Skip to content
Security & Trust Talking point

When an action's outcome is unknown, send it to a person, not a retry loop

At Perpetual, an action with an unclear result goes to a person, which shifts who carries the cost of failure.

W
WebPulse Newsroom
AI-assisted · 2 min read
Share on X LinkedIn
When an action's outcome is unknown, send it to a person, not a retry loop
In brief
  • Karthic Rao of Perpetual described workflows where an action with an unknown outcome stops for manual review, with no automatic retries.
  • That is a choice about who absorbs a failure: a staff reviewer's time, or the client on the receiving end of a repeat.

Karthic Rao of Perpetual described a workflow design on the AI Engineer show. In it, an action whose outcome is unknown stops for manual review, with no automatic retries. He was talking about finance, wealth and insurance. We think that is a sound default, and it shows who pays when things go wrong.

What was said

Rao described a state machine, a plain program that runs an approved plan step by step. Model calls happen only at set points, and each returns a typed answer the program then acts on.

Actions on the outside world, such as sending an email or a signature request, get their own lifecycle. Rao listed the stages: prepare, dispatching, succeeded. If the outside world gives no clear answer, the action is marked outcome unknown and goes to manual review. Rao said the system makes no automatic retries.

The same habit shows up elsewhere in his design. If a client answers only some of what was asked, a model labels the answer incomplete. Rao called that label provisional. The step then waits, and a person must confirm the label before work moves on. He summed up the model's role this way: "The model steers. It never drives."

Why it matters

Our reading: every system with an uncertain step makes a choice about who absorbs the doubt. An automatic retry bets that trying again is harmless. If the bet is wrong, the person on the receiving end pays. Manual review moves that cost to staff time and a slower queue.

For buyers, this gives a concrete question to put to any vendor. What happens when a send or a signature request gets no clear answer? Does the system guess, or does it stop and show a person?

For builders, Rao's design shows the answer can be built in. Each outside action has its own approval gate. The approval is tied to the exact payload and checked again just before sending.

The other side

Rao described one narrow case: an unclear answer from the outside world. The excerpts do not say what happens to an action that clearly failed. They also do not say why he ruled out retries, or how often a case lands in manual review. The reasoning about duplicate messages and who bears them is ours, not his.

A human queue is not free. It costs staff time, and an unresolved case can sit while a client waits. The cost moves rather than disappears. The excerpts also do not show how a reviewer settles an unknown outcome.

Rao's setting is regulated work, where the evidence trail matters. A low-stakes workflow might reasonably choose differently.

Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.

The conversation this talking point comes from

Share this insight