AI Notebook] Agentic AI in Finance: Notes from the CFA Institute Roundtable

EJ's note

CFA Institute's Research and Policy Center convened Brian Pisaneschi and James Tait for a roundtable on how agentic AI is reshaping investment workflows, data practice, and governance. Worth 26 minutes of your time — but if you don't have them, here's my summary.

Bottom line: The barrier to agentic AI in investment workflows has collapsed. You no longer need Python to build autonomous workflows — you need well-documented process knowledge. That shifts the advantage from technologists to domain experts, and shifts the governance question from "should we use AI" to "how do we test, de-bias, and trust it at scale."

My takeaway: for investment organisations, the frontier has moved from access to process. Everyone will have capable models. Differentiation will come from how well you encode your domain expertise into workflows, how rigorously you test for the biases you've now inherited, and whether your governance keeps pace with what your analysts are already automating.

Enjoy!

I. Skills and MCP: agentic AI without code

The most practical development discussed is "skills" — essentially master prompts in structured markdown files that define a specific workflow, bundle resources (say, a valuation template), and connect to outside data via MCP servers. The AI now sits above your existing stack — Excel, market data terminals, internal databases — as the interface layer, deciding when to invoke which tool.

Two points stood out. First, no programming required: any analyst who can document a repeatable process can automate it. Second, skills compound. When the model errs, you encode the fix into the skill file. Across a team of analysts iterating on the same workflows, you get steadily more robust automation even if the underlying models never improve. Your firm's process knowledge becomes the asset.

II. Data gaps and synthetic data

On the data side, the toolkit runs from classical fixes (interpolation, k-nearest-neighbour imputation) to the more interesting frontier: generative models trained on historical returns plus macro variables. Because these learn relationships empirically — no imposed functional form or distribution — they can simulate return paths under specified macro scenarios. The applications are stress testing, scenario analysis, and backtesting that go beyond bootstrapping empirical returns, which assumes regimes don't change. They do.

III. Open source vs. proprietary: a cost question, not a capability question

Open-source models lag the frontier by only a few months, and much of the value has migrated to the "harness" — the agentic architecture around the model — where the panel sees winner-take-all dynamics favouring the frontier labs. The pragmatic answer for firms with data-confidentiality constraints: use proprietary models as your benchmark, test open-source alternatives against it, and train small local models for high-volume repetitive tasks where API costs would balloon. Proprietary wins on scale (50,000 batch requests, results in 24 hours); local models win on cost and privacy.

IV. Governance: the machines inherit our biases

The sharpest warning: delegating to an LLM does not remove human bias — it can encode it. Trained on human-generated data, these models exhibit documented behavioural biases (including loss aversion — holding losers too long) and skew toward Western, US tech-heavy training data in their default recommendations. Left unspecified, a model screening stocks will implicitly choose its own criteria.

The mitigation playbook: make instructions explicit (your ratios, your criteria — in the skill file), red-team models for bias the way cybersecurity teams attack their own systems, and build trust the way you would with a new analyst — many tasks, systematic evaluations, reviewed output distributions. Not vibes; evidence.

Resources

Homepage | CFA Institute Research and Policy Center

Explore expert CFA Institute Research and Policy Center analysis on AI, private markets, and the future of investment management.

CFA Institute Research and Policy Center