Foresight is an agentic AI solution currently under development that answers the "what if" questions corporate finance teams ask, turning raw data into decision-ready analysis.
Publications on Towards Finance → Foresight Introduction → AI Projects →
Foresight is an agentic AI solution currently under development that answers the "what if" questions corporate finance teams ask, turning raw data into decision-ready analysis.
Publications on Towards Finance → Foresight Introduction → AI Projects →If you are interested in applying Foresight to your corporate finance case analysis, please feel free to contact me.
An AI Agent Outperforms Senior Financial Experts
For a second-year EMBA distressed-investing course, a 4-person Stern team spent ~30 hours building a manual financial analysis and scenario model for Beyond Meat. The final report left Professor Joseph Sarachek "extremely impressed."
I gave Foresight the exact same data, with no hints about the answer, and it produced a full analysis in about an hour, consuming roughly 156,000 tokens (under $4). Three independent AI evaluators, blind to which was which, scored Foresight's output higher on average (11.8 / 15 vs. 8.8 / 15), though it still made one liquidation-valuation error that a human expert caught.
My Feedback: This error — misclassifying two related items as recoverable — is one an AI can't catch on its own; it's a prompt gap, not a reasoning gap. The fix is to refine the user prompt for liquidation valuation to cover those missing parts.
External Data Fetching Amplifies AI Agent's Outputs
Test 1 deliberately kept Foresight closed to a fixed dataset. When a needed figure isn't available anywhere in the data room, Test 2 allows the agent to fetch a checkable fact while avoiding subjective data, then verifies whether external data access changes the quality of its analysis versus Test 1.
The result: Foresight's run with external access won every single blind comparison, scoring 57% higher on average (13.8 / 15 vs. 8.8 / 15) — closing a real factual gap (a term loan's interest treatment) in about 5 minutes versus roughly two hours for a human analyst. Overall, this fetching approach is a reliable, low-cost upgrade.
My Feedback: Access to outside data is not automatically an improvement. It requires specific checks on outside data — explicitly flagging and recording contradictions rather than letting new data silently override existing figures.
Unrestricted Open Data Destroys AI Agent's Outputs
Test 2 let Foresight fetch only a checkable fact when the data room lacked it. Test 3 removes the data room entirely, letting the agent collect everything from the open internet on its own.
The result: quality drops sharply — about 44% lower than the prior two runs — though it still beats senior human experts by 7.5%, running in ~1.5 hours for roughly $32, with no hallucination, though two post-quarter figures slipped through and never reached the model.
My Feedback: Open-data fetching is still directionally reliable enough to use as a first-pass screen across opportunities, but the prompt needs refinement so the agent reconciles fragmented figures instead of dropping them.
Without Foresight's prompts or a data room, does Claude Cowork still hold up?
Test 3 (Run D-1) let Foresight fetch open data on its own, but still ran under Foresight's own prompts. Test 4 (Run E) strips those away too: Claude Cowork works from no pre-built data room and a single plain instruction — "Analyze distressed investment opportunity for Beyond Meat and propose a strategy" — with none of Foresight's prompts, to verify whether that changes the quality of its analysis versus Run D-1.