Accuracy
How accurate are we? Here is the one number no one else will show you.
Anyone can grade a call. The question that matters is whether the grade is right, and the only honest way to answer it is to check every grade against what your firm actually did. That is what an independent audit is for. Here is how we measure ourselves, in public, sample size and all, including the error we are hardest on.
We grade ourselves against your own decisions
Every call we grade gets compared to what the firm did with it weeks later: signed, declined, or referred out. That gives us three numbers we can hold ourselves to.
When we say SIGN
how often the firm signs
Of the calls we grade as a signable case, the share the firm actually signed.
When we say PASS
how often the firm passes
Of the calls we grade as a decline, the share the firm passed on or referred out.
Our wrongful-decline rate
the error we watch hardest
The cases we told a firm to pass on that they signed anyway. We weight this ten times heavier than any other error, and we show it.
Still building. We do not publish a headline accuracy rate until enough cases have reached a decision to make the number mean something, and when we do, it will always carry its sample size and a confidence range. A percentage without its test set is exactly the kind of thing you have been pitched before. As the founding cohort logs real outcomes, the numbers above turn on, here, in public. That is the whole point of an independent rater: you get to watch the number accrue, not take our word for it.
No AI receptionist, agency, or CRM publishes how often it is right, let alone how often it is wrong. We do, because we are paid the same flat fee no matter what the number says.
How the grade is built
- Every recorded intake call is transcribed.
- The transcript is graded against a fixed rubric: the same rubric on every call, so grades don’t drift from one review to the next.
- A call is flagged as a signable case that walked when it grades above the threshold, wasn’t converted, and is still inside the callback window.
- Every flag carries the transcript evidence behind it, so you can check the call yourself, and a former PI paralegal reviews every score before you see it.
Where a model like this gets it wrong
Two failure modes, and we track the rate of each rather than pretend they don’t happen:
- Missed flag: a signable caller the model graded too low, usually when the caller downplays the injury or the liability language is ambiguous.
- Wrongful decline: a call we graded as a pass that turned out signable. This is the one that can cost a firm a real case, so we weight it heaviest and report it by name.
A human at your firm approves every callback, so a false flag costs a moment of a reviewer’s time, not a wrong message to a caller.
How we express confidence
Every call maps to one of five plain-English tiers. We use the same five definitions on every call, so a Tier 4 means the same thing in January as it does in June, and we never mix a confidence level with a likelihood in the same sentence.
| Tier | What we call it | What it means | Likelihood |
|---|---|---|---|
| 5 | Very likely signable | Clear liability and injury, no disqualifier heard. | 80-95% |
| 4 | Likely signable | Strong indicators, one minor open question. | 55-80% |
| 3 | Roughly even | Genuinely mixed signals. | 45-55% |
| 2 | Unlikely signable | Weak indicators, or a likely disqualifier. | 20-45% |
| 1 | Very unlikely | Clear disqualifier: prior counsel, no injury, or property damage only. | 5-20% |
We state the transcript moment behind each judgment, and we never mix a confidence level with a likelihood in the same sentence. A call is flagged for follow-up at Tier 4 and above.
How we estimate missed signable case value
How we estimate missed signable case value: we count the signable cases our model flags that didn’t sign, then multiply by your firm’s own average fee per signed case for that case type. Where you haven’t given us your average fee, we substitute a named, sourced benchmark (e.g., auto soft-tissue ~$16,000; serious injuries $55,000+) and label every substituted figure. Estimates are estimates, not a promise of recovered fees; our model’s precision and recall will be published on this page the day the test corpus is documented, and not before.
Every estimate describes value identified in your own calls, never revenue you recover. It’s a claim about what the audit finds, not a promise about outcomes.
I charge nothing until the number survives your scrutiny.