Hear the question
The transcript must show the exact claim-specific question. A plausible answer to an unasked question is held.
Evidence before automation
Kol uses CALL-E to chase claim status, remembers the phone-tree route, and blocks every result that cannot prove what was asked, who answered, what they said, and how the call got there.
Public demo uses fictional claims. No EHR writes. Live calling is PIN- and destination-gated.
Claim CLM-4471
$1,240 paid · Aug 12, 2026
Question witness
Claim-specific question found
Destination witness
Claims status department
Answer witness
Amount and full date grounded
Route witness
Fixture receipt: 2 → 1
Decision
All independent witnesses agree. Safe for a biller to review.
25 min
Manual claim-status inquiry
CAQH 2024 Index
$11B
Annual medical claim-status spend
CAQH 2024 Index
2,000
Deterministic adversarial verdicts
720 single-claim · 1,280 multi-claim
Not another voice bot
A correctly shaped answer can still be wrong. Kol turns each field into a claim that must survive independent checks before it moves downstream.
The transcript must show the exact claim-specific question. A plausible answer to an unasked question is held.
Status, amount, full date, denial code, and department each need payer-side evidence.
Model-reported keys must agree with a fixture log, decoded DTMF audio, or independent provider event.
Route Atlas
Kol replays a reviewed IVR route only while the live prompts still match. Drift quarantines the map and forces exploration next time.
Versioned payer routes
Every correction creates a new, inspectable route version.
Fail-closed drift
A remapped key never silently teaches the atlas.
Measured time, honest scope
Exploration and replay timings stay separate; no invented savings.
Replay the fictional evidence lab, inspect route drift, then place one controlled test call when you are ready.