Dan Shipper of Every on using ChatGPT Agent to autonomously analyze feedback for his email AI app

Summary

Dan Shipper, CEO of Every, used pre-launch access to ChatGPT Agent to run a customer intelligence task that would have otherwise consumed days of manual work. He fed the system roughly 1,500 support emails and 500 forum posts from the past two months for Kora, Every's AI email management app, directing the agent to identify customer archetypes, classify promoters and detractors, cross-reference those users on LinkedIn for demographic and professional context, and compile a structured research report. The entire workflow required little more than a prompt and a pre-authorized Gmail OAuth connection.

The agent spins up a cloud-hosted virtual machine, browses the web autonomously, and pulls from connected data sources without the user managing any infrastructure. The one friction point Shipper flagged was LinkedIn authentication, which required a manual password entry via a browser takeover in the virtual machine. He describes the overall experience as consumer-grade, contrasting it directly with Anthropic's Claude Code, which runs locally in the terminal, requires API-level thinking, and offers more control and composability but is effectively inaccessible to non-developers.

Shipper's framing of the OpenAI versus Anthropic agent architecture split is worth noting. ChatGPT Agent abstracts the compute layer entirely, optimizing for ease of use across any device. Claude Code exposes that layer, which makes it more powerful for technical users but limits the addressable audience. He believes Claude Code is currently underrated, noting it handles non-programming tasks well despite its terminal interface, but concedes ChatGPT Agent will reach a far broader user base.

The scheduling capability is a meaningful differentiator. Shipper plans to run the same customer feedback analysis automatically on the first of every month, a task no one on a small team would realistically prioritize manually at that cadence. Recurring autonomous research reports, web monitoring across paywalled sources, and personalized daily news digests are the use cases he expects to gain the fastest traction with consumers.

Shipper is explicit that ChatGPT Agent is not a daily driver. For most tasks, GPT-4o remains faster and sufficient. His mental hierarchy runs from 4o for quick lookups, to o3 for more considered responses with web context, to Deep Research for comprehensive reports, and finally to Agent only when the task requires taking real-world actions across systems. The agent tier is reserved for multi-step, multi-source workflows that would otherwise require delegating to a human.

The behavioral challenge he flags is harder to solve with a product update. Shipper found himself watching the agent's animated UI in real time rather than letting it run in the background, describing it as an attention drain. He compares the trust calibration required to the learning curve new managers face when delegating to employees, arguing that users will become meaningfully better at deploying agents in three to six months as intuitions about when to intervene versus when to step back become more developed.