When people say "I trained my bot," they usually mean something different from what the engineers mean. Both kinds of learning are real. Only one is yours.
Training: the schooling you didn't pay for
A model is trained before you ever use it. The company behind it feeds it huge amounts of text and tunes it. That's where it gets its general knowledge, its writing skill and its ability to reason.
You don't change that. It's a bit like hiring someone with a degree. You don't get to redo their schooling. You get what they walked in with.
Training also has a date on it. xAI's docs list a knowledge cutoff for each model, for example May 2026 for Grok 4.7 (docs.x.ai). After that date, the model only knows what it's told or what it looks up.
Instructions: the onboarding you do
The second kind of learning is yours. You tell the bot:
- what its job is
- how you want things done
- what it must never do
- where the files are
That's onboarding. In an organization, this is the job description, the procedures and the work instructions. For a bot, it's written instructions, playbooks and skills.
| Training | Instructions | |
|---|---|---|
| Who does it | The model's maker | You |
| When | Before you start | Every day |
| Can you change it? | No | Yes |
| ISO parallel | The education a hire brings | Training and work instructions on the job |
Teaching sticks only if it's written down
Here's the catch. If you correct a bot in a chat, that correction lives in the chat. When the chat ends or gets shortened, it can disappear.
A good bot notices this. When you correct it, it asks some version of: "Should this be the new default?" If yes, it writes the change into its standing instructions. Now it's learned for real. See Fix the cause, not the answer.
A real example
Our AN Creation office bots run on a set of skill files: one for saving records, one for the project manager role, one for shared principles. When a rule needs to change, the skill file changes, and it gets a new version number. The bot didn't get "retrained." Its work instruction got updated. That's the kind of learning you control.
What this means for you
- Don't wait for the model to get smarter at your business. Teach it, in writing.
- Keep your instructions short and specific. A model already knows how to write a letter. It doesn't know your customer's name.
- Review the written instructions now and then. They're the part of the bot that's truly yours.
Training gives you capability. Instructions turn it into a worker.