> ## Documentation Index
> Fetch the complete documentation index at: https://docs.brilo.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Choose your AI model
> Pick the AI model, the transcriber, and the thinking effort for your Brilo agent, and understand the trade off between how fast it answers and how well it reasons.
Three settings under **Models** in your agent's **Advanced** tab decide how your agent thinks and how it
hears. Most businesses never need to touch them, and the defaults are chosen to be fast enough for a
phone call. Change them when you have a specific problem: replies that are too slow, or reasoning that is
not good enough for a complicated conversation.
## Before you start
* An agent that has taken real calls, so you have something to compare against.
* Ten minutes, and a phone.
* A specific complaint to fix. Changing models without one usually makes things worse.
* You will be working in the **Models** section of your agent's **Advanced** tab.
## Steps
Open your agent, go to **Advanced**, and expand **Models**.
Click **AI Model**. The list is grouped by provider and filtered to the models available in your
workspace region, with a short description of each.
This is the model that decides what your agent says. A larger model reasons better and answers more
slowly, and on a phone call the delay is the thing callers notice.
Click **Transcriber**. This is what turns the caller's speech into text before your agent reads it.
If your problem is that your agent misunderstands callers rather than answering badly, this is the
setting to change, not the model.
**Thinking effort** runs from **Off** through **Low**, **Medium**, **High**, **Extra High**, to
**Max**. Higher effort lets the model reason more before replying, and every step makes replies
slower. **Off** is the fastest.
On voice calls the available range is capped, because a caller will not wait. Only the levels a model
can sustain in a real time conversation are offered.
Save, then call your agent and run through your most complicated scenario. Time the pause before it
answers. See [Test your agent yourself](/go-live/test-your-agent-yourself).
## You will know it worked when
Your agent answers your hardest question correctly, and the pause before it speaks is short enough that
you do not wonder whether the line has dropped.
## The trade off, in one table
| Symptom | What to change |
| ------------------------------------ | --------------------------------------------------- |
| Long pause before it answers | Lower **Thinking effort**, or choose a faster model |
| Gets confused by multi step requests | Raise **Thinking effort** one step |
| Mishears words and numbers | Change the **Transcriber**, not the model |
| Answers the wrong question | Neither. Fix the instructions and knowledge |
| Invents facts | Neither. Add the facts to knowledge |
The last two are worth reading twice. Most problems that feel like the model being not clever enough are
actually a missing document or a contradictory instruction, and no model setting will fix those. See
[Why it answered wrong](/knowledge/why-it-answered-wrong).
## How to change these safely
Change one setting at a time and call your agent after each change. Model settings interact, so changing
the model and the effort together tells you nothing about which one helped.
If your agent is live, use a version rather than editing in place, so you can compare properly and switch
back in one click. See [Save and switch agent versions](/agents/versions), and
[A/B test two versions of your agent](/agents/ab-test-your-agent) if you want the two running side by
side on real traffic.
## What else you can do here
* Control recordings and how long calls are kept, in the **Privacy** section on the same tab. See
[Recordings, transcripts and retention](/compliance/recordings-and-retention).
* Reduce the pause without touching the model, by shortening what your agent says. See
[Give your agent a personality](/agents/give-your-agent-a-personality).
* Tune the pause your agent leaves before replying, which is a different delay entirely. See
[Pauses, silence and interruptions](/calls/pauses-and-silence).
* Fix specific words it mishears rather than changing the transcriber. See
[Teach it names and terms](/knowledge/teach-it-names-and-terms).
## If it did not work
**Replies got slower and no better.** Put **Thinking effort** back down. Effort buys reasoning on
genuinely hard problems and costs you latency on every ordinary call, which is a bad trade for a
receptionist.
**The model list is short or empty.** Models are filtered to your workspace region. If the list is empty,
your region has none available and support can tell you what is coming. Email
[support@brilo.ai](mailto:support@brilo.ai).
**You cannot select the effort level you want.** Voice calls cap effort at whatever the model can sustain
in real time. The missing levels are not available for voice, by design.
**It still misunderstands callers.** Change the transcriber, turn on **Advanced Noise Cancellation**, and
add the words it gets wrong to your vocabulary. See
[Sound quality and background noise](/calls/sound-quality).
**It still gets facts wrong.** That is knowledge, not the model. See
[Why it answered wrong](/knowledge/why-it-answered-wrong).
## FAQ
The default, unless you have a specific complaint. Open **Advanced** and expand **Models** to see what
your workspace region offers, with a description of each. Change it only when replies are too slow or
your agent cannot follow a complicated conversation.
Usually **Thinking effort** set higher than the call needs, or a long answer that takes time to
produce. Lower the effort a step, and set **Answer Length** to **Concise**. Note this is separate from
the deliberate pause your agent leaves after you stop talking, which is
[Patience Level](/calls/pauses-and-silence).
It gives the model more room to reason before it replies. Higher effort means better handling of multi
step requests and slower responses. **Off** is fastest. On voice calls the range is capped to what can
be sustained in a live conversation.
It converts the caller's speech to text before your agent reads it. Change it when your agent
misunderstands what callers say. If it understands fine but answers badly, the transcriber is not your
problem.
No. Invented facts come from missing knowledge, and a bigger model invents more fluently. Upload the
facts and tell your agent never to quote anything not in its knowledge. See
[Why it answered wrong](/knowledge/why-it-answered-wrong).
Not today. You choose from the models Brilo offers in your region. Email
[support@brilo.ai](mailto:support@brilo.ai) if you have a requirement that needs it.
Change it on a version rather than in place, so you can compare and roll back in one click. See
[Save and switch agent versions](/agents/versions).
## Related
* [Recordings, transcripts and retention](/compliance/recordings-and-retention) for the Privacy section
* [Save and switch agent versions](/agents/versions) to change models safely
* [Pauses, silence and interruptions](/calls/pauses-and-silence) for the other kind of delay
* [Why it answered wrong](/knowledge/why-it-answered-wrong) for problems the model cannot fix
* [Sound quality and background noise](/calls/sound-quality) for callers being misheard