Why AI should be a form, not a conversation

Maybe it was a decade ago when it seemed like every single game developer started believing that all games should be open world. Infinite possibilities, player freedom, emergent storytelling, who wouldn't want that? But open world design introduced new problems that linear games never faced. Players would get lost in meaningless side quests while the main story waited forgotten. Developers would pile on features that didn't serve the core experience. Worst of all, players often couldn't tell when they'd actually "won." I sank an unhealthy amount of time into Metal Gear Solid V when it came out, wandering its open world long after I'd completed every meaningful objective, never quite sure if I was done or just... tired.

The problem with open world design is scope creep disguised as feature richness. When everything is possible, nothing feels special.

Today, with this new wave of AI, we're making the exact same mistake. We're taking the "open world" approach to problem-solving, and the results are just as messy.

When the chatbot craze was at its peak, I worked for a startup where we automated customer service. Many of our competitors were building what I now recognize as "open world chatbots". AI systems with infinite conversational possibilities. These chatbots would greet you, ask about your day, make jokes, and try to make you forget that you had a very specific reason for starting the session. Mind you, this was before ChatGPT, when LLMs weren't widely available.

Each competitor's demo showcased the chatbot's vast capabilities through highly choreographed scenarios. Like game developers showing off sprawling landscapes and complex skill trees, they were proud of the sheer scope of what their AI could theoretically handle. The problem was identical to open world games: the moment real users engaged with these systems, everything collapsed. Customers, like players, are unpredictable. They don't follow the golden path you've designed.

The linear gameplay

Our approach was fundamentally different. We built what I'd call a "linear gameplay" AI. Our chatbot didn't try to converse with you. It didn't greet you, didn't chat about the weather, didn't pretend to be your friend. It appeared as a form or email address, a clear single point of entry with one mission: solve your customer service problem.

When you sent your fully formed message, our application took over with purposeful, sequential steps. It read the message, classified it, analyzed sentiment, then moved through our integrations (Zendesk, Shopify, Magento, USPS) like a player progressing through carefully designed levels. When it retrieved your tracking information or order details, it crafted a response that directly answered your questions. If at any point it couldn't complete the mission, it executed a clean handoff to a human agent who received a comprehensive summary. Like a save file with all the relevant progress data.

This agent was designed for one specific quest: resolve customer service problems. Nothing else. Just like how the best linear games focus relentlessly on their core mechanics rather than trying to be everything to everyone.

A lot of our "open world" competitors have either pivoted or gone out of business since then. Which makes it even more surprising to watch Taco Bell's drive-thru AI failing in exactly the same way.

When open world design meets fast food

Just a few weeks back, Taco Bell went through a PR nightmare when a customer ordered 18,000 cups of water, and the system dutifully added every single cup to the order. This wasn't malicious hacking; this was a classic open world design failure. The system was built to handle "anything," so it handled everything, including completely absurd requests that broke the entire experience.

Taco Bell uses a service called Omilia to power their AI drive-thru. On their website, they explicitly describe their open world approach:

Taco Bell recognized the importance of providing a seamless, human-like experience for customers, so Omilia's Voice AI Solution was meticulously tuned to provide clear, accurate responses, delivering a conversational experience that mirrors human interaction and improves the customer's ordering experience.

Notice the language: "human-like experience," "conversational," "mirrors human interaction." They built an open world when they needed linear gameplay. The conversational nature invites exactly the kind of scope creep that breaks systems. Regular customers report their perfectly normal orders failing spectacularly, with the AI getting stuck in loops, asking "Would you like a drink with that?" even after drinks were clearly specified.

A linear approach works

I couldn't find a Taco Bell with AI drive-thru in my neighborhood, but I did find Rally's. They use a different company called HI Auto, and to my surprise, it worked flawlessly. The experience felt like a well-designed level progression: clear prompts, sequential steps, defined objectives.

"What would you like to drink?" "What size?" Next level unlocked.

It wasn't a conversation, it was a voice-powered form. No philosophical debates, no jokes, no attempts at charm. Just a transaction with guided rails and clear success criteria. The software knew exactly what it was supposed to accomplish, and users knew exactly what was expected of them.

This is the difference between open world chaos and linear focus. A drive-thru isn't a space for exploration and emergent dialogue. It's a hyper-linear experience where success is defined as "correct order, minimal time." Any system that invites deviation from this core mission is poorly designed for its context.

Pretend you are filling out a form

You could theoretically plug ChatGPT into a drive-thru experience, just like you could theoretically add an open world to any game. But you'd be guaranteed to receive unexpected inputs that break the core experience. Instead, treat AI applications like carefully designed HTML forms with proper validation, clear fields, and predictable outcomes.

The goal is always the solution (a correct order, a resolved support ticket, a completed task), not the conversational medium. In fact, I think voice itself is not the most optimal for these linear experiences. The most efficient drive-thru model might be a QR code at the menu board that lets you complete the entire "form" on your phone before reaching the speaker.

---

Choose linear when story and progression matter, open world when exploration serves the core purpose. The best AI applications should match their interaction model to their mission.

Maybe OpenAI's goal is to build an AI. But your drive thru AI is perfectly fine being specialized. Start building AI systems that do one thing exceptionally well, with clear boundaries, defined success criteria, and linear paths to completion. Your users will thank you, and you won't end up in viral videos featuring 18,000 cups of water.

---