Three Kinds of AI Products That Currently Work
From: https://www.seangoedecke.com/ai-products/
Sean Goedecke identifies three categories of AI products that have demonstrated viability and success in the market: chatbots, completion tools, and agents. While many “AI products” are merely variations of chatbots, Goedecke argues that only specific implementations of these three archetypes truly deliver value. He also discusses two emerging categories—AI-generated feeds and AI-powered games—that show future potential.
“As far as I can tell, there are only three types of AI product that currently work.”
The Three Working AI Product Archetypes
Goedecke categorises successful AI products into three distinct types:
Chatbots:
- Description: The most popular LLM product, exemplified by ChatGPT. These allow users to interact with an LLM in natural language for a wide range of tasks.
- Challenge: The best chatbot is often the model itself (e.g., ChatGPT), which has advantages in model access and simultaneous harness development.
- Variants:
- Explicit Roleplay: Chatbots that cater to niches (e.g., adult content) that mainstream models avoid. Goedecke notes ethical concerns and the likelihood of large labs eventually entering this space.
- Chatbots with Tools (AI Assistants): These allow the model to perform actions (e.g., booking meetings). However, they are prone to “jailbreaking” for malicious use (e.g., refunds) and often provide a worse user interface than direct actions (e.g.,
Ctrl+Plusfor font size).
Completion Tools:
- Description: Products that act as “smart autocomplete,” suggesting completions as a user types. GitHub Copilot is the prime example.
- Advantage: Users do not need to interact with a chatbot, seamlessly integrating AI power into existing workflows without changing user interface habits.
- Observation: While highly successful in coding, their adoption in other areas like professional writing has been slower, possibly due to user familiarity with existing autocomplete in code editors.
Agents:
- Description: Systems where the AI takes an initial natural language request and autonomously implements and tests the solution. Coding agents (e.g., Claude Code, GPT-5-Codex) are the most prominent example.
- Distinction from Chatbots with Tools: Agents manage a sequence of actions, rather than just single tool calls, making them effective for complex tasks.
- Why Coding Agents Work:
- Verifiability: Changes can be easily validated through tests or compilation.
- Incentive: AI labs are motivated to develop effective coding models for their own use.
- Future Potential: Research agents (e.g., for skimming search results or data analysis) are emerging, with potential in specialised fields like medicine or law.
Emerging AI Product Categories
Goedecke also identifies two categories that are not yet widely successful but show significant future potential:
- LLM-Generated Feeds: Infinite, personalised content feeds generated by AI, similar to social media feeds but with AI-created content. This leverages existing user habits of scrolling feeds.
- AI-Based Video Games: Games that incorporate LLMs for content generation, dialogue, or even full world simulations. Challenges include long development cycles, gamer resistance to AI-generated content, and the difficulty of making AI-generated content engaging and challenging.
Commentary
Sean’s analysis highlights a crucial distinction between AI as a feature and AI as a standalone product. Many “AI products” are simply chatbots, which often struggle to compete with general-purpose models like ChatGPT or face usability issues when trying to integrate tools. The success of completion tools like GitHub Copilot underscores the value of AI that enhances existing workflows without forcing a new interaction paradigm.
The rise of agents, particularly in coding, points towards a future where AI handles more complex, multi-step tasks autonomously. This shift from conversational interfaces to task-oriented execution is a significant development. The discussion on “AI-generated feeds” and “AI-based video games” also touches on the broader impact of generative AI on content creation and consumption, suggesting a future where AI plays a more direct role in shaping our digital experiences.
A key takeaway is that the most successful AI products often integrate AI capabilities seamlessly into existing user behaviours or solve specific, verifiable problems, rather than trying to replace established user interfaces with chat. The Hacker News comments further elaborate on this, pointing out successful AI applications in areas like grammar checking (Grammarly), translation (DeepL), and document processing, which often operate as embedded features rather than overt “AI products.” This suggests that the most impactful AI might be the one users don’t explicitly notice, working behind the scenes to enhance functionality.
