Your scorecard, read the way your compliance team reads it.
Two advisers can read the same exclusions out in different words. One compliance officer counts a fact find from an earlier call as covering the customer's budget; another wants it repeated at the point of sale. When your compliance team corrects a verdict like that, CallGuard AI uses your most recent corrections as examples the next time it scores that criterion. Growth and Pro
See what's included on Growth and Pro, or ask us about upgrading.
A scorecard is not self-explanatory. Your compliance team already knows what it means.
An AI that only reads the words of your scorecard still has to settle judgement calls somehow, without knowing which reading your compliance team would choose. "Explained the exclusions" can mean read verbatim from the policy document to one reviewer, and summarised in the customer's own understanding to another. Both are defensible. Only one is yours.
A generic reading
Scored against the scorecard's words alone, an AI has no way to know which side of a judgement call your firm sits on — so every borderline criterion gets a generic answer, not your firm's answer.
Calibrated to your corrections
When your compliance team corrects a verdict and gives a reason, that correction — with the transcript evidence it was made on — is shown to the AI as an example the next time it scores that criterion, on a different call.
Examples for the AI to weigh, not rules for it to follow.
Correct a verdict — choose Pass or Fail, add a reason if you want to, and save — and it's stored with the transcript excerpt the AI based its verdict on, plus your reason if you gave one. The next time CallGuard scores that same criterion — on a different call — it's shown up to five of the most recent corrections on that criterion. They're shown as examples of how your firm reads that criterion, not as rules the AI must follow: it can still reach a different verdict on a call that genuinely differs.
Exemplar calls work alongside corrections. Mark a call or a sale as an example of what good looks like at your firm, and the AI is shown the first 1,500 characters of the two most recently marked exemplars — for a sale, that's its closing call — alongside any written guidance from your knowledge base: your product rules, your house style for explaining exclusions, whatever your compliance team has documented.
- Unclear consent goes to a person
- Where it is not reliably clear which speaker on a call is the customer, consent items go to your review queue and a person decides — the routing is about telling the speakers apart, not, by default, the AI's confidence in its verdict.
- Your best calls as the benchmark
- Mark a call or sale as good practice, and the AI is shown the first 1,500 characters of the two most recently marked exemplars as what good looks like at your firm.
- Coaching that remembers
- On calls scored individually, each coaching draft is written with that adviser's last three coaching drafts in view. An insights brief, generated on request over a period you choose, summarises patterns for your compliance team.
Calibrated by your team, not by itself.
Calibration is deliberately narrow about what it changes. Here's what it does not do.
Examples, not rules
Corrections shape what the AI is shown, not what it must conclude. It can still score a similar call differently if the transcript genuinely warrants it — the one exception is the same sale that was corrected: if that sale is re-scored, the correction's ruling stands.
No fine-tuning, no model per firm
Nothing is fine-tuned, and there's no separate model per tenant. Every firm's corrections, exemplars and guidance are fed to the same underlying model as context, each time it scores.
Consent routing isn't a confidence dial
The AI-confidence floor that could route uncertain scores to review is off by default. Whatever that setting says, a consent item goes to your review queue when the customer and the adviser can't be reliably told apart on the call it is judged from. On a sale, that is the call its evidence comes from.
No accuracy or agreement figure, published
We don't publish a figure for how closely calibrated scoring tracks a human reviewer's judgement. What you get instead is the transcript evidence each verdict was decided on, or a note that none was found, and any reason your reviewer gave for a correction, so your team can check it directly.
If you're weighing this against another vendor's approach to per-tenant scoring, see how we compare to Aveni and Recordsure, or browse the full comparison hub.
Questions about scoring calibration, answered.
What happens when my compliance team disagrees with a verdict?
They choose Pass or Fail, add a reason if they want to, and save. That correction is stored with the transcript excerpt the AI based its verdict on, plus the reason if one was given. The next time CallGuard scores that criterion on a different call, it's shown up to five of the most recent corrections on that criterion, as examples of how your firm reads it, not rules it must follow: the AI can still reach a different verdict on a call that genuinely differs.
Which plans include scoring calibration?
Growth and Pro. Starter doesn't include calibration from corrections, exemplar calls or coaching memory. See what else each plan includes on the plans page.
Does CallGuard ever send a checkpoint to a person on its own?
Yes, in two situations, both driven by speaker attribution rather than the AI's confidence. When CallGuard can't reliably tell which speaker is the customer, a consent question goes to your review queue. When a sale's main call can't be attributed at all, every checkpoint on that sale goes to review. Routing works per checkpoint, not per call. There's also a threshold on the AI's own confidence, but it's off by default.