AI and ML
Understanding agents, their five controllers and one graph
For decades, investment research was a manual grind. An analyst’s worth was judged by how many hours they spent sifting through dense financial filings to find a single missing detail. Modern AI can compress parts of that document-review process from hours or days into minutes.
However, a major trap has emerged across the industry: speed is not alpha.
Summarizing filings faster does not give you an edge. It simply gets you to the market consensus at the same time as everyone else. When every firm uses AI to process information instantly, speed becomes basic table stakes, not outperformance.
To build a real, defensible investment edge, financial institutions must look past generic AI chatbots and focus on building enterprise-grade AI platforms. True alpha comes from engineering systems that solve three fundamental problems:
Clean the data plumbing: Fix fragmented data feeds, compute missing metrics accurately and stop bad data before it reaches portfolio models.
Challenge the consensus: Use AI agents to actively red-team investment ideas, stress-test assumptions and spot where the market narrative is wrong.
Protect the trade: Link thesis generation directly with execution inputs (like market impact, liquidity limits and portfolio rebalancing parameters), ensuring theoretical signal gains aren't eroded by real-world friction.
The future of investment research isn't about summarizing text faster. It's about building the platform architecture needed to turn raw data into sustainable outperformance.
The three pillars of modern factor research
Modern investment research expands far beyond traditional price signals. By pairing domain-specific natural language processing (NLP) with agentic workflows, investment teams can track three distinct dimensions of momentum simultaneously:
Price momentum: The traditional quantitative baseline, tracking statistical price trends and historical return persistence over defined time horizons.
Fundamental momentum: Tracking the direction and acceleration of underlying business health, such as earnings revisions, operating margin trajectories and balance sheet stress.
Narrative momentum: Using specialized financial NLP pipelines and goal-oriented agentic AI to evaluate the degree of positivity, negativity or structural shift in an issuer's public coverage.
Agentic pipelines can orchestrate the collection and analysis of unstructured text, evaluate it in context and translate narrative signals into quantitative features that can then be validated as potential factors.
None of these signals should be treated as alpha simply because an AI model can generate them. They must first survive out-of-sample testing, controls for look-ahead and survivorship bias, regime changes, turnover and transaction costs.
Related: Evaluating AI agents in production: A practical framework
The efficiency paradox: Why speed is not alpha
Summarizing a 100-page filing in 30 seconds does not create an alpha-generating edge. As access to the same AI capabilities becomes widespread, the advantage from processing consensus information faster is likely to compress. The differentiator shifts toward proprietary data, better models, better interpretation and better execution.
To convert research productivity into true outperformance, institutional frameworks must focus on two higher-order capabilities:
Counter-consensus hypothesis generation: Alpha does not come from summarizing consensus narratives; it comes from identifying where the market narrative is wrong. Advanced AI systems must be configured as red-teaming agents designed to actively challenge consensus assumptions, stress-test management guidance against supply chain metrics and flag mispriced risk.
Execution & signal decay mitigation: Paper alpha frequently vanishes in market execution. High-performing quantitative engines pair idea generation directly with portfolio sizing, trade timing and Transaction Cost Analysis (TCA) to ensure the theoretical edge survives actual market impact.
Navigating the real-world hurdles
This new workflow is powerful, but implementing it isn't without challenges.
The "black box" problem: Many advanced AI models can be difficult to interpret, which is a roadblock for regulatory approval and risk management.
Cultural resistance and talent gap: Seasoned professionals may face pushback or skepticism, and there is a scarcity of talent combining financial expertise with AI skills.
Legacy technology: Connecting modern AI to decades-old financial systems is complex and often stalls projects due to integration challenges.
Regulatory uncertainty: Regulation is evolving, but firms do not operate in a regulatory vacuum: existing obligations around supervision, recordkeeping, communications, privacy, fiduciary duties and model governance already apply to AI-enabled workflows.
Data hygiene as a prerequisite: Fixing the "patchwork" trap
Sophisticated AI models and factor algorithms are useless if the underlying data architecture is a fragmented, unverified patchwork.
Consider a common institutional failure mode: a firm attempts to build an advanced portfolio management system, but relies on fragmented data feeds, unverified statistical inputs (such as missing volatility metrics or factor betas) and unreviewed corporate action adjustments. When foundational data feeds break or lack oversight, downstream portfolio models produce flawed risk signals.
AI can help address this structural bottleneck by acting as an automated data steward across three critical layers:
Autonomous schema mapping: AI-assisted schema mapping can identify field correspondences, detect schema changes and propose transformations into a common internal model, reducing (but not eliminating) the need for manually maintained ETL logic.
Deterministic math engines: When key statistical parameters are missing from a feed, the system does not guess or hallucinate numbers. An agent can call governed calculation services or generate code that is executed in a controlled environment and validated against approved formulas and test cases.
Continuous anomaly auditing: Using statistical anomaly detection, business-rule validation and cross-source reconciliation, AI auditors continuously scan incoming feeds for data drift, unadjusted stock splits or extreme outliers before flawed data reaches the portfolio engine.
Context engineering & citation-grade RAG
To move from generic internet knowledge to institutional precision, AI systems must be constrained by strict context engineering. For document-grounded AI workflows, a key control is RAG with traceable source attribution and verification.
An analyst cannot put their name on unverified AI output. If an AI summary states that "Leverage is capped at 4.5x," the system must provide an exact, clickable audit trail pointing directly to Page 42, Paragraph 3 of Loan_Agreement.pdf.
By defining explicit analyst personas, negative constraints and few-shot formatting examples, context engineering can reduce unsupported outputs and lower the analyst's verification burden.
The workflow:
Step 1: INPUT
Analyst Query: "What are the covenants?"
+
System Retrieval: [Finds Page 42 of Loan_Agreement.pdf]
Step 2: CONTEXT PACKET
The Prompt sent to AI = {
"Role": "Credit Analyst",
"Task": "Answer the query using ONLY the attached text.",
"Attached Text": "Page 42 content...",
"Query": "What are the covenants?"
}
Step 3: OUTPUT
AI Answer: "According to the attached text, leverage is capped at 4.5x."'
Define the task, evidence and constraints
General models try to be helpful to everyone, which often leads to vague outputs. You must narrow their scope.
The prompt: Start by explicitly defining task, evidence, definitions, output schema, permitted sources and decision constraints.
The constraint: Add negative constraints to reduce noise: "Do not summarize the business overview. Only extract financial covenants that are at risk of breach."
Teach by example ("Few-shot" learning)
When you need the AI to structure data in a specific format, showing is better than telling.
Method: Instead of writing complex instructions, provide two or three examples of the exact input and output format you want (e.g., "Deal Name | Interest Rate | Maturity").
Benefit: This gives the model concrete examples of the firm's preferred reporting style, increasing the consistency of its outputs.
Taken together, these components form a practical workflow: start with trusted data, generate and test differentiated signals, ground AI outputs in traceable context, then carry validated insights through portfolio construction and execution. The playbook is less about deploying one model and more about building a repeatable system that moves from raw information to investable decisions.
The smart adopter's playbook
For any firm looking to gain an AI advantage, the path forward doesn't require a billion-dollar budget.
Leverage commercial tools first: Master the AI embedded in existing platforms for quick wins in efficiency.
Build selectively on open models: Don't assume the answer is always a generic commercial chatbot. Depending on the use case, consider open models, retrieval from proprietary data, task-specific fine-tuning or models designed specifically for financial problems.
Train people to ask better questions: Invest in training analysts to formulate insightful queries and challenge AI outputs.
Use the right tool for the job:
Classification: Use for sorting data, such as flagging suspicious AML transactions or risk levels.
Regression: Use for predicting numbers, such as revenue forecasts or expected credit losses.
Commercial ROI: Measuring operational & financial value
Deploying institutional AI requires clear operational justification. Recent industry data and market benchmarks highlight the hard commercial value of structured AI adoption vs. the pitfalls of isolated pilots:
Commercial bottleneck | Traditional reality | Institutional AI solution | Metric to measure |
Verification overhead | Analysts re-read filings to verify AI text | Citation-grade RAG: sentence-level mapping to source documents. | Analyst verification time per memo |
Computational cost bloat | Unmonitored API calls and oversized model selection. | Model benchmarking & token logging: Dynamic routing based on speed/cost/variance. | Cost per research task |
Institutional knowledge loss | Analyst turnover wipes out historical deal context. | Private data layer & vector stores: Indexing internal research into enterprise memory. | Time to proficiency for new analysts |
Coverage capacity constraints | Human teams capped at ~150 issuers per analyst. | Parallel batch execution: Multi-agent universe scanning across thousands of issuers. | Issuers screened per analyst |
Conclusion: The new formula for alpha
AI can make investment research faster, broader and more systematic, but efficiency alone is not alpha. A faster summary, a cleaner forecast or a new factor score is only a starting point. The real test is whether the resulting insight is differentiated, survives rigorous validation and still holds after portfolio constraints, transaction costs and market impact.
This creates two potential sources of advantage. Insight Alpha comes from identifying signals, relationships or outcomes that the market may be mispricing, then validating that those signals hold up out of sample and in real portfolio conditions. Synthesis Alpha comes from the analyst’s ability to combine those signals with proprietary context, historical perspective, contradictory evidence and market judgment to form an investment thesis others have not reached.
AI can strengthen both, but it does not create either automatically. The durable edge comes from the system around the model: reliable data, traceable evidence, disciplined evaluation, execution-aware portfolio construction and human judgment at the point of decision. That is the playbook: a repeatable workflow that moves from trusted data, to differentiated and validated insight, to portfolio decisions and execution.
The new formula for alpha is therefore not simply faster analysis. It is better insight, better synthesis and better execution, amplified by AI.
Disclaimer: The statements and opinions expressed in this article are those of the author(s) and do not necessarily reflect the positions of Thoughtworks.
