---
title: Reverse underwriting: rating AI products by task
subtitle: How consumers can assess AI products by task and decide how much authority to delegate.
kind: Essay
date: 2026-04-30
author: Ron Bronson
revised: 2026-10-16
---

People already make judgments about which AI products to trust. Someone might use Claude to edit a contract, ChatGPT to plan a trip, and Gemini to search a document. Their confidence in each product may differ by task. Vloora calls the process of making those judgments explicit **reverse underwriting**.

## Evaluate the AI product the customer uses

The product being evaluated is the service the consumer recognizes. An individual need not know which model revision, internal agent, or tool-routing system handled the request. The provider remains accountable for what its product does.

A performance history can record the task, the person's acceptance or correction of the result, and any outcome independently confirmed by a receiving service. Each is a different observation. A user accepting a draft shows that they accepted it and says nothing about whether it was accurate.

## How performance history informs a customer's permissions

Consumer agent credit describes a person's assessment of an AI product's performance in a particular domain. Delegated authority is a separate decision: what the product is permitted to read, submit, change, or spend.

Someone may trust a travel assistant to compare fares while requiring approval before booking. Good performance can justify recommending an expanded permission, but granting it remains the person's decision. Poor performance can support a narrower permission under limits the person has already established.

It works like underwriting in reverse: The customer assesses the software supplier and decides how much discretion to extend.

## An experimental score tracks performance by task

Vloora's experimental credit library keeps a locally stored performance ledger and a guarded-session mechanism. Its current 0–10 score responds to interaction signals including rejection, acceptance, and edits. The score is experimental, and no one has validated it as a measure of correctness. Enforcement applies only when a host routes calls through its guard.

A useful personal credit file would show the underlying tasks and evidence, disclose why an assessment changed, and give the consumer a way to contest an error. Any score it shows would come second to that history.

## A customer can revisit permissions when a product changes

Products change. A provider may update a model, integrate tools, or change how work is routed. A consumer should be able to respond to the results without reconstructing the provider's internal architecture.

The person decides which requests that product can still make on their behalf. A portable record of outcomes informs that decision, and permission controls carry it out.

## Sources

- Vloora, Stenodisc, decision models, and reverse underwriting, research brief, updated October 16, 2026
- Vloora, [When an agent should ask you](/writing/when-an-agent-should-ask-you/), January 15, 2026, on delegated authority and approval