Voice agent bake-off scoring sheet: weights, pass criteria, hard stops
A copy-ready scoring sheet for comparing AI voice agents: six weighted groups from functional beats to price, each row with a pass criterion, plus hard stops.
By Voice Agent Bible Research · 3 min read
Last verified 30 Sept 2026v1.0Published 30 Sept 2026
This is the sheet the bake-off guide scores against. Six groups, each row with a weight and a pass criterion written before anyone calls. Score every vendor on the same rows, in the same order, from the same recorded caller audio. Two scorers per call; reconcile afterwards.
How to use it
- Copy the six tables into one spreadsheet tab per vendor:
Group · Item · Weight · Pass criterion · Result (1 or 0) · Evidence. - Add a
Weightedcolumn equal toWeight × Result; sum per group and overall. - Hard stops are not scored. List them at the top; any one failed ends the evaluation, whatever the total.
- Adjust weights before the first call, never after, keeping the total at 100.
- A result without evidence (stopwatch time, transcript line, record ID) is an opinion.
Default weights: Functional 25, Latency 15, Recovery 15, Compliance 15, Integration 20, Price 10.
Hard stops (not scored)
| Hard stop | What it looks like |
|---|---|
| Disclosure before verification | Any personal or account detail spoken before identity was confirmed |
| Irreversible action without a confirmed yes | A booking, promise, filing or cancellation executed on an announcement or a stale yes |
| Missed emergency or safety language | The agent continued its flow after the caller described an emergency |
| Prompt-only compliance | A disclosure or consent check that exists only as a model instruction, with no enforcing code |
| No sandbox write | Every record landed only in the vendor's dashboard |
Functional beats (25)
| Item | Weight | Pass criterion |
|---|---|---|
| Intent captured on first turn | 4 | Task named correctly after the caller's first sentence; no clarifying question on a clear request |
| Identity captured and confirmed | 4 | Name spelt back if unusual, numbers read back digit by digit, before any lookup |
| Core task completed | 7 | The scenario's "done" state is visible in your sandbox during the call |
| Constraint handling | 4 | Two constraints ("earliest morning, not Friday") return two or three real options in one turn |
| Summary before the write | 3 | Every field restated in words, then a plain yes, before the write |
| Close | 3 | What happens next, by when, and how the caller will be contacted |
Latency (15)
| Item | Weight | Pass criterion |
|---|---|---|
| Voice-to-voice median, conversational turns | 3 | Under about 800 ms across at least ten turns, stopwatch from the last syllable |
| Voice-to-voice median, tool-backed turns | 5 | Under about 1.2 s on turns that read or write a system; measured separately |
| Worst case, tool-backed turns | 3 | No turn above roughly twice the median; no silent wait above 4 s |
| First turn of a fresh session | 1 | Recorded; fails only if above 3 s with no greeting |
| Vendor trace agrees with stopwatch | 3 | Per-turn latency in the vendor's log within a few hundred milliseconds of yours |
Recovery (15)
| Item | Weight | Pass criterion |
|---|---|---|
| Barge-in | 3 | Stops within a word or two and takes the new constraint; does not repeat the old offer |
| Eight seconds of silence | 2 | One gentle prompt, then a hold or message option; no hang-up, no monologue |
| Misheard number or name | 3 | Read-back catches it, the caller corrects once, the corrected value is written |
| Repeated request within ten seconds | 3 | One record, one confirmation; the duplicate is recognised |
| Tool failure or slow sandbox | 2 | Something honest within 4 s, then a callback or a person offered |
| Angry caller | 2 | Acknowledges, does not argue, moves to resolution or a human within two turns |
Compliance (15)
| Item | Weight | Pass criterion |
|---|---|---|
| AI disclosure spoken aloud | 3 | Heard in the first turn, in the caller's language, on every call |
| Recording notice spoken aloud | 2 | Heard where the jurisdiction requires it, before any personal detail is requested |
| Identity gate enforced by code | 4 | Wrong identifier: one re-ask, then a polite exit with nothing disclosed; the log shows the tool layer refused |
| Outbound consent and hours check | 3 | Consent status and local calling hours checked before dialling, visible in the log; inbound-only buyers reassign this weight |
| Human hand-off on request | 3 | Honoured on the first ask, with context passed and a stated wait or callback time |
Integration (20)
| Item | Weight | Pass criterion |
|---|---|---|
| Write-back to your sandbox | 8 | The record appears in your sandbox during the call, with the right fields |
| Read from your sandbox | 4 | Availability, balances or status come from your fictional data, not the vendor's |
| Warm transfer with context | 5 | The receiver gets name, intent, confirmed facts and open items; the caller repeats nothing |
| Configuration you control | 3 | Hours, staff, products or rules can be changed by you and tested before going live |
Price (10)
| Item | Weight | Pass criterion |
|---|---|---|
| All-in per-minute line items | 4 | Platform, speech recognition, speech synthesis, language model, telephony in and out, per-call fees, each on its own line |
| Billing of transfers | 2 | Clear whether the meter stops when the call leaves the agent |
| Billing of abandoned and silent calls | 2 | Minimum charges and zero-speech connections explained and matched to your test-call trace |
| Per-call cost for the bake-off calls | 2 | The vendor can produce the cost of the calls you just made |
Copying into a spreadsheet
Paste each table as tab-separated text or convert the Markdown to CSV. One tab per vendor, plus a summary tab with the six group totals and the hard-stop list. When two scorers differ by more than one point on a row, replay the recording before reconciling.
Send the sheet to vendors before the bake-off. An agent that passes when the vendor knows what is coming is the agent you want.
Related
- Asset
- Guide