AI - chat or more?
Intro
Most of us know AI through chat clients and web and mobile apps, yet many people only use the chat. Usually because there is no time, or it just doesn't feel convenient. Below are five types, from simplest to most complex.
By "AI," I mean LLM chat clients like ChatGPT, Claude, or DeepSeek. I will take an everyday user's perspective on using these systems to gain knowledge or speed up day-to-day work.
Copilot is familiar, since many companies roll it out across the Microsoft Office suite. You have probably tried an app or web client. There is usually a next step worth knowing.
Types of AI use cases
Looking at the broad picture of how we can use AI, I distinguish the following types:
- Chat text
- Chat + doc as base
- Projects (partial RAG)
- RAG
- Agent AI
Quick Comparison Table
| Use Case | Best for | Requires set‑up? | Example | Cost |
|---|---|---|---|---|
| Chat text | Quick Q&A | No | “Explain recursion” | free |
| Chat+doc | Analyzing one file | No (upload) | Summarize a PDF | free |
| Projects | Ongoing work | Yes (folder) | Writing a report from your own notes | free/paid |
| RAG | Many documents | Yes (index) | Company policy bot | free or paid (check current price) |
| Agent | Task automation | Yes (permissions) | Email follow‑ups | free or paid (check current price) |
Chat
The easiest way is to visit chatgpt.com, claude.ai, or deepseek.com directly and use their chat interfaces. If one of those platforms interests you, you can install a mobile app for easier access.
You can also use visual models to generate images by describing them.
Chat with 'documents'
This is the next step in using AI: you can upload a document to the chat and talk with the AI. The chatbot can analyze your document and add comments. Some chatbots can also search the web, so ask them to check facts, and verify important answers yourself. This could be a presentation to be created or an official document that is hard to understand.
Another example of such a 'document' is a URL to a video, which can be analyzed and summarized. The user can then chat based on that summary.
Before you upload a file, check whether it contains personal data, customer data, or company information. Some providers may use your inputs to train their models, unless you turn this off in the settings. Check the privacy settings and terms of the provider first. In Germany and the EU, personal data is covered by the GDPR (DSGVO). For work documents, check your employer's or client's rules before you upload anything.
Projects
This sounds technical, but the idea is simple. Developers often use AI on their own code to check best practices and suggest changes. Google launched an interesting tool called NotebookLM. It allows users to create an online project, upload many files or links as sources, and get details based on them.
The advantage (for some) here is that answers are mostly based on the sources you upload, not just general internet data. Those summaries can even be generated in the form of a podcast.
RAG (Retrieval-Augmented Generation)
This approach goes a step further. I explain it here for orientation. The user starts the AI and provides access to their own documents, like an Obsidian vault with mixed notes. They can be analyzed at once or left until needed.
With such an approach, it is also good to set up an Agent 'Persona' (more on that: Agent.md). Such a file instructs the AI about its role, how the user expects to receive answers, where the files are located, and what the structure is. This is already moving into the next step, but it can be kept as is.
At this level, we can already play with different desktop solutions from AI vendors (Codex, Claude, Gemini, etc.) and OpenCode, which allows users to select their preferred model provider and, by using API keys, access one or many models depending on their needs.
Agent AI
Here we reach a deeper level when working with large amounts of data. This level becomes necessary when there is a need for specialized agents for certain tasks. Some agents can read data and inform the user of what can be done, similar to standard chat, while others can have access to user data and create, write, delete files, or even perform system actions. For example, OpenClaw. A great write-up is available in the OpenCode documentation: Agents
Agent tools can read and change files. Give them access only to folders you can restore.
Useful resources at this stage include model comparisons that provide model accuracy for specific tasks as well as cost data.
Cost
Cost will vary depending on the user. We can stay on free plans across all the above solutions. This can be limiting, as access to the best-performing models is usually paid. However, for simpler tasks, free tiers can be fully sufficient.
Paid plans often cost around 20 USD per month (October 2026, check the current price). However, solutions like OpenRouter or OpenCode Zen offer pay-per-use pricing, which can help limit spending.
Conclusion
If you've reached this note, I assume you are into LLMs and AI and are likely already aware of these use cases. If, by any chance, this is not the case, I encourage you to test these tools.
Try a free tier first. Pick one task from your week and test it without sensitive information.
If you want the same steps for your whole team, start with a short review of your processes. See Workflow Automation.