Francesco Di Costanzo
Back to articles

(15) Agents Are All You Need: Why the Next Great Hire in Finance Is a One-Person Agent Team

/21 min read

Download paper (PDF)

In article (10), "The Agent Layer Is Becoming the Control Plane of the AI Economy," I argued that the organisations accruing durable advantage from AI are not those adding AI to existing workflows but those restructuring the workflow itself around the agent layer. This article extends that argument from the macroeconomic to the practitioner level — specifically the capital markets context I know directly from years at Moody's, Société Générale, and Hoist Finance, and now as CEO of Jaja Finance.

The honest version of this thesis is narrower than the headline implies. The claim is not that agents replace analysts universally, or that human judgment is no longer necessary, or that the model is ready for autonomous execution of investment decisions. The claim is that for a well-defined category of analytical work — document-heavy, monitoring-intensive, template-driven, high-repeatability workflows — a single senior professional with the right profile and a well-architected agent team can match the analytical throughput of a much larger conventional team, at a fraction of the cost, while keeping human judgment precisely where it belongs: at every consequential decision node.

The Margin Compression That Makes This Structural

The economic context matters because it changes "interesting" into "necessary." UK investment management operating margins fell to 18% in 2023–2024, down from approximately 29–32% in 2019–2021, as costs rose 36% against revenue growth of only 15%, according to the Investment Association's annual survey. European asset managers' average operating profit dropped to 11.1 basis points of AUM — the lowest level since the 2008 financial crisis, per EFAMA data reported by ETF Stream. Meanwhile, equity mutual fund expense ratios have fallen 62% since 1996, from 1.04% to 0.40% by 2024, per the Investment Company Institute's March 2025 fee data. Lower fees with rising costs and compressed margins produce a single imperative: increase analysis per unit cost without increasing governance risk.

The headcount data already expresses this logic. UK investment management direct employment fell 1% in 2023, even as UK-managed assets recovered from the prior year's sharp decline, with the Investment Association noting that technology, outsourcing, and efficiency measures may have contributed to the divergence. Lyneer Search Group's labour market study, as reported by Forbes, found hiring for 22–25-year-olds in finance fell 6% between late 2022 and July 2025, while hiring for workers aged 35–49 grew over 9%. Entry-level finance positions have fallen by approximately 24% since the generative AI boom began, according to research by Rezi.ai cited by Forbes. The talent market is already bifurcating — firms are retaining senior judgment and automating the throughput that junior staff previously provided. The one-person orchestrator model is not a prediction; it is an acceleration of an observable trend.

The Cost Gap and What It Actually Represents

A fully loaded 5-person mid-market UK analyst team — using benchmarks from Selby Jennings's 2024 Global Investment Management Compensation Guide, the CFA Institute's 2024 Compensation Study, and BLS employer cost data showing benefits at approximately 30% of total compensation — costs approximately £1.0–1.3M per year. Scaling to eight analysts, the range rises to £1.9–2.3M. A single senior professional at £200,000–£350,000 base, plus AI infrastructure at £20,000–£50,000 annually, produces a fully loaded cost of approximately £330,000–£600,000, with the lower end of that range representing minimal bonus realisation. The gap is 2.2x to 7x depending on configuration, with a median around 3–4x.

The comparison requires precision. The senior orchestrator is not replacing the entire intellectual contribution of a five-person team. What agents can automate is the analytical throughput that currently requires team size: data extraction from earnings releases and regulatory filings, credit monitoring across a broad portfolio, first-draft memo production, regulatory change tracking, portfolio surveillance triggers, and due diligence data gathering. McKinsey's "Seizing the Agentic AI Advantage" documents a bank credit-risk memo workflow where agent deployment produced 20–60% productivity increases and a 30% reduction in credit turnaround time. A separate McKinsey case study of a 500-person research firm found multi-agent pipeline deployment delivered over 60% productivity gains and over $3M in annual savings. These are case-study estimates, not randomised trials. The floor is the BCG and Harvard Business School preregistered experiment with 758 consultants: 12.2% more tasks completed, 25.1% faster completion, and over 40% quality improvement on tasks inside the AI capability frontier.

What the Architecture Actually Looks Like

The architecture of a functional one-person agent team for capital markets work is not a chatbot and a spreadsheet. It is a structured multi-agent system in which a human orchestrator sets objectives, constraints, and verification criteria at the top of the stack, with specialised agents executing defined sub-tasks beneath, a quality assurance layer catching errors before human review, and an explicit human decision checkpoint before any output enters a consequential workflow. The following diagram renders this structure for a capital markets context:

╔══════════════════════════════════════════════════════════════════╗
║              HUMAN ORCHESTRATOR (Domain Expert)                  ║
║   • Sets objectives, constraints, and verification criteria      ║
║   • Defines task scope (inside vs. outside capability frontier)  ║
║   • Accountable for all outputs under SM&CR / fiduciary duty     ║
╚══════════════════════════════════════════════════════════════════╝
                              │
          ┌───────────────────┼───────────────────────┐
          │                   │                       │
          ▼                   ▼                       ▼
  ┌───────────────┐  ┌────────────────┐  ┌──────────────────────┐
  │ RESEARCH      │  │ DUE DILIGENCE  │  │ PORTFOLIO MONITORING │
  │ AGENT         │  │ AGENT          │  │ AGENT                │
  │               │  │                │  │                      │
  │ • Earnings    │  │ • Document     │  │ • Daily P&L alerts   │
  │   extraction  │  │   review       │  │ • Covenant tracking  │
  │ • Sector      │  │ • Management   │  │ • Rating migration   │
  │   synthesis   │  │   background   │  │ • Trigger events     │
  │ • Peer comps  │  │ • Legal entity │  │ • Exposure drift     │
  └───────────────┘  │   mapping      │  └──────────────────────┘
          │          └────────────────┘            │
          │                   │                    │
          ▼                   ▼                    ▼
  ┌─────────────────┐  ┌──────────────────┐  ┌──────────────────┐
  │ MARKET          │  │ REGULATORY       │  │ MEMO DRAFTING    │
  │ SURVEILLANCE    │  │ TRACKING AGENT   │  │ AGENT            │
  │ AGENT           │  │                  │  │                  │
  │ • Price/spread  │  │ • FCA/ESMA       │  │ • IC memo v1     │
  │   anomalies     │  │   rule changes   │  │ • Credit summary │
  │ • Vol regime    │  │ • EU AI Act      │  │ • Risk narrative │
  │   shifts        │  │   compliance     │  │ • Client update  │
  │ • Positioning   │  │ • Reporting      │  │   drafts         │
  │   signals       │  │   deadlines      │  └──────────────────┘
  └─────────────────┘  └──────────────────┘
          │                   │                    │
          └───────────────────┴────────────────────┘
                              │
                              ▼
          ╔═══════════════════════════════════════════╗
          ║        VERIFICATION / QA LAYER            ║
          ║  • Source citation cross-check            ║
          ║  • Numerical consistency validation       ║
          ║  • Hallucination flag via secondary LLM   ║
          ║  • Completeness against task specification║
          ╚═══════════════════════════════════════════╝
                              │
                              ▼
          ╔═══════════════════════════════════════════╗
          ║          HUMAN REVIEW CHECKPOINT          ║
          ║  • Orchestrator reviews flagged items     ║
          ║  • Calibration against domain knowledge   ║
          ║  • Approval required before downstream    ║
          ║    use in investment decisions or client   ║
          ║    communications (SM&CR accountability)  ║
          ╚═══════════════════════════════════════════╝

This architecture is not theoretical. Man Group's AlphaGPT, deployed from July 2025 and described by Bloomberg as "the arrival of agentic AI at the world's largest listed hedge fund," operates a three-agent system — named the Idea Person, Implementer, and Evaluator in Man Group's own November 2025 publication — that generates, codes, and backtests quantitative trading ideas before a human research team reviews viable concepts. Citadel's AI Assistant is used by nearly all equity investors at the $71B fund. Morgan Stanley's AskResearchGPT placed GPT-4 over 70,000 proprietary research reports, producing a 3x increase in questions asked relative to the prior AI tool and reducing response time to one-tenth of its prior level. JPMorgan's LLM Suite has been made available to 250,000 employees and is projected to generate $2B in annual value. These are production deployments, and they share a consistent structural feature: a human expert layer that sets objectives and owns outcomes, with agents executing the document-heavy and data-intensive sub-tasks beneath. In practice, the most useful way to think about such a system is not as a loose collection of prompts, but as a lean operating model with hierarchy, bounded cost, and explicit escalation logic.

From Architecture to Operating Model

In my own framing, the architecture above behaves like a compact, high-performance firm. The human remains the managing director, but only at the edge of the workflow. The day-to-day control plane is Atlas: the executive director and orchestrator that receives work, decomposes it, routes it, checks it, and decides whether the task can be solved inside the base layer or requires escalation.

  • The human managing director is the final escalation point and should handle only strategic, ambiguous, or genuinely unsolved problems.
  • Atlas is the central coordinating brain. It owns task decomposition, delegation, sequencing, parallelisation, output aggregation, quality control, and escalation decisions.

Below Atlas sits a fixed-cost base team. The important point is that this base layer is intentionally not frontier intelligence everywhere. To preserve cost discipline, the always-on system must be good enough for repeatable work, but cheap enough that it can run continuously without turning the economics back into a traditional analyst bench.

  • Researcher is the VP-level intelligence layer. It produces deep analysis, raw insight, and first-pass thinking on complex but well-scoped tasks.
  • Writer is the associate-level layer. It turns research into reports, memos, content, and communication, but can degrade when the assignment becomes too ambiguous, too broad, or too structurally complex.
  • Operator is the analyst or junior layer. It handles simple, repeatable downstream work such as formatting, sending, routing, and workflow execution after the higher-level reasoning has already been done.

This is the critical shift. Atlas is not merely delegating in the mechanical sense of passing work around. It is replacing what would traditionally be a coordinator, project manager, and quality-control function all at once. It breaks large problems into smaller tasks, decides which tasks should run sequentially and which in parallel, recombines the outputs, and tests whether the result meets the original specification.

The boundedness of the base team is a design choice, not a bug. If the task is too complex, too large, or too poorly specified, the fixed layer predictably fails. The failure modes are familiar:

  • The task times out because the reasoning chain is too long or the workflow too sprawling.
  • The agent loops because the objective is underspecified or the decomposition is poor.
  • The output quality collapses because the task sits outside the capability frontier of the base model.

That is why a serious agent operating model needs a contractor layer rather than pretending the base team can do everything. In my framework, Francis is the PhD-level external contractor: the highest-reasoning agent in the stack, invoked only when Atlas detects that the permanent team has reached its limit. Francis is not part of the fixed monthly cost base. It is activated on demand and paid for on a per-use basis, which keeps high-cost intelligence out of routine workflows and deploys it only when complexity justifies it.

The escalation flow is therefore straightforward:

  1. Atlas receives the task and decomposes it into sub-tasks.
  2. Atlas assigns those sub-tasks across the Researcher, Writer, and Operator layers.
  3. Atlas runs those agents sequentially or in parallel, depending on the structure of the work.
  4. Atlas aggregates the outputs, validates quality, and checks whether the result is fit for use.
  5. If an agent times out, loops, or produces weak output, Atlas escalates to Francis.
  6. If Francis still cannot resolve the problem, the work escalates to the human managing director, which is a signal that the task requires strategic judgment or a reframing of the system itself.

This operating model clarifies the economics. Researcher, Writer, and Operator form the predictable fixed-cost layer. Francis is variable cost. The governing principle is simple: keep expensive intelligence out of the base architecture and bring it in only when the marginal value of better reasoning exceeds the marginal cost of invoking it.

It also clarifies the scalability model. A traditional organisation scales by hiring, which creates lag, management overhead, and chronic risk of overcapacity. An agent organisation scales by replication. One research pass can generate ten insights; Atlas can then spin up five Writers against those insights in parallel, consolidate the drafts, and return a finished output without changing the permanent cost base. The structural advantage is the move from fixed hierarchy and limited throughput to an elastic workforce with minimal human intervention between framing and final approval.

Read that way, the model is easy to summarise. Atlas is the operating system. The core agents are the execution layer. Francis is the escalation valve. The human managing director remains the final authority, but not the operator. And that, in turn, is why the human profile at the top matters so much.

The Professional Profile That Makes It Work

The efficiency claim depends entirely on who is sitting at the top of that architecture. The relevant empirical evidence here comes from two directions. Sarkar's 2025 study, analysing an AI coding agent rollout across 1,000 organisations, found experienced workers were 5–6% more likely per standard deviation of experience to accept agent-generated code and integrate it effectively. The BCG study identifies distinct integration modes — including deeply integrated "Cyborg" approaches — and subsequent analysis suggests seniority correlates with more effective adoption of these modes than junior counterparts. The pattern is consistent: agents are most effective in the hands of practitioners who have a sufficiently precise mental model of what good output looks like to calibrate and correct the agent when it drifts.

From my own background in structured credit and consumer lending, the domain expert's primary contribution to agent orchestration is not the ability to write code — it is the ability to write a precise objective. Knowing that a credit monitoring agent should flag a rating migration from B2 to B3 but not just a negative outlook revision, or that an earnings extraction agent should treat segment-level EBITDA differently from group-level reported EBITDA, or that a regulatory tracking agent needs to distinguish between consultation papers and final rules — these distinctions require years of market experience to encode correctly. The senior orchestrator does not replace a junior analyst by working faster; they replace the requirement for a junior analyst by knowing precisely what the agent needs to know to do the work correctly.

Andrew Ng, as cited in Tech Founder Stack, puts this precisely: "10x engineers don't write code 10 times faster. Instead, they make architecture decisions that result in dramatically better downstream impact." The finance equivalent is the professional who designs the workflow, sets the constraints, specifies the verification criteria, and reviews outputs with judgment that only accumulated market experience can provide.

The Strongest Counterargument and Why It Does Not Disqualify the Model

The most credible objection is that AI agents are not reliable enough for investment-grade work without continuous human intervention — and that the intervention cost erodes the productivity gain. This concern is grounded in real evidence. Research by Patronus AI on finance-specific LLM tasks suggests hallucination rates of 15–25% without structural safeguards. The BCG study demonstrated a 19-percentage-point decline in correctness for tasks outside the AI capability frontier. The METR randomised controlled trial, published in July 2025, found that experienced open-source developers in familiar codebases took 19% longer when using AI tools — with researchers documenting a 40-percentage-point gap between self-assessed and actual performance. If the METR dynamic generalises to senior credit analysts in their core sectors, the productivity thesis is meaningfully weakened.

The counterargument is correct in its diagnosis but misdirects its conclusion. The reliability concern is an argument for the governance architecture described above — not against the model itself. The verification layer and human review checkpoint are not optional additions; they are what separates a viable investment-grade workflow from a liability. For structured data extraction tasks — earnings figures from SEC filings, covenant metrics from credit agreements, regulatory deadline tracking — retrieval-augmented generation against auditable structured data sources can achieve accuracy rates above 99% at commercial scale, as Daloopa documents for its structured data pipeline. The METR finding, read carefully, is that AI slows experts on tasks they already know how to do quickly. The orchestrator model is not designed to make a senior analyst faster at their core competence work. It is designed to eliminate the surrounding documentation-heavy and monitoring-intensive workload, so the senior professional focuses entirely on the judgment that agents cannot supply.

The perception gap documented by METR — participants believed AI made them faster even when it did not — is a real operational risk. The mitigation is straightforward: measure throughput empirically on the tasks being automated, and gate agent deployment on demonstrated output quality rather than assumed improvement.

What the Regulatory Environment Permits and Requires

The regulatory landscape for agent-based investment workflows in the UK and EU is more permissive than headlines suggest, but not permissive in ways that remove human accountability. The FCA's principles-based, technology-neutral approach means there are no AI-specific rules — but Senior Managers and Certification Regime accountability applies fully to AI-driven investment decisions. The individual named under SM&CR cannot delegate accountability to an agent; they can only delegate execution. ESMA's May 2024 statement confirms that MiFID II obligations apply without modification to AI use in investment services. Separately, the EU AI Act's Article 12 requires providers of high-risk AI systems to maintain logs covering data sources, decision processes, and all system modifications over time. The EU AI Act classifies creditworthiness assessment as high-risk AI under Annex III, with full compliance obligations — including Article 14 human oversight requirements — effective August 2026; AI-driven investment advice tools face analogous obligations under MiFID II.

Read together, these frameworks are not obstacles to the orchestrator model; they are a description of it. The architecture in which a named senior professional sets objectives, reviews all outputs before consequential use, and maintains auditable records of agent actions is exactly what FCA principles and MiFID II requirements converge on. The EU AI Act's human oversight mandate is the human review checkpoint in the diagram above. The SM&CR accountability requirement is the human orchestrator. Firms treating regulatory compliance as a constraint on AI deployment have misread the frameworks; they describe a deployment model, not a prohibition.

The Honest Limits of the Claim

There is an evidentiary gap that deserves acknowledgement. No published controlled study has compared the output of a single senior AI orchestrator against a conventional analyst team in a production investment management context, held all other variables constant, and measured throughput and decision quality over an extended period. The directional evidence is strong — from BCG's RCT, McKinsey's case data, the institutional deployments at Man Group, Citadel, and JPMorgan, and the talent market bifurcation data — but the specific claim of full analytical equivalence at team scale remains inference from consistent secondary sources rather than primary experiment. I present it as a structurally well-supported thesis with one key evidentiary gap.

The key-person risk concern is also real. An intellectual infrastructure embedded in custom agent workflows, RAG architectures, and prompt engineering frameworks is not easily transferable to successors. The mitigation — documented workflow architecture, explicit succession planning, redundant tool access — reduces concentration risk but does not eliminate it. This is structurally identical to key-person risk at any small GP structure, which the institutional allocator community has long experience pricing.

The implication that flows from the evidence is not that every investment firm should immediately reduce its analyst headcount to one. It is that the firms now competing hardest for senior AI-literate capital markets talent have understood what they are actually hiring: not a person, but a team — assembled from specialised agents, governed by domain expertise, and accountable through the professional judgment that no agent can supply.

Sources

AI Productivity Evidence (Academic Studies)

  1. Dell'Acqua et al., "Navigating the Jagged Technological Frontier," HBS Working Paper 24-013 https://www.hbs.edu/faculty/Pages/item.aspx?num=64700

  2. Dell'Acqua et al., SSRN version of BCG/HBS study https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4573321

  3. METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity" https://letsdatascience.com/blog/developers-thought-ai-made-them-faster-the-data-said-otherwise

  4. Reuters, METR AI developer productivity study summary https://www.reuters.com/business/ai-slows-down-some-experienced-software-developers-study-finds-2025-07-10/

  5. Sarkar / University of Chicago Booth, AI coding agent rollout study (AI Insider) https://theaiinsider.tech/2025/11/17/study-ai-agents-are-quietly-delivering-the-productivity-gains-the-hype-cycle-forgot/

  6. International Center for Law and Economics, "AI Productivity and Labor Markets: A Review of the Empirical Evidence," February 2026 https://laweconcenter.org/resources/ai-productivity-and-labor-markets-a-review-of-the-empirical-evidence/

  7. Brynjolfsson, Li, Raymond, AI customer support productivity study, Quarterly Journal of Economics https://academic.oup.com/qje/article/140/2/889/7990658

  8. Professor KL, "Discovering AI's Jagged Frontier," updated BCG/HBS analysis, March 2026 https://professorkl.substack.com/p/discovering-ais-jagged-frontier-and

  9. arXiv, LLM knowledge of financial data accuracy study, March 2025 https://arxiv.org/html/2504.00042v1

Fund Economics and Fee Compression

  1. The Investment Association, "Investment Management in the UK 2023–2024," Chapter 6 https://www.theia.org/sites/default/files/2024-10/Investment%20Management%20in%20the%20UK%202023-2024%20Chapter%206.pdf

  2. The Investment Association, "Investment Management in the UK 2024–2025" https://www.theia.org/sites/default/files/2025-10/Investment%20Management%20in%20the%20UK%202024-2025_0.pdf

  3. ETF Stream / EFAMA, "Asset Management Profit Margins Fall to Lowest Level Since GFC" https://www.etfstream.com/articles/asset-management-profit-margins-fall-to-lowest-level-since-gfc-efama-finds

  4. Investment Company Institute, Fund Fees Decline Press Release, March 2025 https://www.ici.org/news-release/25-news-fund-fees-decline

  5. McKinsey, "Asset Management 2025: The Great Convergence," September 2025 https://www.mckinsey.com/industries/financial-services/our-insights/asset-management-2025-the-great-convergence

  6. McKinsey, "Beyond the Balance Sheet: North American Asset Management 2024" https://www.mckinsey.com/industries/financial-services/our-insights/beyond-the-balance-sheet-north-american-asset-management-2024

Compensation and Labour Market Data

  1. Selby Jennings, 2024 Global Investment Management Compensation Guide https://hub.selbyjennings.com/hubfs/Selby%20Jennings%202024/North%20America/Selby-Jennings-Investment-Management-Compensation-Guide-Global.pdf

  2. CFA Institute, 2024 Compensation Study https://www.scribd.com/document/816673019/CFA-2024-Compensation-Study-1

  3. U.S. Bureau of Labor Statistics, Employer Costs for Employee Compensation https://www.bls.gov/news.release/pdf/ecec.pdf

  4. U.S. Bureau of Labor Statistics, Occupational Outlook Handbook: Financial Analysts https://www.bls.gov/ooh/business-and-financial/financial-analysts.htm

  5. Forbes, "The Finance Talent Arbitrage: Why Entry-Level Jobs Are Disappearing," March 2026 https://www.forbes.com/sites/jonmarkman/2026/03/06/the-finance-talent-arbitrage-why-entry-level-jobs-are-disappearing/

  6. Hunt Scanlon, "AI's Growing Threat to Entry-Level Finance Hiring" https://huntscanlon.com/ais-growing-threat-to-entry-level-finance-hiring/

  7. The Digital Banker, "AI Is Coming for Wall Street: Banks Are Reportedly Weighing Cutting Analyst Hiring by Two-Thirds" https://thedigitalbanker.com/ai-is-coming-for-wall-street-banks-are-reportedly-weighing-cutting-analyst-hiring-by-two-thirds/

Agent Orchestration Frameworks

  1. McKinsey, "Seizing the Agentic AI Advantage," June 2025 https://www.mckinsey.com/capabilities/quantumblack/our-insights/seizing-the-agentic-ai-advantage

  2. PwC, "AI Agents for Finance" https://www.pwc.com/us/en/tech-effect/ai-analytics/ai-agents-for-finance.html

  3. Alvarez & Marsal, "Demystifying AI Agents in 2025" https://www.alvarezandmarsal.com/thought-leadership/demystifying-ai-agents-in-2025-separating-hype-from-reality-and-navigating-market-outlook

  4. The Data Score, "Can AI Match Sell-Side Analysts? Testing OpenAI Deep Research" https://thedatascore.substack.com/p/can-ai-match-sell-side-analysts-testing

  5. Microsoft, "AI Transformation in Financial Services: 5 Predictors for Success in 2026," December 2025 https://www.microsoft.com/en-us/industry/blog/financial-services/2025/12/18/ai-transformation-in-financial-services-5-predictors-for-success-in-2026/

Institutional AI Adoption (Case Studies)

  1. Morgan Stanley, "AskResearchGPT" press release, October 2024 https://www.morganstanley.com/press-releases/morgan-stanley-research-announces-askresearchgpt

  2. The AI Insider, "Morgan Stanley Brings OpenAI-Powered ChatGPT Tools to Investment Banking and Trading," October 2024 https://theaiinsider.tech/2024/10/24/morgan-stanley-brings-openai-powered-chatgpt-tools-to-investment-banking-trading/

  3. JPMorgan Chase, 2025 Annual Report — CEO Letter to Shareholders, April 2026 https://www.jpmorganchase.com/ir/annual-report/2025/ar-ceo-letters

  4. Best Practice AI, "JPMorgan COiN: 360,000 Annual Lawyer-Hours Saved" https://www.bestpractice.ai/ai-case-study-best-practice/jpmorgan_reduced_lawyers'_hours_by_360,000_annually_by_automating_loan_agreement_analysis_with_machine_learning_software_coin

  5. Bloomberg, "JPMorgan Marshals an Army of Developers to Automate High Finance," February 2017 https://www.bloomberg.com/news/articles/2017-02-28/jpmorgan-marshals-an-army-of-developers-to-automate-high-finance

  6. Reuters, "Citadel Debuts New AI Tool for Equities Investors," December 2025 https://www.reuters.com/business/citadel-debuts-new-ai-tool-equities-investors-cto-subramanian-says-2025-12-03/

  7. AI Street, "Inside Citadel's AI Assistant," December 2025 https://www.ai-street.co/p/citadel-reveals-ai-assistant

  8. Bloomberg, "Man Group Says Agentic AI Is Now Devising Quant Trading Signals," July 2025 https://www.bloomberg.com/news/articles/2025-07-10/man-group-says-agentic-ai-is-now-devising-quant-trading-signals

  9. AI Street, "Inside Man Group's AlphaGPT," December 2025 https://www.ai-street.co/p/inside-man-group-s-alphagpt

  10. Reuters, "Citigroup's AI Usage Frees Up 100,000 Hours of Developers per Week," October 2025 https://www.reuters.com/business/citigroups-ai-usage-frees-up-100000-hours-developers-week-2025-10-14/

  11. Sify, "The Dawn of Hedge Agents: How Agentic AI Is Transforming Hedge Fund Operations," April 2026 https://www.sify.com/ai-analytics/the-dawn-of-hedge-agents-how-agentic-ai-is-transforming-hedge-fund-operations/

Industry Surveys and Adoption Data

  1. EY-Parthenon, "Generative AI in Wealth and Asset Management" survey, 2024 https://www.ey.com/content/dam/ey-unified-site/ey-com/en-gl/industries/wealth-asset-management/documents/ey-gl-genai-wam-survey-highlights-03-2024.pdf

  2. EY, Gen AI in Wealth and Asset Management Survey, 2025 https://www.ey.com/en_us/insights/wealth-asset-management/gen-ai-in-wealth-asset-management-survey

  3. IIF / EY, 2024 Annual Survey Report on AI/ML Use in Financial Services https://www.iif.com/portals/0/Files/content/Innovation/2024%20IIF-EY%20Survey%20Report%20on%20AI_ML%20Use%20in%20Financial%20Services_Public%2001.08.25.pdf

  4. AIMA, "Getting in Pole Position: How Hedge Funds Are Leveraging Gen AI to Get Ahead" https://www.aima.org/article/press-release-getting-in-pole-position-how-hedge-funds-are-leveraging-gen-ai-to-get-ahead.html

Regulatory and Governance

  1. Financial Conduct Authority, "AI and the FCA: Our Approach," September 2025 https://www.fca.org.uk/firms/innovation/ai-approach

  2. A-Team Insight, "FCA AI Update 2025: How the Regulator Is Embedding AI Oversight into UK Financial Rules" https://a-teaminsight.com/blog/fca-ai-update-2025-how-the-regulator-is-embedding-ai-oversight-into-uk-financial-rules/

  3. ESMA, "ESMA Provides Guidance for Firms Using Artificial Intelligence in Investment Services," May 2024 https://www.esma.europa.eu/press-news/esma-news/esma-provides-guidance-firms-using-artificial-intelligence-investment-services

  4. Regulation Tomorrow, "ESMA Issues Initial Guidance for Firms Using AI in Investment Services," May 2024 https://www.regulationtomorrow.com/2024/05/esma-issues-initial-guidance-for-firms-using-ai-in-investment-services/

  5. Goodwin Law, "Key Points for Financial Services Businesses Under the EU AI Act," August 2024 https://www.goodwinlaw.com/en/insights/publications/2024/08/alerts-practices-pif-key-points-for-financial-services-businesses

  6. Freshfields, "Navigating the New Regulatory Momentum: AI in UK Financial Services" https://riskandcompliance.freshfields.com/post/102lo9s/navigating-the-new-regulatory-momentum-ai-in-uk-financial-services

  7. NIST, Generative AI Profile (AI RMF 1.0 Companion) https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf

Historical Parallels

  1. Tech Founder Stack, "Rethinking the 10x Engineer," September 2025 https://www.techfounderstack.com/p/rethinking-the-10x-engineer-from

  2. CB Insights, "Bloomberg Terminal Disruption" https://www.cbinsights.com/research/report/bloomberg-terminal-disruption/

  3. Acquired.fm, "Renaissance Technologies" episode https://www.acquired.fm/episodes/renaissance-technologies

Hallucination Rates and Reliability Evidence

  1. Ankur's Newsletter, "Unveiling the Challenges: Why Large Language Models Struggle with Financial Data" (Patronus AI FinanceBench study) https://www.ankursnewsletter.com/p/unveiling-the-challenges-why-large

  2. Four Dots, "Business Impact of AI Hallucinations: Rates and Ranks" https://fourdots.com/business-impact-of-ai-hallucinations-rates-and-ranks

  3. Daloopa, "Pros and Cons of Using LLMs for Financial Analysis" https://daloopa.com/blog/analyst-best-practices/pros-and-cons-of-using-llms-for-financial-analysis