Perpetual's design for finance automation: people approve the plan, not fill the form

Karthic Rao of Perpetual described a design for automated finance work on the AI Engineer show. In it, the model interprets messy requests. People approve plans and outgoing actions, and they confirm the model's uncertain calls. Fixed software, not people, holds the evidence trail.

What was said

Rao began with a one-line request: fill out a form for Mr. Harper. A lookup finds three Harper households, so someone must ask the advisor which one is meant. The form then lacks three values. The assistant emails the client, waits, follows up, requests an e-signature and files the result. One sentence becomes days of work.

Rao said this is not simple data entry. "Every form is a small program," he said. A W9 tax form, for example, asks who the taxpayer is, how they are taxed, which number applies and who signs.

In his design, deterministic software, meaning code that acts the same way every time, controls the order of steps, the evidence and any action that changes the world. Sending an email is one such action. The model only makes small, typed judgments, such as whether a client's reply is complete. A person signs off on the whole plan up front. Each outgoing email or signature request then needs its own approval, tied to its exact content.

Why it matters

In our reading, the human work in this design is judgment, not typing. People decide whether a plan is right, read the exact email before it leaves, and settle doubtful model readings. The software writes the record.

Rao argued that a fixed flowchart breaks on ambiguity, while a freely improvising agent is something a compliance team cannot sign off. Named approval points are his answer to both problems. For buyers and managers, the practical questions follow. Where are the approval points? What does a person see before approving? Does the system generate the evidence trail automatically?

The other side

The excerpts do not say how people learn this judging skill. If software does the form-filling, they do not say where staff gain the experience to approve a plan or overrule a model. They also do not say how firms would stop approvals from turning into routine clicks. The excerpts have gaps, so the full talk may cover more.

The numbers are design goals, not results. Rao said the design aims for about 95% of model decisions to need no frontier model, meaning the most capable and costly tier. His example was an illustrative run of 20 decisions. He said logging and evaluation exist today, but the training loop that would shrink the frontier share comes next. This is also a builder describing his own design, and the excerpts show no live results.