Choose the model for a bot
Set per bot, on purpose. A deflection bot and a qualifier do not need the same thing — and the more capable model is not always the better one.
Model choice is per bot rather than per workspace, which is a small decision with a genuinely counter-intuitive answer.
The instinct is to pick the most capable model available. But a more capable model is also more willing to improvise, and for a bot whose job is to stay on the rails and fetch a human when unsure, willingness to improvise is a defect rather than a feature. On voice there is a second cost: a slower model produces pauses, and a pause on a phone call reads as the line going dead.
Steps
Find the model setting on the bot
Per bot, not per workspace — which is deliberate, because a deflection bot answering five FAQs and a qualifier handling an open-ended sales conversation have different requirements.
There is no single right model for your business. There is a right model for each job.
Match it to the job
Answering the top five FAQs from a knowledge base rewards consistency. Qualifying a complex lead across an unpredictable conversation rewards reasoning. These are different requirements and there is no reason to buy both everywhere.
Prefer the smaller model for anything scripted
Deflection bots, appointment reminders, order status. Staying on script is the whole job, and a smaller model is often better at it as well as faster.
Watch latency hard on voice
Anything you gain in answer quality can be lost several times over in people hanging up. Test on a real call, not in a chat window — the tolerance for silence is completely different.
Change one thing at a time
If you change the model and the instruction together and the bot gets worse, you have learned nothing and now have two suspects.
Check it worked
Run the same five real questions past the bot before and after a model change, and read the answers side by side. If they are indistinguishable, keep the cheaper, faster one.
The parts people get wrong
- Model choice does not change your calling-minute or AI-message allowance. Those are counted the same way regardless.
- If a bot invents things, consider a smaller model before a bigger one. Improvisation is the problem you are trying to remove.