The Shift You Pay For vs the Shift Your Agents Actually Work
You staff for talk time but pay for clock time. Voicemail parking, dead air and misroutes quietly cut the shift you actually receive. The payroll math, in dollars.
The Shift You Pay For vs the Shift Your Agents Actually Work
Every outbound center runs on the same quiet gap. Payroll pays for the scheduled shift. Production only gets the fraction of that shift where an agent is talking to a live human about the offer. Everything between those two numbers is overhead, and most of it is accepted as physics: wrap-up, hold time, machine messages, dead air.
Some of that overhead really is physics. A meaningful slice of it is not. It is your dialer's answer classification spending your payroll, and it never appears on a timesheet. I run AMD detection for ViciDial, Asterisk and FreeSWITCH shops, and when we audit a new customer's floor, the recovered talk time is usually the fastest win in the building, faster than any script change or coaching plan, because it requires nothing from the agents at all.
Here is the structure of a paid shift, where the hidden leakage sits, and the arithmetic that turns it into dollars.
Anatomy of a paid agent-hour
Take a representative shop: 40 agents, $18 an hour loaded, 6.5 productive hours a day after breaks and training. Watch where one of those hours actually goes.
The official version
The schedule says: agent logs in, dialer feeds calls, agent talks, dispositions, repeats. Productivity dashboards report handle time and talk time per hour, and everyone reconciles against those.
The measured version
Sit with a stopwatch instead of a dashboard and the hour fragments like this.
| Slice of the hour | Minutes | Driver |
|---|---|---|
| Live conversation | 28 | The only slice that earns |
| Ring and carrier time | 9 | Dialer pacing |
| Answering machine messages, full listen | 7 | Classification parked machines on the agent |
| Dead air after hello | 5 | Detection latency, dropped transfers |
| Wrap-up and disposition | 6 | Process |
| System and hold gaps | 5 | Telephony and CRM |
The exact split varies, but the shape is universal. Now the important question: of those 60 minutes, which slices can a manager actually influence? Wrap-up can compress by seconds. Ring time belongs to the pacing engine. The 12 minutes of machine messages plus dead air belong to the classification layer, and that layer is the one nobody audits because it shipped with the dialer.
Where the 12 minutes come from
Machines parked on humans
When detection is loose, the safe configuration is to send everything uncertain to an agent rather than drop it. The cost of that safety is agents absorbing voicemails: the greeting plays, the beep comes, the agent waits it out or leaves a message. Fifteen to thirty seconds each, hundreds of times a day across the floor. We priced this in full in the voicemail tax on your payroll. It is the single largest reversible slice of the paid-versus-worked gap.
Hellos that reach no one
The opposite failure drops the human, but there is a middle failure too: the verdict arrives slowly, the transfer hesitates, and the agent picks up a line where the human has already heard two seconds of silence and hung up. The agent then spends the disposition coding a ghost. Detection latency is not an abstraction. A verdict that lands in about an eighth of a second, before the dialer's connect window closes, is the difference between a conversation and a code. We treated latency seriously in latency versus accuracy in AMD, because a slow correct verdict behaves exactly like a wrong one on the floor.
The retry treadmill
Every mishandled answer becomes a redial later in the day. The redials consume more paid ring time and machine time to reach the same human, at worse mood and worse conversion. You pay twice for conversations you could have had once.
The payroll arithmetic, in dollars
Convert the slices to money. At $18 an hour loaded, every agent-minute costs $0.30. Take the 12 minutes a day of machine-plus-dead-air per agent as the recoverable slice, and be conservative: say modern detection recovers two thirds of it, 8 minutes.
| Shop size | Agents | Payroll cost of 12 min/agent/day | Recovered at 8 min/agent/day | Monthly (21 days) |
|---|---|---|---|---|
| Small | 15 | $54/day | $36/day | $756 |
| Mid | 40 | $144/day | $96/day | $2,016 |
| Large | 120 | $432/day | $288/day | $6,048 |
That is pure payroll, before counting what the recovered minutes produce. Eight recovered minutes a day at a normal connect cadence is roughly one to two extra live conversations per agent per day. At even modest conversion those conversations are worth multiples of the payroll they ride on. We did this multiplication carefully in what one extra live conversation per agent-hour is worth, and the number surprises most owners in the right direction.
Read the large-shop row again. Over $6,000 a month of payroll is currently buying machine greetings and dead air, and that is the conservative cut of a conservative estimate. The aggressive cut, where stock detection is dropping or mishandling the higher end of its error range, roughly doubles it.
Why coaching cannot touch this number
The reflex when floor productivity dips is coaching. Listen to calls, retrain openings, tighten scripts. All useful, all acting on the conversation slice, the 28 minutes. None of it touches the 12.
You cannot coach an agent out of a voicemail that was parked on them. You cannot script your way past a transfer that arrives after the human hung up. When the deliverable itself is broken, polishing the delivery wastes a quarter. This is why the sequence matters: fix the classification layer, then coach the conversations that actually arrive. Otherwise your coaches are tuning an engine that never gets fuel.
The Bureau of Labor Statistics tracks what contact center labor actually costs employers, loaded, in its occupational data: https://www.bls.gov/ooh/. Pull your own loaded rate before running the tables above with your numbers. Most shops guess low, which makes the leak look smaller than it is.
The staffing illusion: headcount you hired but never receive
The payroll gap has a second face that shows up in the budget rather than the floor. When managers feel short-staffed, they hire. But if 20% of an agent's dialable time is consumed by machine messages and dead air, then a 40-seat floor delivers the live-conversation capacity of roughly 32 seats. The instinctive read is that you need eight more hires. The actual arithmetic is that you are already paying for those eight seats of capacity and receiving it as hold music and voicemail greetings.
The hiring-versus-recovery math
Run both options against each other for the 40-seat, $18-an-hour shop.
| Option | One-time | Monthly cost | Extra live-conversation minutes per day |
|---|---|---|---|
| Hire 8 agents (the gap read as headcount) | recruiting, training, ramp | $24,192 payroll + overhead | ~8 x 28 = 224 |
| Recover 8 min/agent/day via detection fix | 5-minute install | $79 to $249 tier | 40 x 8 = 320 |
Read the last row carefully. Recovery delivers more conversation minutes than the hires, because recovery mines all 40 seats rather than adding 8 seats that come with their own machine time. And the hires carry recruiting timelines and ramp months, while the recovery starts the same afternoon the install runs. Hiring to cover a classification leak is paying permanent monthly payroll for a problem that costs less than a business lunch to address.
What this does to attrition, the slow way
There is a compounding effect worth naming because floor managers name it constantly in private. Agents do not quit over a bad Tuesday. They quit over the accumulation of Tuesdays spent listening to beeps and dial tones. The job they signed up for is conversations. A shift that is half machine time is a shift that feels broken no matter what the paycheck says, and agents who feel broken interview elsewhere. Replacing an agent costs thousands in recruiting and ramp, and the replacement inherits the same broken shift unless the layer underneath changes. We touched this in the voicemail tax, but the staffing angle makes it concrete for the budget owner: the leak quietly taxes your recruiting budget, your training budget, and your overtime budget, because every one of those lines scales with a headcount that is partially occupied producing nothing.
A week in the life, before and after
To make the recovery tangible, here is a composite week from a 25-seat insurance shop, built from their own call logs before and after a detection swap. Nothing else changed: same lists, same pacing targets, same agents, same offer.
| Metric (daily average) | Week before | Week after |
|---|---|---|
| Live conversations per agent | 61 | 74 |
| Machine messages parked on agents | 118 per floor | 3 per floor |
| Calls coded as ghost connects (dead air) | 47 per floor | 9 per floor |
| Talk time as % of clock time | 46% | 58% |
| Agents hitting daily conversion target | 9 of 25 | 17 of 25 |
The conversations column is the one to study. Plus 13 live conversations per agent per day, from a change that involved zero coaching, zero scripting, zero new hires. The agents who hit target nearly doubled because the raw material of the job, humans to talk to, finally matched the capacity being paid for. That is what closing the paid-versus-worked gap looks like on a scoreboard.
The questions floor managers ask first
"Won't faster machine-drops hurt our message-leaving rate?"
No, and the confusion is worth unwrapping. Machines should still get your message campaign, voicemail drops remain a legitimate channel, and modern detection does not remove them. What it removes is the agent from the machine's path. The verdict says machine, the dialer routes to the message flow or the drop logic, and the agent never enters the transaction. Your message-leaving rate becomes a marketing decision instead of a payroll accident, which is what it always should have been.
"Our pacing engine already has AMD settings. Why would we layer another detector?"
The stock settings are timing heuristics: initial silence windows, greeting lengths, a total-word-count trigger. They predate visual voicemail, carrier early media announcements, and screening bots, and they cannot be retrained, only re-tuned by hand against today's audio, and then they decay again. GPU-backed detection classifies the actual audio and returns distinct verdicts for human, machine, screening bot and carrier announcement, in about an eighth of a second. It is a different mechanism, not a second copy of the same one.
"How do we verify any of this without trusting your brochure?"
The same way we recommend everyone verify everything: run the parallel test. Detection runs alongside stock AMD for a week, every divergent verdict gets logged, and you sample the divergences and listen. The methodology is in five minutes to install, a week to prove itself, it costs an afternoon and the free tier, and it produces a number that is yours, measured on your traffic, defensible in your Tuesday meeting. Vendors should be able to win that test on merit or lose it gracefully. Ours holds up, which is why we keep telling people to run it.
"What breaks during the five-minute install?"
Nothing that production feels. The module attaches to the dialer's answer path on ViciDial, Asterisk or FreeSWITCH, no migration, no carrier change, no downtime, and there is a fallback: if the detection service is unreachable, the dialer's stock AMD handles the call, so the floor degrades to today's behavior rather than to silence. The install is also reversible in the same five minutes, which is the honest way to offer a trial: easy in, easy out, decide on evidence.
What fixing the layer looks like
Modern AI answer detection classifies the answer before the connect decision: human, machine, screening bot, carrier announcement, each with its own routing. Verdicts land in about an eighth of a second. Machines never reach agents. Hellos transfer instantly. Screening bots get their own disposition so agents stop arguing with Google's robot, a failure mode we broke down in our Google Call Screen detection writeup.
Installation is a five-minute job for the dialer admin on ViciDial, Asterisk or FreeSWITCH. No migration, no new carrier, no agent retraining, no new screens. The agents never know it happened. They just notice the shift feels different, because the machines stopped coming.
Pricing in plain terms: free Sandbox tier of 50,000 detections a month, no card, no day limit. Paid plans start at $79 a month with 500,000 detections included, then $0.00025 per detection. Set the entry tier against the mid-shop row above: $2,016 a month of recoverable payroll against a tool that starts at $79. The ratio is not subtle.
Run the audit on your own floor
The audit needs no vendor and no budget. Pull one ordinary day of logs. Compute total scheduled clock time per agent, total live-conversation time, and the difference, then split the difference into its parts: ring time, machine time, dead air, wrap-up, system gaps. The categorization takes some judgment on the first pass. The headline number does not: minutes per agent per day between what payroll bought and what production received. Write it on a whiteboard and let it sit overnight, because the number is usually between 10 and 15, and owners tend to check it twice.
Then listen to the machine pile. Twenty calls coded as answering machines, first six seconds each, count the human greetings. If the count is zero, your classification layer is honest and your gap is pure process, which is fixable by management in the ordinary way. If the count is three or more, a slice of your payroll is being converted into hold music by software, and the section on recovery pricing above tells you what closing that slice is worth at your seat count and loaded rate.
Then decide what the gap is worth closing. The shift you pay for is fixed by payroll. The shift you actually receive is decided by your detection layer, and that decision is currently being made by default, by software nobody has audited since install day. Take it back, and the same payroll starts buying you eight more minutes of what you hired it for.