OpenAI built ChatGPT Work for people who already juggle a dozen tools. Independent researcher and writer Shlok Khemani thinks that framing undersells what it actually is.
In a detailed breakdown of the product, Khemani argues Work is not a side product for specialists. It is a preview of the default ChatGPT experience.
What Khemani is saying
Khemani’s central claim is direct:
“Work, then, is not just a niche product for power users, but a preview of how ChatGPT’s billion weekly users will soon use the app.”
The idea is that Work’s model, letting ChatGPT plan a task, call out to installed plugins, and pull in outside tools rather than just answer from its own training, is not a niche feature set. It is the shape every ChatGPT session eventually takes once the app is doing real work instead of answering trivia.
But Khemani also tested the feature closely enough to find where it breaks. When he asked ChatGPT Work to search for flights and hotels, it skipped installed plugins built for exactly that task:
“When I asked it to search for flights and hotels, it ignored several available but uninstalled travel plugins in favour of web search, even though a plugin might have used fewer tokens, returned better results, and let me complete a booking directly.”
He pushed further, naming a specific travel plugin by name. It still did not offer it: “Even naming Expedia outright didn’t prompt it to offer the plugin.”
Who Khemani is and why the take carries weight
Khemani describes himself as an independent builder and writer based in Mumbai, previously writing and building products in the crypto space at Decentralised.co. His recent work focuses on AI memory systems and personal AI products, including public research talks on how AI labs build memory into their models.
His critique comes from testing the product himself: the specific tasks he ran, the specific plugins he expected to see, and the specific results he got back.
What the piece gets right, and where it is incomplete
Khemani’s strongest point is specific and testable: he names the exact failure mode (installed plugin available, not suggested), the exact workaround he tried (naming the tool directly), and the exact result (still ignored). That is a concrete product bug, not a vague complaint about AI assistants being unreliable.
The piece is less strong as a general theory. Khemani generalizes from his own testing session to a claim about how a billion people will use the app. That is a reasonable bet given ChatGPT’s direction, but it is Khemani’s forecast, not something OpenAI has confirmed or something verified across many users or sessions.
Why it’s notable
Plugin and tool discovery is the same problem builders hit inside their own AI coding agents: an agent with access to ten tools that reaches for the wrong one, or no tool at all, is worse than an agent with no tools, because it fails silently instead of telling you it’s missing something. Khemani’s complaint is that ChatGPT Work has the same blind spot at consumer scale. It knows a tool exists in the user’s account, and still does not reach for it.
Khemani is right that routing through installed tools, rather than answering from training alone, is the correct direction for ChatGPT to move in. The failure he found is not a reason to doubt that direction. It is a reason to doubt that OpenAI has solved the harder problem underneath it: reliably knowing what it has access to before it decides how to answer.
What it means for builders
If you are building on top of an agent platform, plugin or tool visibility is not a detail to leave until launch. An agent that silently falls back to a worse method, like web search instead of a purpose-built tool, erodes trust faster than an agent that admits it does not know how to do something.
The practical check: test your agent’s tool selection the way Khemani tested ChatGPT Work. Give it a task where the right tool is installed but not the obvious first guess, and see what it reaches for. If it reaches for something worse without saying why, that is a discovery problem worth fixing before more people depend on the default path working.
End of article