CFA Institute · Published 2026-09-01

How Agentic AI Is Reshaping Finance Workflows, Processes, and Governance | CFA Institute

privateReview: unreviewedConfidence: high
agentic aifinance workflowssynthetic dataai governancelarge language modelsagentic-aiai-governancefinance-workflowslocal-llmsmodel-context-protocolred-teamingskill-filessynthetic-data
Open on YouTube ↗

Summary

Overview

  • Speaker: Rhodri Preeces, Brian Pisaneschi, James Tait
  • Channel: CFA Institute
  • Main topic: Agentic AI in Finance
  • Purpose: To provide investment professionals and researchers with insights into the tools, data strategies, governance challenges, and ethical considerations surrounding agentic AI adoption in financial organizations. A roundtable discussion hosted by the CFA Institute exploring how agentic AI and large language models are reshaping workflows, investment processes, data handling, and governance in finance.

Topic Map

Agentic AI and Skills in Finance Workflows

  • Explanation: Discussion on how agentic AI workflows, particularly master prompts or skill files, enable semi-structured task automation without requiring deep programming skills.
  • Key claims:
    • Skills/master prompts provide sound structure for autonomous workflows.
    • They allow non-technical practitioners to leverage AI for repetitive tasks.
    • Advanced users can connect MCP (Model Context Protocol) servers for native API integration with outside data.
  • Examples:
    • Using markdown files to outline specific financial workflows.
    • Pre-populating Excel models inside skill files.
  • Terminology:
    • agentic AI
    • skills
    • master prompts
    • markdown files
    • MCP servers
  • Why it matters: Lowers the technical barrier to adopting advanced automation in investment management while allowing iteration and standardization across firms.

Data Strategy: Imputation, Missing Data, and Synthetic Data

  • Explanation: Examining data limitations in finance and how machine learning methods like nearest neighbors imputation and generative AI models address data gaps.
  • Key claims:
    • AI models are only as good as the data they access.
    • Traditional linear interpolation and KNN imputation handle basic missing data.
    • Generative AI models trained on historical time series can generate synthetic data for scenario simulation, stress testing, and portfolio backtesting.
  • Examples:
    • Using generative AI to simulate portfolio returns under different inflation and GDP scenarios.
    • KNN imputation for macroeconomic datasets.
  • Terminology:
    • linear interpolation
    • KNN imputation
    • synthetic data
    • generative AI
    • scenario simulation
    • backtesting
  • Why it matters: Empirical models have limits; generative models offer greater flexibility and non-linear relationship modeling without restrictive distributional assumptions.

Open Source vs. Proprietary Models and Cost Optimization

  • Explanation: Comparing open source large language models against proprietary enterprise offerings regarding data privacy, capability, and API costs.
  • Key claims:
    • Proprietary models (e.g., OpenAI) offer high capability but raise data privacy and cost concerns when making frequent API calls.
    • Running smaller open-source language models locally can achieve cost optimization and protect proprietary data.
    • Infrastructure investment is required to run open source models effectively at scale.
  • Examples:
    • Running local 36B parameter models to avoid per-call API fees.
    • Proprietary architectures like Quadcode prompting open-source alternatives.
  • Terminology:
    • open source models
    • proprietary models
    • API calls
    • cost optimization
    • scalability
  • Why it matters: Firms must balance data confidentiality, performance capability, and infrastructure costs when selecting AI tooling.

AI Governance, Fiduciary Responsibility, and Biases

  • Explanation: Addressing the ethical considerations, loss aversion biases, and regulatory compliance required when deploying AI in investment decision-making.
  • Key claims:
    • LLMs exhibit human-like biases, including loss aversion bias.
    • Firms must uphold fiduciary obligations and client data confidentiality.
    • Red-teaming AI models for bias and setting up rigorous evaluation frameworks are necessary to build trust and control.
  • Examples:
    • LLM-driven screening favoring certain stock ratios over others due to training data skew.
    • Red-teaming models to mitigate systematic investment errors.
  • Terminology:
    • fiduciary responsibility
    • loss aversion bias
    • red-teaming
    • governance frameworks
    • client data confidentiality
  • Why it matters: Ensures that AI-driven investment recommendations comply with professional standards and do not amplify harmful human or training biases.

Key Points

Shift from manual multi-terminal analysis to centralized AI orchestration

  • Explanation: Traditionally, analysts switched between Bloomberg, Excel, and various databases; AI tools now sit above these systems to orchestrate queries and workflows.
  • Evidence: Discussion on practitioners shifting from manual terminal navigation to unified skill-based prompts.
  • Practical implication: Reduces task execution time and streamlines investment research workflows.

Cost optimization through local small language models

  • Explanation: Instead of costly proprietary API calls for large datasets, running smaller open-source models locally saves expenses while maintaining privacy.
  • Evidence: Mention of training small language models on proprietary data sets for local execution.
  • Practical implication: Makes advanced AI workflows economically viable for mid-sized financial firms.

Mitigating human and model biases via structured red-teaming

  • Explanation: AI models mirror human-generated data biases; systematic red-teaming and prompt controls are essential to correct them.
  • Evidence: Comparison of LLM loss aversion bias to investor behavioral biases.
  • Practical implication: Protects portfolios from automated systemic errors and improves output reliability.

Frameworks, Models & Processes

Skill-Based Agentic Workflow Framework

  • How it works: Utilizes semi-structured markdown prompt files containing workflow steps, optional Excel models, and external API connectors to execute repetitive tasks autonomously.
  • Components:
    • Master prompts
    • Markdown instructions
    • Excel file templates
    • MCP API connectors
  • When to use: For automating repetitive investment analysis, valuation modeling, and data extraction tasks.

Examples & Case Studies

An analyst uses linear interpolation and KNN imputation to fill missing macroeconomic and return data points.

  • Illustrates: Basic data gap mitigation techniques in financial modeling.
  • Lesson: Ensures datasets are complete enough for downstream quantitative analysis.

A firm runs a small language model locally on historical time series data to simulate portfolio stress tests under shifting inflation and GDP metrics.

  • Illustrates: Empirical vs. generative simulation capabilities.
  • Lesson: Generative models capture non-linear relationships without rigid distributional assumptions.

Actionable Takeaways

  • Immediate:
    • Standardize repetitive workflows using skill files and master prompts.
    • Assess internal data privacy constraints before adopting proprietary cloud AI tools.
  • Strategic:
    • Incorporate synthetic data generation for robust portfolio stress testing and scenario analysis.
    • Establish internal AI governance and red-teaming protocols to monitor model bias and uphold fiduciary duties.
  • Questions to investigate:
    • How do open-source and proprietary models compare on complex multi-step reasoning tasks over time?
    • What liability frameworks apply when an AI-driven investment decision results in significant losses?

Claims Worth Verifying

  • LLMs exhibit loss aversion bias mirroring human investor behavior. (empirical observation)
  • Running smaller open-source language models locally can substitute for proprietary cloud APIs at a fraction of the cost. (technical claim)

Notable Quotes

"we can really create incredibly robust workflows just by our own knowledge and iteration." (at 0:00) "all they are are just master prompts." "the biases in these large language models basically mirror investor biases."

Compressed Summary

  • Agentic AI workflows use master prompts and skill files to streamline financial analysis.
  • Synthetic data generation improves portfolio scenario simulation and stress testing.
  • Open-source models offer cost efficiency and data privacy compared to proprietary APIs.
  • Fiduciary duty requires robust red-teaming to manage LLM biases and errors.
  • Keywords: agentic ai, finance workflows, synthetic data, ai governance, large language models
  • Core insight: Agentic AI transforms finance workflows by combining prompt-driven automation and synthetic data, requiring rigorous governance to manage model biases and uphold fiduciary standards.

Core insights

5
Mechanismmedium noveltymoderate evidence

In agentic finance systems, a recurring task can be packaged as a 'skill' or 'master prompt'—a markdown file that states workflow steps, can carry prepopulated Excel model templates, and can include MCP server connectors—instead of being written as code. This makes workflow automation a structured text artifact that non-programmers can author and review and that can be standardized across a firm.

Why it matters

It changes the agent architecture: the unit of automation is an editable, versionable prompt-plus-template artifact, and the runtime is a skill loader/executor with tool connectors rather than bespoke integration code. Governance can review markdown diffs rather than application code.

Generalization

Any domain with repetitive expert workflows can use declarative 'skill files' that mix natural-language instructions, embedded templates, and declared external tools as the canonical automation interface.

Skills/master prompts provide sound structure for autonomous workflows.
Open source video
They allow non-technical practitioners to leverage AI for repetitive tasks.
Open source video
Using markdown files to outline specific financial workflows.
Open source video
Pre-populating Excel models inside skill files.
Open source video
Advanced users can connect MCP (Model Context Protocol) servers for native API integration with outside data.
Open source video
Failure Modemedium noveltymoderate evidence

LLMs trained on human-generated data reproduce human cognitive biases—notably loss aversion—and this can create systematic errors in autonomous stock screening or investment workflows. Finance-specific agent deployments therefore need red-team testing and evaluation frameworks that treat behavioral biases as a release-blocking failure mode, not an abstraction to be handled after deployment.

Why it matters

Autonomous agents operating under fiduciary duties can convert subtle training-data skew into repeated, systemic portfolio errors. Because there may be no human in the loop, bias must be a first-class evaluation gate in the agent development process rather than an ethical afterthought.

Generalization

For LLM-based agents in consequential decision-making, bias evals need to test the same systematic cognitive biases humans exhibit (e.g., loss aversion), not just generic safety/toxicity categories.

LLMs exhibit human-like biases, including loss aversion bias.
Open source video
LLM-driven screening favoring certain stock ratios over others due to training data skew.
Open source video
Red-teaming models to mitigate systematic investment errors.
Open source video
Tradeoffmedium noveltymoderate evidence

For an agentic workflow that makes frequent LLM calls on data-heavy, repetitive finance tasks, backend model selection is an economics and data-privacy decision: proprietary models offer high capability but incur per-call cost and client-data exposure, while smaller locally deployed open-source models (e.g., a 36B-parameter model) can reduce cost and protect proprietary data, at the price of infrastructure investment.

Why it matters

Agent engineers should treat model routing and deployment tier as first-class architecture, splitting high-frequency tasks to local small models and complex low-frequency reasoning to larger cloud models. Marginal-cost optimization and confidentiality can drive better architecture than defaulting every agent call to one large API.

Generalization

Any high-volume, data-sensitive agent deployment should consider a portfolio of local/open and cloud/API models rather than one default provider, with routing based on task complexity, sensitivity, and cost.

Proprietary models (e.g., OpenAI) offer high capability but raise data privacy and cost concerns when making frequent API calls.
Open source video
Running smaller open-source language models locally can achieve cost optimization and protect proprietary data.
Open source video
Infrastructure investment is required to run open source models effectively at scale.
Open source video
Running local 36B parameter models to avoid per-call API fees.
Open source video
Mechanismmedium noveltymoderate evidence

Generative AI can function as an environment simulator rather than only as a chatbot: trained on historical financial time series, it can generate synthetic data for scenario simulation, stress testing, and portfolio backtesting (e.g., returns under changing inflation and GDP), capturing non-linear relationships without restrictive distributional assumptions.

Why it matters

For engineers, this adds a world-model branch to the agent/data stack: conditional synthetic-data generation can supplement historical data and imputation methods, enabling robust scenario analysis when history is sparse or regime-changing.

Generalization

Generative models can be repurposed as domain simulators in any data-scarce area that requires counterfactual scenarios or stress testing.

Generative AI models trained on historical time series can generate synthetic data for scenario simulation, stress testing, and portfolio backtesting.
Open source video
Using generative AI to simulate portfolio returns under different inflation and GDP scenarios.
Open source video
Empirical models have limits; generative models offer greater flexibility and non-linear relationship modeling without restrictive distributional assumptions.
Open source video
Architecturelow noveltymoderate evidence

The emerging architectural pattern in finance is an orchestration layer that sits above existing terminals and databases (Bloomberg, Excel, various databases) rather than replacing them. The analyst's workflow, encoded as a skill or prompt, becomes the programmable unit that orchestrates queries and actions across these backend systems.

Why it matters

Enterprise integration shifts from replacing monolithic tools to building connectors and a shared workflow layer. This preserves existing data infrastructure while improving task execution time and making workflows repeatable and auditable.

Generalization

Enterprise agents can be deployed as a control plane over heterogeneous internal tools rather than as a new data plane that replaces legacy systems.

AI tools now sit above these systems to orchestrate queries and workflows.
Open source video
Reduces task execution time and streamlines investment research workflows.
Open source video

Deep dives

5

Behavioral-bias red-teaming for LLM-driven finance agents

Research question

What evaluation protocol, metrics, and red-team workflow can systematically detect and measure loss-aversion and related behavioral biases in an LLM-based stock-screening or investment workflow before deployment?

Why

LLMs trained on human-generated content mirror human cognitive biases, and autonomous finance agents may execute without continuous human oversight. Under fiduciary obligations, converting subtle training-data skew into systematic investment errors is a release-blocking defect, not an afterthought.

LLMs exhibit human-like biases, including loss aversion bias.
Open source video
LLM-driven screening favoring certain stock ratios over others due to training data skew.
Open source video
Red-teaming models to mitigate systematic investment errors.
Open source video
Source video

Cost, privacy, and routing thresholds for local open-source versus proprietary models in agentic finance workflows

Research question

At what call volumes, latency requirements, data-sensitivity levels, and task complexity profiles does a locally hosted 36B-parameter open-source model become the economically and governance-preferred choice over repeated calls to a proprietary API?

Why

High-frequency agent calls to proprietary APIs create per-call expense and client-data exposure; local open-source models trade infrastructure investment for privacy and marginal-cost control. Engineering teams need a defensible decision framework rather than defaulting every call to one large model provider.

Proprietary models (e.g., OpenAI) offer high capability but raise data privacy and cost concerns when making frequent API calls.
Open source video
Running smaller open-source language models locally can achieve cost optimization and protect proprietary data.
Open source video
Infrastructure investment is required to run open source models effectively at scale.
Open source video
Running local 36B parameter models to avoid per-call API fees.
Open source video
Source video

Validation of generative synthetic financial time series for scenario simulation and stress testing

Research question

What validation criteria can establish that generative-model-simulated scenarios (e.g., portfolio returns conditioned on inflation and GDP shocks) preserve tail risk, non-linear dependencies, and regime behavior well enough to inform portfolio backtests and stress tests?

Why

Generative AI appears to offer flexible, assumption-light simulation, but unvalidated synthetic data can silently corrupt stress-test conclusions, especially when trained data is sparse, missing, or biased.

Generative AI models trained on historical time series can generate synthetic data for scenario simulation, stress testing, and portfolio backtesting.
Open source video
Using generative AI to simulate portfolio returns under different inflation and GDP scenarios.
Open source video
AI models are only as good as the data they access.
Open source video
Empirical models have limits; generative models offer greater flexibility and non-linear relationship modeling without restrictive distributional assumptions.
Open source video
Source video

Portable markdown skill-file and MCP contract for executable finance workflows

Research question

What minimal schema, versioning, and runtime contract allows a markdown skill file—containing workflow steps, embedded Excel model templates, and MCP server connectors—to be executed reproducibly across independent agent runtimes?

Why

Skill files are proposed as the low-code unit of automation, standardization, and governance in finance. Without a portable contract, firms cannot exchange workflows, compare outputs, or treat a skill file as an auditable governance artifact across different stacks.

Skills/master prompts provide sound structure for autonomous workflows.
Open source video
Using markdown files to outline specific financial workflows.
Open source video
Pre-populating Excel models inside skill files.
Open source video
Advanced users can connect MCP (Model Context Protocol) servers for native API integration with outside data.
Open source video
Source video

Orchestration control-plane pattern over legacy finance data systems

Research question

What integration and observability patterns should an AI orchestration layer implement to coordinate queries and workflows across existing Bloomberg, Excel, and database systems without undermining data governance or traceability?

Why

The realistic enterprise architecture is not replacing incumbent terminals and data stores but placing an orchestration control plane above them. This changes integration work from replacing monolithic tools to building stable adapters and a workflow layer that preserves existing infrastructure while making tasks faster and auditable.

AI tools now sit above these systems to orchestrate queries and workflows.
Open source video
Reduces task execution time and streamlines investment research workflows.
Open source video
Source video

Article ideas

4

Behavioral Bias Is a Release-Blocking Bug, Not an AI Ethics Topic

Agentic finance workflows should not be deployed until they pass adversarial bias tests derived from behavioral finance, because removing the human from the loop makes LLM training-data skew—especially loss aversion—a direct fiduciary and portfolio risk.

Angle

Governance and model-risk engineering

Source video

Your High-Frequency Finance Agent Should Probably Run on a Local 36B Model

For frequent, data-heavy finance tasks, defaulting every call to a proprietary API is both a privacy leak and a cost anti-pattern; routing repetitive work to small local open-source models is the economically and governance-sound architecture.

Angle

Engineering economics and agent architecture

Source video

The Skill File Is the New Governance Artifact

When investment workflows become editable markdown files rather than bespoke code, governance review shifts from opaque application code to readable workflow diffs, letting compliance and non-technical stakeholders shape automation directly.

Angle

Workflow automation and governance

Source video

Synthetic Market Scenarios Need a Crisis Bench

Generative scenario simulation is only as credible as its ability to reproduce the non-linear tails of historical crises, so every synthetic-data pipeline for stress testing should be validated against held-out regime-shift periods before it informs capital decisions.

Angle

Quantitative methodology and model validation

Source video

Project ideas

4

Adversarial Loss-Aversion Screening Probe

beyond-evals

An LLM configured for stock screening will display statistically significant loss-aversion bias when equivalent expected-value investment cases are framed as gain-seeking versus loss-avoidance, and a red-team probe can reliably surface that bias before deployment.

Proof of concept

Construct paired screening cases with identical expected values but opposite gain/loss framing; ask the LLM to rank or screen the opportunities; also run adversarial variants designed to amplify loss aversion; compare with a neutral baseline prompt set.

Measurement

Proportion of loss-framed vs gain-framed choices; chi-square or logistic-regression p-value; agreement rate with neutral baseline recommendations.

Source video

Local 36B Cost-Router for High-Frequency Finance Tasks

movement-lab

For a representative high-frequency finance data-extraction task, routing calls to a locally hosted 36B open-source model cuts cost per 1,000 calls by at least 70% versus a comparable proprietary API while maintaining at least 95% field-level agreement and acceptable p95 latency.

Proof of concept

Deploy a 36B open-source model locally; build a task router that selects the local model for repetitive calls; run 5,000 extraction tasks against both the local model and a proprietary API; compare outputs, cost, and latency.

Measurement

Cost per 1,000 calls; p95 latency per call; aggregate field-level agreement between local and API outputs.

Source video

Synthetic Scenario Crisis Validation Harness

new

A generative time-series model trained only on pre-crisis macro history will fail to reproduce the return distribution of a held-out historical crisis period, with a Kolmogorov-Smirnov test rejecting distributional equivalence at p < 0.05.

Proof of concept

Train a generative model on macroeconomic and equity-return history before a known turbulent period; generate scenarios conditioned on negative growth and rising inflation; compare the simulated return distribution to the actual crisis-period distribution.

Measurement

Kolmogorov-Smirnov p-value; simulated versus actual tail-mass ratio below the 5th percentile; maximum-drawdown exceedance rate.

Source video

Portable Skill-File Interoperability Test

gatehouse

A single markdown skill file that embeds an Excel model template and uses MCP connectors will, when executed by two independently built agent runtimes, yield equivalent outputs within a 1e-6 numerical tolerance for a representative financial workflow.

Proof of concept

Author a minimal skill file describing a monthly data-pull, forecast-parameter, and Excel-model-population workflow; implement two lightweight runtimes (e.g., a Python executor and a CLI agent executor); execute the same file across both and compare produced spreadsheets.

Measurement

Field-level numerical tolerance; number of runtime execution errors; schema validation pass rate.

Source video

Architectural implications

5

Skill/master-prompt files can specify financial workflows with embedded Excel models and external API connectors, making the workflow definition a file artifact.

Before

An analyst manually runs Bloomberg/Excel/database steps or developers hard-code integrations for each task.

After

An agent runtime interprets a markdown skill file, populates a model template, and calls MCP-connected external data APIs.

Consequence

Teams can standardize and iterate on workflow text; the runtime must execute templates and APIs and track file version and workflow provenance in each run.

Source video

Analysts are shifting from manual multi-terminal navigation to centralized AI orchestration where agents sit above Bloomberg, Excel, and databases.

Before

Workflows execute in the analyst's head or runbooks, with manual context switching across tools.

After

AI orchestration sits above legacy tools and performs queries and workflow steps through unified skills and connectors.

Consequence

The architecture needs stable adapters to legacy/terminal systems plus observability of cross-system actions and task outcomes.

Source video

A firm can run smaller open-source models locally on proprietary data to avoid expensive per-call API fees and to protect confidential client information.

Before

Every task calls a large proprietary API, creating increasing per-call expense and data-exposure risk.

After

High-frequency or sensitive tasks are routed to locally hosted open-source models, while larger proprietary models are reserved for tasks where their additional capability is needed.

Consequence

The agent stack must support multiple model endpoints, local inference, routing, and cost/latency monitoring; infrastructure budgets shift from per-token opex to managed local-model capacity.

Source video

Generative AI trained on historical time series can produce synthetic scenarios for stress testing and backtesting, complementing traditional data handling such as linear interpolation and KNN imputation.

Before

Stress tests rely on historical data plus simple empirical models, interpolation, or imputation that may not capture non-linear relationships or unusual scenario combinations.

After

A generative model produces many synthetic paths conditioned on scenario variables such as inflation and GDP, enabling richer portfolio stress tests.

Consequence

Data and simulation pipelines need controls to generate synthetic data, validate its plausibility, and label it distinctly from observed data.

Source video

LLMs exhibit human-like behavioral biases, including loss aversion, in finance-related outputs, so governance and risk management must be integrated into the AI system itself.

Before

Investment professionals rely on model outputs without systematic adversarial review for behavioral bias.

After

Red-teaming and rigorous evaluation frameworks become release gates for AI-driven workflows, checking for systematic investment errors before deployment.

Consequence

Agent CI/CD should include adversarial finance-bias test suites and evaluation checkpoints, similar to audit controls in established investment governance.

Source video

Tradeoffs and failure modes

3

Local open-source vs proprietary cloud models

Benefit

Running smaller open-source language models locally can achieve cost optimization and protect proprietary data, especially for high-frequency calls.

Cost or risk

Proprietary cloud APIs may be easier to operate at high capability, but frequent API calls raise privacy and cost concerns; local open-source models require infrastructure investment and internal ML operations.

Running smaller open-source language models locally can achieve cost optimization and protect proprietary data.
Open source video
Source video

Synthetic/generative scenario generation vs empirical model fidelity

Benefit

Generative models can simulate portfolio returns under different inflation and GDP scenarios without restrictive distributional assumptions, enabling richer stress tests and backtests.

Cost or risk

The usefulness of generated synthetic data is bounded by the quality of the underlying data; models are only as good as the data they access, so missing or biased training data can propagate into simulations.

AI models are only as good as the data they access.
Open source video
Source video

Speedy centralized orchestration vs fiduciary accountability

Benefit

Central AI orchestration above data terminals reduces task execution time and streamlines investment research workflows.

Cost or risk

Automation can amplify training-data skew into systematic investment errors (e.g., favoring certain stock ratios), so firms must uphold fiduciary obligations and deploy red-team controls before trusting autonomous workflows.

LLM-driven screening favoring certain stock ratios over others due to training data skew.
Open source video
Source video

Open questions

4

What validation protocol best establishes that LLM/generative-model synthetic financial time series are safe to use in backtests and stress tests, especially with respect to tail events and regime shifts?

Why unresolved

The summary states generative models can create synthetic scenarios and that empirical models have limits, but it provides no fidelity or validation criteria for synthetic data.

Research direction

Benchmark synthetic scenarios against held-out historical crisis periods and compare tail-risk, correlation, and scenario-conditional distributions with traditional stress-test assumptions.

Source video

How should red-teaming and governance gates be instantiated for loss-aversion-like behavioral biases in an automated investment pipeline?

Why unresolved

The speakers call for red-teaming and rigorous evaluation frameworks but do not define the specific evaluation protocol, metrics, or red-team workflow.

Research direction

Build a behavioral-finance adversarial benchmark for LLM stock screening, with tests that detect whether loss aversion, anchoring, or framing skews investment decisions.

Source video

What infrastructure budget, latency, and operational thresholds make locally hosted 36B open-source models more economical than proprietary API calls for a high-volume agent workload?

Why unresolved

The summary asserts local open-source models reduce per-call cost but requires infrastructure investment; no operational break-even analysis is provided.

Research direction

Measure throughput, latency, SLOs, and cost per 1K calls for local 36B inference versus large proprietary APIs across representative finance tasks.

Source video

How portable and standardizable are markdown skill files across different firms or agent runtimes, and what is the minimal contract needed between skill files, MCP connectors, and execution environments?

Why unresolved

The summary mentions standardization across firms and MCP connections, but no skill-file schema, versioning, or interchange mechanism is specified.

Research direction

Design an open representation/schema for agent skill files and test whether two independent agent runtimes can execute the same skill file and produce equivalent outputs.

Source video

Key claims

8
opinionVerification needed

Skill/master-prompt workflows allow non-technical practitioners to leverage AI for repetitive finance tasks.

Evidence

They allow non-technical practitioners to leverage AI for repetitive tasks.

Question

Can non-programmers successfully author and maintain these skill files in practice without engineering support?

Source video
factualVerification needed

Advanced users can connect MCP servers for native API integration with outside data.

Evidence

Advanced users can connect MCP (Model Context Protocol) servers for native API integration with outside data.

Question

Does the MCP architecture support all the authentication, pagination, and query modes needed for production finance data providers?

Source video
comparativeVerification needed

Proprietary models offer high capability but raise data privacy and cost concerns when called frequently, while local smaller open-source models can optimize cost and protect proprietary data.

Evidence

Proprietary models (e.g., OpenAI) offer high capability but raise data privacy and cost concerns when making frequent API calls.

Question

What is the measured total-cost-of-ownership difference between proprietary API calls and local 36B open-source inference for representative high-frequency tasks?

Source video
predictionVerification needed

Generative AI models trained on historical time series can generate synthetic data for scenario simulation, stress testing, and portfolio backtesting.

Evidence

Generative AI models trained on historical time series can generate synthetic data for scenario simulation, stress testing, and portfolio backtesting.

Question

Can generative time-series models produce realistic conditional return distributions under inflation and GDP shocks that pass standard financial backtesting validation?

Source video
factualVerification needed

LLMs exhibit human-like biases, including loss aversion bias.

Evidence

LLMs exhibit human-like biases, including loss aversion bias.

Question

Do controlled experiments on LLM finance outputs systematically reproduce loss aversion in investment decisions?

Source video
causalVerification needed

LLM-driven screening can favor certain stock ratios over others because of training-data skew, creating systematic investment errors.

Evidence

LLM-driven screening favoring certain stock ratios over others due to training data skew.

Question

Can training-data composition be causally linked to specific screening biases in finance LLMs, and can they be mitigated by red-teaming?

Source video
factualVerification needed

AI tools can sit above Bloomberg, Excel, and databases to orchestrate queries and workflows for investment research.

Evidence

AI tools now sit above these systems to orchestrate queries and workflows.

Question

Which production finance workflows currently demonstrate this control-plane orchestration pattern, and what is the measured time savings?

Source video
comparativeVerification needed

Generative models offer greater flexibility and non-linear relationship modeling than empirical models without restrictive distributional assumptions.

Evidence

Empirical models have limits; generative models offer greater flexibility and non-linear relationship modeling without restrictive distributional assumptions.

Question

Across what financial datasets and conditions do generative models outperform empirical models in out-of-sample scenario accuracy?

Source video

Connections

5