Fine-tune Riley on your best calls

Fine-tuning teaches Riley from your own shop's best calls. You grade real calls, start a training run built from those grades, and then compare the fine-tuned Riley against the baseline on your Analytics page. Over time, Riley sounds less like a generic receptionist and more like your best day on the phones.

Step 1: grade your calls

Training data comes from the quality labels you put on real calls. On any call's detail page (open it from /dashboard/calls) you'll find the quality rating control with five labels:

• Exemplar: exactly how you want every call handled. These are the gold examples training learns from.
• Good and Acceptable: fine, unremarkable calls.
• Poor and Bad: calls that went wrong. Marking these teaches training what to avoid.

You can add optional notes with each grade. The more calls you grade - especially at the Exemplar and Bad ends - the more signal a training run has to work with. Grading a handful of calls a week is enough to build a useful set.

Step 2: start a training run

1. Go to Fine-tune in the sidebar (/dashboard/shop/training).
2. Choose a provider from the provider select.
3. Click Start training.

Your run appears in the job status list on the same page. Training runs move through statuses: Queued, Training, Ready, and occasionally Failed or Paused. The same statuses are visible in the fine-tune panel on Shop config's Advanced (AI) tab - though that tab is flagged for team use only, so there's no need to change anything there.

Step 3: compare against baseline in Analytics

Once a run is Ready, open Analytics (/dashboard/analytics) and look at the A/B baseline vs fine-tuned comparison. It shows how calls handled by the fine-tuned Riley perform against the baseline, on your real traffic. Let the comparison accumulate calls before judging - a single afternoon of data is noise, a couple of weeks is a verdict.

Good habits

• Grade first, train second. A training run is only as good as the grades behind it. If you've barely graded anything, spend a week labeling calls before starting a run.
• Be stingy with Exemplar. If everything is an Exemplar, nothing is. Reserve it for calls you'd genuinely play to a new hire as "do it like this".
• Grade Bad calls too. Negative examples matter as much as positive ones.
• Pair with mystery shops. After a run goes live, dial a few mystery shop scenarios to hear the difference and get it scored on the 12-point rubric.

If a run shows Failed, just start another - and if it fails repeatedly, we're happy to look at it with you.

Still stuck? Submit a ticket from this help center and we'll take it from there.