# Five styles of 'AI Product Development' > Source: > Published: 2026-08-05 10:02:15+00:00 # Five styles of 'AI Product Development' Me and lots of people I work or interact with are using AI to [build things](https://www.elliotcsmith.com/you-can-just-do-things/). The kind of things they are building vary a lot so I figured I would write down a few of the archetypes and some brief observations on each. ## Type 1 : A long running conversation with an agent This one is the most common and the simplest. You start a chat with your LLM of choice about a topic and the thread becomes the 'product'. A personal example of this one is that I have a Claude thread that's been looking at some of my running data and recommending adjustments to my training plan based on linked research. This kind of 'product' works well when the thing you want is hard to put into repeatable rules. Doubly so when its built for just one person. If my Claude thread disappeared tomorrow I would be mildly annoyed but honestly its no great loss. It's a very disposable form of a product and you could probably argue it isn't a product at all. Despite that there are no doubt many of these in existence now. I do think one important caveat to this form of product is it works well when you know enough about the thing to be skeptical. LLMs are far from flawless, even when you ask them to do research, asking them probing questions is a must. ## Type 2 : Artifacts This form is becoming more common all the time and I feel that the configuration of tools like Claude are now biased to create more. This kind of 'product' is one level above a chat thread. It's more interactive, you can share it with other people and it can function somewhat on its own. In reality these are simple small web apps but the one's I've seen and used weren't built with the goal of building an 'app' in mind. These artifact products are still highly disposable. Many of the ones I've seen or used are for data visualization or forms to collect data internally. At work I get sent these a few times a week. Both sides of the transaction are pretty aware that they're a short lived thing but I do wonder if there will be an increasing expectations that these are treated like internal apps, or at least like the fancy spreadsheets of yesteryear. ## Type 3 : Building standard products with an agent Software factories, vide-coding, agent assisted development all fall into this bucket. You're building software products but the code is increasingly written by AI. The important differentiation between this and type 4 is that I am specifically not talking about building software products that have an AI feature. You can consider that it's own type or maybe just a subtype of this one. There's a lot changing in this world right now. [Bottlenecks](https://www.elliotcsmith.com/all-the-other-bottlenecks/) have shifted and companies need to adjust. Folks are struggling to work out if they should review code or let AI do that too. Mostly though, from the outside the end product here is ideally the same as before. Ideally faster and without a compromise in quality but these kinds of products are one's we're already familiar with. ## Type 4 : Building products that use AI as a feature for humans These kinds of products (or perhaps features in existing products) are all the new and shiny things with a 'AI' icon next to them. Your product makes calls to an AI tool, probably an LLM, and the resulting response is somehow shown to your customer. To use some examples I have worked on this might be structured data extraction from PDFs, being able to ask questions about your data or searching large data sets for common themes. These products are interesting to me because they tend to break a lot of the previous pricing assumptions around SaaS. Previously the costs associated with more customer use didn't scale linearly. You may need more infrastructure eventually but for the most part your margins were high. Now that features have a linear cost to their use the margins, growth assumptions and relations to valuation are all being reassessed. ## Type 5: Building product for AI to use My original draft of this post had four styles but this one has popped up recently. These are tools specifically designed for an AI agent to interact with. I'm not saying only AI use these products but they're built with the assumption that they'll be used by an agent. To give an example of this one I am currently working on a tool that collates customer feedback and internal conversations and tries to link snippets back to work being planned. This could become a Type 4 product but it's designed to be a companion tool used when working with AI in other ways. This one in particular will be called when working on scoping PRDs. Since starting on this tool I've talked to a lot of people working on similar variants of this. The term Retrieval-augmented generation (RAG) gets used in this space but I wouldn't limit it to just databases. These tools can also fall into the broad bucket of 'give the agent a capability it doesn't natively have'. This could be farming out questions to humans or doing work in the 'real' world. I am sure this grouping will continue to evolve over time but it is interesting to me to already see how many different types of 'AI product development' exist.