedge funds and Citadels of the world?" — has a clean answer: yes, and they're piling up in research-augmentation roles, not in trade execution. The structural diagnosis — mostly white-collar / back-office today, not mission-critical regulated workflows — holds, supported by both the latency math and the regulatory carve-out. The investment thesis line is unchanged: own the picks-and-shovels for the AI-infra buildout that feeds all of this; the FS vertical is one of many demand buckets, and its uniquely large compliance friction means it'll be a steady, fat demand stream, not a step-function one.

Sources

§2.1 — JPMorgan

§2.2 — Goldman Sachs

§2.3 — Morgan Stanley

§2.4 — BlackRock

§2.5 — Bridgewater

§2.6 — Citadel

§2.7 — Vanguard / Fidelity

§2.8 — Data vendors

§2.9 — Startups

§3 — Latency / HFT

§4 — Regulation

§5 — FS spending

§6 — Insurance

Sister deliverable