Three Common AI Product Patterns
PMs often hear three terms:
<ul>
<li>RAG: retrieve relevant information and provide it to the model at runtime</li>
<li>fine-tuning: further train or adapt a model for more consistent behavior on a task</li>
<li>tool use: let the model call external systems or functions to fetch data or take actions</li>
</ul>
They solve different problems. Confusing them leads to weak product decisions.

Users want an assistant that answers questions over a frequently changing internal policy library. Which approach is usually the best first fit?

Fine-tuning is usually the best first answer when the knowledge changes daily.

Tool Use Is for Actions and Live Data
If the product needs to:
<ul>
<li>check account status</li>
<li>create a ticket</li>
<li>send an email</li>
<li>update a CRM record</li>
<li>trigger a workflow</li>
</ul>
then the assistant often needs tools or integrations, not just a prompt.

A customer asks, "Can you check my order and start a refund if it has not shipped yet?" What capability is most important?

RAG helps ground answers in retrieved information instead of relying only on model memory.

Where Fine-Tuning Helps
Fine-tuning can help when prompting alone is not consistent enough and the product needs:
<ul>
<li>a very specific response style</li>
<li>a custom classification taxonomy</li>
<li>repeated formatting behavior</li>
<li>domain-tuned patterns of output</li>
</ul>
It is not usually the first tool for fresh private knowledge.

A support platform needs extremely consistent classification into a custom ticket taxonomy, and prompting baselines are still noisy. What might help next?

Tool use often needs permissions, validation, and retry or error-handling logic.

Many Good Products Combine Patterns
A strong AI product might use:
<ul>
<li>RAG for current knowledge</li>
<li>tools for action-taking</li>
<li>prompting for behavior and structure</li>
<li>fine-tuning only where consistency gains justify the effort</li>
</ul>
The PM job is to choose the simplest architecture that reliably solves the user problem.

An HR assistant must answer handbook questions and also file PTO requests. What is the best pattern?

PMs should start with the simplest architecture that can reliably solve the task, not the most technically elaborate one.