Can AI users judge an app first?

Every change to a product like Narrative Coach raises the same question: will it help the owner who uses it? Real people answer that best, but recruiting them for every change is slow. So our CTO, Chonghuan Wang, has been asking whether AI users could judge an app first.

What simulated users can do

An AI user can be given a business profile and a goal, such as a café owner worried about a quiet January, and asked to work through the app. Many such users can run through a journey in the time it takes to schedule one interview, across many kinds of business. They are good at finding steps that confuse, answers that miss the question and journeys that lead nowhere.

What they cannot do

A simulated user does not run a business. It cannot tell us whether an owner would trust an answer, act on it or come back the next day. For those questions we rely on people, as in our paid pilot, which an independent research agency ran and evaluated with small business owners using their own accounts.

How this fits with our testing

Simulated users would add a layer to testing we already do. Before any change to Coach’s models or prompts is released, it runs against fixed test sets. Each response is judged twice: once for whether it is technically correct, and once for whether a small business owner would find it useful. We also test adversarially, and the product has been reviewed against the OWASP Top 10 and the OWASP Top 10 for generative AI.

The aim is to catch problems early, so that the time we spend with real owners goes on the questions only they can answer.