Intel / video

Reliable, Smart Workflows for Finance, Wealth and Insurance — Karthic Rao, Perpetual

Source: youtube.com
Author: AI Engineer
Date: Published Oct 12, 2026

Context

Perpetual's workflow design gives AI room to interpret a financial-service request while keeping execution under explicit rules. The talk's publisher description describes a planner that converts a request into connected steps. A compiler checks those steps against the firm's permitted actions and returns a valid plan, a clarification request or a refusal.

Once approved, the plan cannot change, and external actions require approval tied to their exact contents. Waiting and follow-ups become recorded workflow states. An uncertain action outcome goes to manual review rather than an automatic retry. The intended benefit is a process that can handle ambiguity without letting each model judgment rewrite the rules for execution.

Key takeaways · AI-distilled

Rao's planner turns a request into a semantic graph whose nodes point back to the original request, so every planned step stays traceable to what the user actually asked.
The compiler has three outcomes, not two: a valid plan, a clarification request, or a refusal. A judgment made mid-run can spawn a smaller child plan, but that plan must pass the same compilation and approval gates.
External actions get their own approval gate tied to the exact payload, and waits and client follow-ups are modeled as state machine steps. Uncertain action outcomes go to manual review instead of automatic retries.
Rao's illustrative 20-decision run routes 19 decisions to a smaller model and one to a frontier model. He presents that ratio as a design target, not a measured fleet result.

Read the source www.youtube.com