AI for Investing
From demo to real: the production gap
The agent patterns live in a sandbox: simulated or open data, you in a notebook, an "interesting error" and no consequences. Production is real market feeds, clients and advisers and traders, financial liability, and real money. It is worth seeing the terrain before someone sells you a bridge.
The patterns don't change. Everything around them does.
Every tool gets a risk level
In production the risk level of a tool decides what happens when the agent wants to call it. Low risk, such as fetching a price or running an optimisation, runs automatically. Medium, such as sending an alert or exporting data, runs but is logged. High, such as placing an order or moving funds, requires a human. A sandbox demo lives entirely in the low-risk column. Production needs the whole spectrum.
Protection at every layer
It also needs controls at each layer. Input validation, to block dangerous requests before they reach the agent. Tool classification, so the risk level decides approval. Execution controls, such as position limits, rate limits and circuit breakers. Output validation, such as disclaimers, scrubbing personal data, and compliance checks. And audit logging, to record everything. A demo has none of these.
For high-risk actions the flow changes completely. You say rebalance my portfolio to 60/40. The agent prepares the action - it would sell 50 shares of one holding and buy 100 of another - and presents approve, modify, or cancel. A human decides. The agent never takes a high-risk action on its own: it prepares, presents, and waits. That is a different architecture, not a setting you switch on.
Why projects stall
This is a significant engineering effort, and it is why many finance AI projects fail - not because the patterns are wrong, but because the production infrastructure is underestimated. What you can build with the patterns alone is real: internal tools on read-only data and computation, such as dashboards, research assistants and risk calculators, plus proofs of concept that show stakeholders what is possible. What you are not yet ready for is agents that trade real money, client-facing systems, and the regulatory obligations that come with them. That is fine. Most people jumping into AI for finance do not understand the patterns. The production layer is an engineering challenge, not a conceptual one.
Keep going
Start at AI for investing for the whole picture, and if you work in a regulated firm, the governance side is in AI risk management. The full book, AI Agents for Investing, builds the patterns in code. I am publishing more in the coming weeks. Subscribe to get it.