The friendly draft that keeps reverting to generic
Sara Rojas gets the register she wants, then loses it. She has to re-prompt, the version she liked is gone, and nobody told her settings existed.
ChatGPT cut Sara Rojas's reply time from about ten minutes to under a minute, and it has one persistent failure she raised without being asked: the tone will not hold. She prompts for friendly, gets friendly, and then at some point the model reverts to a generic register. She has to prompt again, and the version she liked is gone.
What the failure looks like
Her description of the loop is exact. "I kind of tell it, be friendly, you know, make it more this, make it more that, and it starts generating, and then I don't know, at some point it decides that this is too much. I'm just gonna do a generic one. So it kind of goes back to it."
Galina Fendikevich names it in the interview as losing the context, and Sara agrees. The consequence is the part that costs her: "It takes me a while because I find one that I really like and then it gets lost."
So the register is not a setting she can lock. It is a state she has to keep re-establishing, and each re-establishment is a fresh negotiation with slightly different output. For a business where every reply goes to a prospective client, the voice is not decoration.
The discoverability problem underneath it
There is a fix, partially. ChatGPT has account settings where standing instructions can be stored. Sara did not know they existed. Her reaction on being told, on the record, is the whole product problem in one line: "I didn't even know you can do that. Like why didn't you tell me? It's like, hey, do you know that if you want, if you don't want us to keep changing it, you can go to the settings. They didn't tell me that. I just found out right now."
Galina's own objection is worth recording alongside it. If you are telling the assistant something in the chat, why does it not store that as an instruction to remember, rather than requiring the user to go and find a settings panel?
Sara's broader version of this is about vocabulary rather than menus. "I feel like there's a lot of ways that you can prompt things that will say better, but you know, I don't really have the keywords and how to like have it say that I how I want it." Learning to work the tool has been trial, error and a lot of clicking.
Galina frames the mismatch directly: the products are sold as assistants, so an operator should be able to give instructions the way they would to a junior colleague and have the background work happen. Instead the operator is being asked to learn a new language.
The volume ceiling
The second failure is arithmetic. Some mornings she wakes up to around ninety-four emails. At a minute each that is still an hour and a half of one-at-a-time drafting, and she calls it a tedious job at that scale. A handful is fine. Ninety-four is not.
What she wants is a pass over the inbox — everything drafted, then reviewed in one sitting, rather than ninety-four separate conversations.
Galina points out on the record that a technical route to this exists, and then rules it out for the right reason: Sara is not technical, she already has a full plate, she is not going to go and configure agents and back-end connections, and paying someone to do it is a significant extra cost. Someone who did not know about the account settings is not going to be building an inbox agent.
That is the gap. The capability exists. The packaging that would put it in the hands of the person who needs it does not.
Where this comes from
S2E2: The secret AI strategy this small NYC business uses to save thousands — the full interview with Sara Rojas and Luisa Ramirez. Listen or watch: YouTube, Spotify or Apple Podcasts.
