Source ledger: Finance firm strategy upstream refresh
Related local coverage: HRT token-burn analysis, buy-side practitioner signals, hedge-fund AI capability audit, quant AI infrastructure moats, and BCG asset-management report.
Source status: First-party trading-firm and asset-manager pages, consulting research, regulator material, and an academic evidence audit reviewed through 2026-08-16.
Confidence: MEDIUM-HIGH for the operating-model pattern; MEDIUM for firm-reported scale; LOW for any inference about incremental alpha or P&L.
The answer in one sentence
The public record supports investing in a governed research-and-engineering system, not buying the idea of an autonomous trading agent: the durable edge is the loop from point-in-time data to hypothesis, simulation, execution, feedback, and monitoring.
Three layers of finance AI
1. Research intelligence: expand coverage, keep the investment gate human
Bridgewater’s PAT combines codified investment knowledge, LLMs, agentic workflows, and software architecture for research. T. Rowe Price describes AI-assisted systematic research on qualitative information, paired with quantitative analysis and fundamental analysts. The local Balyasny, Man Group, Two Sigma, CFM, Citadel, and Jane Street material shows a similar pattern: AI broadens what researchers can read, test, summarize, and code; the portfolio or trading decision still passes through explicit human, model, risk, or governance gates.
This is the strongest near-term hedge-fund use case because the output is a verifiable research artifact: a sourced observation, cleaned data set, feature, hypothesis, code change, backtest, or risk memo. It can be rejected without sending an order to market.
2. Trading and market making: models are core, general LLM autonomy is not
Optiver provides the clearest negative evidence. Its public test found recent LLMs near intern-level performance on theory and basic trading tasks, but weak on sequential belief updates, expected-value maximization, adverse selection, action consistency, modeling other market participants, and game-theoretic reasoning. Those are not cosmetic gaps in a market-making environment; they are the job.
Jump and IMC show what the production stack actually looks like: large-scale simulation, low-latency or real-time inference, specialized ML/RL/deep-learning models, high-performance data and compute, rapid researcher-to-trader feedback, and continuous validation. LLM agents appear as one tool layer for research, coding, NLP, and workflow assistance. They are not presented as a substitute for the market-specific models, simulators, execution controls, or trader feedback loop.
3. Firm operations: broad agents, but bounded authority
The new OpenAI, Anthropic, Google Cloud, Microsoft, and consulting delivery channels are most relevant to firm-wide workflows: knowledge access, software engineering, operations, compliance preparation, surveillance, client service, and internal automation. The Bank of England’s July 2026 stability assessment independently describes trading firms using more autonomous AI mainly in research, coding support, surveillance, and lower-risk operations rather than fully autonomous trading.
That convergence matters. It says the firm-wide platform should be broad, while the live trading authority layer should remain narrow, explicitly permissioned, and separately evaluated.
What differs and contradicts
| Question | Consulting/vendor narrative | Trading-firm/regulator/academic evidence | Strategy conclusion |
|---|---|---|---|
| How much can be automated? | BCG projects substantial capacity unlocks, including 70–80% of standard trade-execution flow and possible Sharpe improvement. | Optiver exposes failures in sequential/adversarial trading reasoning; BoE says autonomy is concentrated outside fully autonomous trading; an academic audit finds weak time-split and cost-model discipline. | Use high automation for bounded, observable workflow steps. Treat live decision autonomy as a separate proof burden. |
| What creates the edge? | More agents, broader coverage, and agentic workflow redesign. | Jump, IMC, Optiver, Jane Street, and the local corpus emphasize data, simulation, specialized models, infrastructure, and researcher/trader feedback. | Build the closed-loop research and execution system. Do not confuse generic agent volume with trading edge. |
| What is the maturity baseline? | BCG presents a large AI-first opportunity and projects aggressive upside. | Mercer finds 55% of asset managers integrated AI in at least one process, 27% piloting, and 18% not integrated; 69% cite data constraints and most have very small dedicated AI teams. | Sequence around data and evaluation readiness. A small number of production-quality workflows beats a large ungoverned catalog. |
| Should the stack be centralized? | Frontier labs and cloud/consulting alliances favor a platform-centered channel. | Trading firms maintain specialized research, execution, simulation, and risk systems; Caylent/NVIDIA material favors federated placement and workload-specific architecture. | Centralize identity, provenance, evaluation, and telemetry; federate model and compute placement. |
| Does a benchmark prove value? | Vendor cases and consulting projections supply directional ROI narratives. | Regulators and academic reviews demand validation under changing markets, transaction costs, and temporal splits. | No alpha or ROI claim enters the investment case without a point-in-time counterfactual and after-cost measurement. |
The strategic architecture
The right architecture is a dual stack with a shared evidence spine:
Shared evidence spine
point-in-time data · lineage · evals · permissions · telemetry
/ \
Research-intelligence stack Trading-performance stack
retrieval · agents · code features · simulators · execution
hypothesis · review · backtest latency · TCA · risk · live monitoring
\ /
Firm control plane
identity · policy · approval · rollback · cost
The research stack can use frontier models, small models, retrieval, and agents to produce artifacts. The trading stack should use the model class that wins the specific latency, calibration, stability, and adversarial test, including classical statistical models, gradient boosting, deep learning, RL, or specialized hardware. The shared spine prevents research acceleration from becoming an untraceable path to live risk.
The operating metrics that matter
Token burn per employee is a useful cost signal, but it is not a firm-strategy metric by itself. Track the following by desk, workflow, model, and authority level:
| Metric | Why it matters |
|---|---|
| Cost per verified research artifact | Converts token/GPU spend into a reviewable output |
| Time from hypothesis to reproducible backtest | Measures research-cycle compression without claiming alpha |
| Artifact acceptance and promotion rate | Separates useful work from fluent noise |
| Human review minutes per accepted artifact | Shows whether automation moves or merely relocates labor |
| False-positive, leakage, and stale-data rate | Protects the research process from attractive but invalid signals |
| Backtest-to-paper and paper-to-live survival | Tests whether the research signal survives each control gate |
| Live drift, decay, turnover, and after-cost TCA | Measures production behavior rather than demo quality |
| GPU, token, and tool-call cost per accepted signal | Makes inference economics comparable with compute and data costs |
| Rollback/interrupt success rate | Tests whether authority controls work under stress |
For the specific token-burn question, the current public record can support a workload budget and an attribution model. It cannot support a universal “tokens per employee” benchmark for hedge funds because desk mix, model routing, context length, tool calls, private inference, and bursty research workloads differ materially.
Recommended firm sequence
- Separate the budgets. Fund research intelligence, trading-performance infrastructure, and firm operations as different portfolios with different acceptance criteria.
- Start with reversible research authority. Allow read, retrieve, summarize, code-draft, and backtest-submit actions first. Add paper trading only after temporal, leakage, and adversarial tests pass.
- Build the evidence spine before scaling agents. Preserve raw inputs, point-in-time versions, lineage, prompts, tool calls, code, evaluations, approvals, and live monitoring.
- Keep desk ownership. A central platform team should provide identity, routing, evaluation, observability, and cost accounting. Desk teams should own hypotheses, domain constraints, and promotion decisions.
- Make the consulting partner document the counterfactual. Every proposed workflow should state baseline time/cost/error, expected mechanism, measurement window, rollback condition, and what evidence would stop the work.
High-value questions for the McKinsey meeting
- Which recommendations are specific to a market-making or systematic-investment workflow rather than generic enterprise agent adoption?
- What public or private evidence supports the claimed impact after data, review, infrastructure, and control costs?
- How do you separate research coverage gains from genuine incremental alpha?
- What is the proposed authority ladder from read-only research to paper trading to live deployment, and who can interrupt it?
- How will the firm preserve point-in-time data, temporal splits, transaction costs, and reproducible model versions?
- What remains portable if the frontier model, cloud, or implementation partner changes?
- Which workflows should not be automated because the cost of a wrong action is nonlinear or difficult to observe?
Bottom line
The major players do not actually disagree on the value of AI-assisted research and workflow automation. They disagree on how much autonomy and standardization buyers should accept before the evidence is in. Consultants and frontier labs describe a large opportunity and a rapidly forming delivery channel. HFT and prop-trading firms expose the harder truth: live edge comes from specialized models, market structure knowledge, simulation, data quality, latency, feedback, and controls.
The firm strategy should therefore be centralized evidence and governance, federated model choice, desk-owned hypotheses, and a much higher proof bar for live trading authority than for research assistance.
Evidence boundaries
Trading-firm pages are first-party capability disclosures, not audited performance reports. BCG and EY projections are consulting estimates. Mercer is a named survey but still self-reported. Vendor and partner cases are selected examples. The BoE and AMF provide independent oversight context, while the agentic-trading paper is an academic evidence-quality audit rather than a universal negative result. No source in this refresh establishes incremental, capacity-adjusted, after-cost alpha from a frontier-model agent.