Why it matters
Scoring every call produces more results than any team can open. Review plans decide which of them reach a person, and a call can be on that list for five reasons: the machine flagged it, an audit sample drew it from the confident and unflagged calls, a client mandate named it, a calibration set includes it, or it was already adjudicated and is being re-reviewed blind. The queue is what those plans produce.
What your reviewer works is a prepared case. Audio, speaker-aligned transcript, the evidence span, and the rubric question that put the call in the queue open together, with playback already sitting on the moment in question. Confirm is one key; override is one key plus a reason code. Submit commits and moves to the next call, and an undo toast stands where a confirmation dialog would.
Every override is a labeled example for the check it disagreed with, so reviewing is also the motion that improves the models. And every consequential line on an agent's scorecard has a person's confirmation and a reason code behind it, which is the answer to give a works council when they ask who graded someone.
How it works
- 01
Write the review plan
A plan holds a selector, a sampler, a cadence, and an ordering. Triage takes everything the machine flagged. An audit plan takes a small random sample of the calls it scored confidently and did not flag, which is the only measurement of what the model got wrong without saying so. A mandate plan takes what a client asked for: every save attempt on the retention campaign, or five calls per agent per month.
- 02
Set the ceiling on each pool
An admin sets what a reviewer pool may open at all, and it is enforced on every call-scoped endpoint rather than on the queue alone. A plan draws from inside that ceiling and can never widen it, so an outsourced pool has no path to another client's calls, including by call ID, search, export, or API.
- 03
Hand out a batch
Work is allocated round robin, by shortest open list, or at random, with weights for part-timers and trainees. Each reviewer holds a stable batch under a lease that expires and returns unworked calls to the pool. Reviewers are handed the next call rather than picking one, which keeps cherry-picking out of throughput numbers and out of the audit sample. Leads browse and search.
- 04
Open the prepared case
The waveform marks every flagged moment as a tick. The transcript is scrolled and highlighted at the span that triggered the flag. The verdict rail lists each rubric question with its verdict, confidence, rationale, and a link to the evidence behind it.
- 05
Confirm, override, or escalate
One key confirms the machine verdict on the focused question. One key opens the override picker, where a reason code is required and a note is optional. Escalating reassigns the call to a named role or pool with a reason, and the reassignment is an audited record.
- 06
Submit and advance
Submit commits the call and drops the reviewer into the next one. Adjudications reach reporting immediately and collect as a labeled set your deployment owns.
Technical shape
The review plan
One saved, versioned object: selector, sampler, fan-out, cadence, lane, ordering, grouping key, deadline, and maximum age before forced promotion. The selector is the same matcher expression language your rubric questions and saved searches already use, so a lead can search first and promote the result into a plan.
Queue position
Order is computed per batch from lane weight, risk, age, and deadline pressure rather than stored as a priority number. A row decomposes into those parts, so “why is this call third” has an answer, and so does “why were these forty reviewed and not those forty.” Lanes merge by weighted share, and each one carries an age ceiling that promotes its oldest work.
Review surface
Waveform with flagged-moment ticks, synchronized transcript, evidence highlights, session header, verdict rail, and a keystroke for each action: play, jump, focus question N, confirm, override, escalate, submit and next.
Decision record
Reviewer identity, timestamp, the machine result it answers, confirm or override, reason code, note, and escalation state. Undo appends a superseding event. The machine verdict itself is never edited.
Data boundary and privacy
No agent is coached off a raw machine score. Every consequential verdict has a human adjudication and a reason code in front of it.
Reviewers sign in through your identity provider and reach the workbench on your network. There is no Call1-hosted login page and no Call1 account inside the workbench: the roles are reviewer, QA lead, and admin, and Call1 employees hold none of them.
Remote reviewers arrive over a path you operate: your VPN, or a Cloudflare tunnel in your own account with your IdP in front of it. Call1 creates none of it and holds no token for it. A tunnel that terminates TLS at a third party can read what a reviewer's session carries, so the method is recorded in the deployment profile and stated in the appliance manual.
Audio plays from the local hot cache or from storage you control. A remote reviewer does not cause the recording to be proxied through a vendor application. Notifications carry a link and permitted metadata, never transcript text or audio.
Your overrides train the adapter behind the question they corrected, and that training runs on the deployment. A retrained adapter scores nothing until someone authorized on your side promotes it.
Performance
Opening a case skips the search
Timestamps and evidence spans are computed before review, so opening a call lands on the reason it was queued instead of starting a scrub through twenty minutes of audio.
Every action has a keystroke
Play, jump, question focus, confirm, override, escalate, and submit-and-next are single keys. Submit advances on its own, and a short undo window replaces the confirmation dialog. This is a tool someone runs the same loop in for a full shift.
Capacity is projected before a plan is saved
The projection runs across every plan you have against your own call volume and your team's throughput, and it shows before you commit. When the portfolio exceeds what the floor can clear, plans downsample by the precedence you declared and the shortfall is named and logged. The audit stream holds a reserved floor and is squeezed last.
The hot cache follows review value
Queued and disputed calls stay immediately playable. Older audio can expire or move to your archive, and a call opened from there takes longer to start.
What it doesn't do
- A reviewer is never assigned a call they handled. The refusal is server-side and not configurable.
- A low-confidence or high-consequence verdict routes to a person rather than publishing as final.
- Bulk actions do not stand in for adjudication on consequential questions.
- Machine verdicts and human adjudications stay distinguishable in the scorecard, the export, and the audit log.
- A rate computed over a targeted sample is not that agent's rate. The sampling method travels with the assignment so any surface reporting a rate can say which one it is showing.
- Urgent work arrives as an offered insert a reviewer can accept, so a batch already in progress does not reorder underneath them.
What to validate before deployment
- 01
Which streams do you need on day one: triage, audit, mandate, calibration, reviewer QA?
- 02
What may each pool open at all, and does an outsourced pool need scoping by end client rather than by campaign?
- 03
Which reason codes must an override offer, and is that one list or one per rubric question?
- 04
What ordering does each plan use, how do lanes share a reviewer's day, and how old may a call get before it is promoted?
- 05
What playback latency and submit-to-next time will you accept on your network?
- 06
Who may expedite a call to the front, and how is that rate reviewed?