How the checks work
KPI Shield compares the calling activity your dialer reports with the call evidence and reports where they disagree: reported activity that the call evidence does not support. It shows the evidence for each finding. People decide what happened.
What “verified” means here: supported by the available evidence and the current analysis method. It is not a guarantee: each finding shows its evidence, what could not be checked, and whether a person has reviewed it.
1. Look at the call before the rep
Every call is first sorted by how it ended on the phone network: failed, busy, rang out, ended while ringing, a machine or menu answered, or answered with 10 seconds or more of talk (call length alone doesn't show who answered). Carrier outages and caller IDs flagged as spam are detected across the whole team and never count against anyone.
2. Compare with real peers
Rates are compared with reps on the same team and dialer mode, because a power-dialer team and a manual team look different on ordinary days. Small samples are pulled toward the team's typical rate, so a rep with 40 dials in a week can't look extreme by chance. Nothing is flagged below 50 attempts a week unless the check counts something that should never happen.
3. The checks
| Check | What it looks at | Other explanations it rules out | Needs |
|---|---|---|---|
| Very short dial attemptsDIAL-01 | Dials the rep ended within seconds of ringing, compared with peers on the same team and dialer mode. | Outages, spam-flagged caller IDs, busy signals, voicemail pickups and power-dialer pacing. | call records |
| Burst dialingDIAL-02 | Bursts of 4+ very short dials within two minutes, or long runs of them. | Dialer-paced sessions and voicemail hang-ups in a row. | call records |
| Rapid redial of the same numberDIAL-03 | The same number dialed 3+ times within 10 minutes with no answered call of 30 s+. | Callbacks, busy lines being retried, and owners of several accounts. | call records |
| Repeated attempts to one numberDIAL-04 | One number dialed many times across separate sessions in a week, rarely answered. | Live deals, busy-line retries and shared business lines. | call records |
| Dials to numbers already reported disconnectedDIAL-05 | Dials to numbers the phone network already reported as disconnected. | Numbers that later reached a person, and outage windows. | call records |
| Calls to own, colleague or company numbersDIAL-06 | Calls to the rep's own line, a teammate's line or company numbers logged as prospecting. | The dialer ringing the rep first, warm transfers, calls logged as internal, and a couple of short test calls. | call records |
| Calls logged by hand with no dialer recordDIAL-09 | Calls logged by hand with no dialer call to the same number around that time. | Double logging of a dialer call, dialer-outage days and team-wide off-dialer calling. | call records + hand logs |
| Automated or scripted dialingDIAL-11 | Gaps between a rep's dials that land on one value far more often than peers' do: a script, an auto-clicker, or a list clicked through on a fixed rhythm. Shown with how many of those dials ended before anyone could answer, and whether dial lengths were as regular as the gaps. | Power, parallel and predictive dialer pacing, a shared auto-advance delay, a steady wrap-up after real conversations, and a steady rhythm of dials that rang long enough to be answered. | call records |
| Overlapping calls from one repDIAL-12 | Calls that start while another call on the same login is still live. | Parallel and predictive dialing, transfers within one call, and clock skew of a few seconds. | call records |
| Reported conversation inconsistent with call evidenceDISP-01 | Calls reported as a conversation that the call evidence doesn't support: the network shows no answer, busy or a machine, or the transcript has only automated audio. | A disposition set on the next call to the same number, and very short conversations. | call records |
| Leads written off without reaching anyoneDISP-05 | Unanswered calls closed out as wrong number, bad number or not interested, compared with peers; numbers a teammate later reached are shown. | Disconnected numbers the network confirmed, and people who said no. | call records |
| Meeting-bridge calls logged as conversationsTALK-06 | Long calls to meeting dial-in numbers (or bridges named in the transcript) logged as reaching a prospect. | Internal meetings and no-shows logged as such. A calendar isn't connected, so ask the rep which meeting it was. | call records |
| Calls between reps logged as prospectingTALK-08 | Calls to teammates' lines logged as prospecting, especially when both reps do it the same week. | Role-play and coaching logged as internal, and warm transfers. | call records |
| Recording paused or offEVADE-01 | Answered calls with recording paused or off, as a share of talk time. These calls are also invisible to transcript checks. | Prospects in all-party-consent states who declined recording. Only dialers that report pause or off states are checked. | call records |
| Call lengths bunched just past a known thresholdEVADE-05 | Call lengths bunched just past a threshold the team knows (60 s by default), compared with the rep's own curve and with teammates. | Script lengths the whole team shares. | call records |
| Extended automated connection (phone menu)TALK-01 | Long phone-menu calls with almost no rep speech and no live reply in the transcript, far beyond what it normally takes to reach someone. | Menus that lead to a person, and verticals where long menus are normal. | transcripts |
| Extended automated connection (hold queue)TALK-02 | Long waits in toll-free hold queues for numbers that aren't on any customer record. | Customers' own toll-free lines and internal or vendor calls. | transcripts |
| Long connection with no transcribed speechTALK-03 | Calls that stay connected for 15+ minutes after the transcribed speech stopped. | Meeting bridges, screen-share demos, holds where talk resumes, and platform session timers. | transcripts |
| Line open after voicemail greetingTALK-04 | Voicemails where the rep left no message but kept the line open after the greeting (measured from transcript timing). | Real messages after the beep and long greetings. | transcripts |
| Short answered call logged as no answerDISP-02 | The transcript shows a live hello, then no rep speech and a hang-up within two seconds, and the call is logged as no answer. | Answering machines, system hang-ups and power-dialer abandons. | transcripts |
| Live greeting with no rep reply, or a 'bad line' endingDISP-07 | The transcript shows a live hello and no rep reply before the other side hangs up, or the rep says the line is bad and ends the call. | Audio faults that hit several reps on the same number at the same time, and the dialer's connect delay. | transcripts |
4. From finding to documented review record
Signal
A pattern check that found something on one day only, or below the minimum sample. Not queued for a manager.
Review required
A core rule's finding on a call, such as a reported conversation the call evidence doesn't support or an extended automated connection. For the pattern checks: a strong finding spread over several days, the same pattern in an earlier week, two different kinds of pattern, or supporting evidence on the same calls. Queued for a manager's review.
Documented review record
Supporting evidence, an acknowledged monitoring notice, the rep's side heard, a comparison with peers and a second reviewer who agrees. It records how the review was done. No finding decides motive, and none authorizes an employment action on its own.
A reviewer records a factual decision on each finding (confirmed, dismissed or inconclusive) and, separately, the business context. Confirming an observation is not a finding of misconduct. Expected patterns, such as a clinic's phone tree, can be suppressed for a destination, rep or team, with an owner, a reason and an expiry; a suppression hides alerts and deletes nothing. Decisions on the pattern checks move a bounded per-check weight; no model learns to decide on its own.
5. Where AI is used
Claude (Haiku 4.5, and Sonnet 5 for a second look) labels what happened on selected calls, for example “phone menu” or “voicemail greeting only”, and writes neutral case summaries from the findings. Emails, phone numbers, addresses and names in common phrases are masked first. It reads the whole redacted transcript when it fits, otherwise the first three minutes and the last minute. It is never asked about emotions, tone, honesty or intent, and answers that go there are rejected. If a call turns out to be personal, only that label is kept and the transcript is deleted.
6. Limits
- Some dialers don't report when a call was answered. Then ringing and talking can't be told apart, and checks that need ring time fall back to call length and status.
- Calls logged by hand can only be checked with a connected dialer that separates dialer records from logged activity (Salesloft does). A CSV export can't show this.
- Low-volume patterns are hard to see: very short dial attempts at around 5% of a power-dialer rep's dials, or two phone-menu calls a week, stay inside the normal range and are often not flagged.
- Transcript checks need your dialer's transcripts. Speech-to-text makes mistakes and speaker labels can be wrong, so talk is measured from speech timing, not word counts. A gap in a transcript is unknown time, not silence: without audio, it doesn't show that the line was quiet.
- There is no calendar connection yet, so a screen-share demo that looks like an open line is marked "not checked" unless the talking resumes before the end.
- So far the checks have been tested on synthetic call data built from published research and field observations. Accuracy on real teams is measured with design partners, who blind-review cases before we quote any numbers.