The promise of a truly helpful AI assistant is tantalizing – imagine simply *telling* ChatGPT to book a dinner reservation and having it done. A future where complex tasks are handled with a voice command feels within reach, yet remains frustratingly just out of grasp.
Major AI developers like OpenAI, Anthropic, and Google are aggressively integrating their models with everyday services. Spotify, DoorDash, Resy – the list grows daily, but the experience is often far from seamless. These integrations rely on “tool calls,” structured requests that allow the AI to interact with the real world, but the execution is proving challenging.
These tool calls are essentially pre-defined functions. For example, ChatGPT’s Spotify integration includes commands like “add_to_library” or “get_currently_playing.” Developers carefully control these functions, prioritizing safety and predictability, but this control often comes at the cost of user experience.
I tested Anthropic’s Claude with the Resy restaurant reservation app. After connecting the services, a conversation about dinner near Midtown Manhattan led to a solid restaurant recommendation. But when I asked Claude to *make* the reservation, things quickly became awkward.
Instead of a simple confirmation, Claude dumped a block of raw JSON code – essentially computer instructions – into the chat window, accompanied by “Always allow” and “Deny” buttons. It felt less like an assistant and more like interacting with a developer’s debugging screen.
Clicking “Always Allow” revealed a grid of available time slots. Asking Claude to book a specific time resulted in another message: “You’ll need to tap the slot in the Resy widget and finish the reservation there.” Suddenly, I was being redirected to Resy, forced to complete the booking myself – defeating the purpose entirely.
This clunkiness isn’t unique to Claude. ChatGPT and Gemini face similar limitations. They can get you 90% of the way there, but that final 10% often feels more laborious than simply using the app directly. The convenience factor vanishes.
The true potential lies in AI agents that can autonomously navigate websites and complete tasks. ChatGPT’s Agent mode, equipped with a built-in browser, attempts this. I tasked it with booking a table at Rosa Mexicano in New York City.
What followed was a surprisingly comical display of ineptitude. The agent struggled with basic web interface elements, like selecting the correct party size from a drop-down menu, repeatedly correcting its own mistakes. It was painfully slow and inefficient.
After five minutes of agonizingly slow clicking, ChatGPT announced the table was “being held,” but then handed the process back to me for personal details. The experience underscored a simple truth: for now, it’s faster to just use OpenTable myself.
The vision of seamless AI assistance is still valid. Eventually, these systems *will* handle our reservations, groceries, and travel plans with ease. But today, the reality is a frustrating mix of promise and clunky execution. For now, the apps remain the more reliable choice.
