In part one we covered what the AI vocabulary actually means. This part is about doing something with it. Almost all practical AI use comes down to three patterns: a prompt, a workflow, or an agent. Understanding the difference is the whole game, because picking the wrong one is how projects get expensive, risky, or quietly abandoned.
Key Takeaways
- A prompt is a single instruction. A workflow is a fixed recipe of steps. An agent is a goal plus the freedom to figure out the steps itself.
- Workflows are predictable, testable, and safe. Most organizations should start there.
- Agents are more powerful and more autonomous, which means they need guardrails. Use one only when the steps genuinely vary case to case.
- There are so many AI models because no single model is best at everything. The smart move is to route each job to the model that fits it.
- Match the tool to the task. The least autonomous option that does the job is usually the right one.
The prompt: one clear instruction
A prompt is simply the instruction you give an LLM. "Summarize these case notes in three bullet points for a supervisor" is a prompt. The better and more specific the prompt, the better the output, which is all "prompt engineering" really means. A strong prompt usually carries four things: a role (who the AI should act as), a task (the specific job), a format (what the output should look like), and context (the material and any constraints). The difference is real. "Summarize this" gets you a shrug. "You are a case-management supervisor. Summarize the notes below in three plain-language bullets for a busy director, and flag anything urgent at the top" gets you something you can use.
The workflow: a fixed recipe
A workflow is a fixed series of steps where AI handles one or more of them. The path is predefined by a human. The AI does not decide what happens next, it just does its assigned job inside the assembly line. A new intake form arrives. Step one, the AI summarizes the narrative fields. Step two, it suggests a program match. Step three, it drafts a welcome email. Step four, a human reviews and clicks send. Every intake follows the same path. Because the path is fixed, you can test it, audit it, and trust it to behave the same way every time. For regulated or high-stakes work, predictable beats clever.
The agent: a goal, not a script
An agent is an AI that is given a goal and a set of tools, and then figures out the steps on its own. Instead of following a recipe, it decides what to do, does it, looks at the result, and decides what to do next. Tools can include searching the web, querying a database, updating records, or sending messages. Give it "find every client due for a 90-day follow-up, check which manager they belong to, and draft a personalized check-in for each," and it works out the order of operations itself. That power comes with a cost: an agent acts on its own, so you decide in advance what it may touch and what it may do without a human's sign-off. Give it permission to draft, not to send. And if the task is the same every time, use a workflow instead. It is safer and cheaper.
A workflow is a fixed line of steps a human laid out in advance. An agent is given the goal and decides each next step itself, choosing tools and looping until the goal is met. Same building blocks, very different control.
Why are there so many AI models?
Open any AI tool and you meet a menu of names: GPT-4o, Claude, Gemini, Llama, Mistral, and a dozen more, each with versions that change every few months. There is a simple reason. No single model is best at everything. One writes a warm client letter beautifully but is slow and pricey. Another is lightning fast and cheap, perfect for quick drafts, but shallow on hard reasoning. Another is built for code, another for reading scanned documents, another for fast translation. Choosing one model for every job is like asking a single employee to be your writer, your accountant, your translator, and your analyst.
There are also two camps. Closed models (ChatGPT, Claude, Gemini) live on the maker's servers, and your data travels to them. Open-weight models (like Llama and Mistral) can be downloaded and run on hardware you control, which is the whole reason private, in-house AI is even possible. The practical answer is not to marry one model forever. It is to have something choose the right model for each request, and quietly swap in better ones as they appear. Different models are different players. You would not send your goalkeeper up to take the free kicks.
Because no single model wins at everything, the smart move is to match each job to the model that fits it. A router that "picks the right model" is how good private-AI platforms spare you from tracking model names at all.
Which one do I need?
Most decisions come down to a single table. Pick the least autonomous option that gets the job done.
| You want to... | Use this |
|---|---|
| Predict a number or risk score from historical data | Machine learning |
| Summarize, draft, translate, or answer questions about text | An LLM |
| Answer questions from your own documents and policies | An LLM with grounding (RAG) |
| Run the same multi-step process reliably, every time | A workflow with AI steps |
| Hand off an open-ended goal and let the system figure out the steps | An agent |
| Keep sensitive client data off public AI tools | Protected or private AI |
That last row is the one that trips up regulated firms, and it is where this series is headed next. If your data cannot go to a public chatbot, you need a different home for all of this. That is part three.