Revenue You Can Recover Before Your Next Campaign

There is a change you can make to your outbound floor that costs one coffee break to install, requires no carrier negotiation, no hardware, and no retraining, and it addresses the single most expensive silent failure in outbound: live people answering your call and getting dropped because the dialer mistook them for voicemail. AMDY's measured figure for how often default detection does this is 10 to 20% of live answers. If you are planning next month's campaign anyway, this is the cheapest incremental revenue available to you, because you have already paid for every one of those contacts.

I am not going to sell you on faith. I am going to give you the before-and-after picture an owner actually needs, a one-week test you can run on your own traffic, and the exact numbers to look at when deciding whether to keep paying. The free tier caps at 50,000 detections a month, so the worst case of running this experiment is that you learn your current detection is better than ours and you spent nothing to find out.

What Changes on the Floor, From the Owner's Chair

Let me describe the change in operational terms first, because owners have been burned by "quick installs" before.

Your dialer admin runs one command. That is the entire installation. It takes about five minutes. Nothing about your telecom setup changes: same carriers, same numbers, same trunks. No server purchase, no seat licenses, no weekend maintenance window. Nothing on the floor changes for your agents, and they do not notice the difference, which is deliberate, because agent behavior changing mid-test would ruin your measurement.

What is actually happening under that one command, in one plain sentence: calls that get answered are now analyzed by an AI classifier that decides person-versus-voicemail, and it returns its verdict in 125 milliseconds, one-eighth of a second, fast enough that routing feels instant to both the prospect and the agent.

The consequence of that verdict arriving fast and correct is simple economics. Live answers reach agents instead of dead air. Voicemails still get skipped, so agent time is not wasted. The two failure modes of old detection, hanging up on buyers and parking agents on greetings, both shrink. You can read the full engineering story of why speed and accuracy have to be traded against each other here: https://amdy.io/blog/amd-latency-vs-accuracy.

The Before Picture, In Your Own Reports

Before the change, this is what your data looks like, whether or not you have ever framed it this way:

• A chunk of answered calls sits in the "answering machine" bucket. Some are machines. Per AMDY's measured estimate, 10-20% of the live humans are in there too.
• Agents burn part of every hour listening to voicemail greetings the old detector passed through as "probably human."
• Your contact rate, the percentage of dials that become real conversations, has a ceiling you have learned to live with and blame on the list.

The After Picture

After the switch, three things move, and you can verify all three yourself within a week:

• Live conversations per agent-hour go up, because humans stop being dropped and voicemails stop reaching agents.
• Your machine-classified bucket gets smaller and cleaner, which is itself diagnostic evidence the old bucket was dirty.
• Agent morale improves marginally, because nobody enjoys saying "hello?" into silence repeatedly, and nobody enjoys being connected to a voicemail box to read a script at it. I count this as a real but unmeasurable bonus, so I do not put it in the decision math.

Run a Fair One-Week A/B on Your Own Traffic

Do not take my word for the numbers. Take yours. Here is the test protocol I recommend, designed so that your own team can defend the result internally.

Setup: Split by Day, Not by Agent

The cleanest structure for a small outbound floor is alternating days: standard detection on Monday, Wednesday, Friday; the new classifier on Tuesday, Thursday, Saturday. Same lists worked in the same rotation, same agents, same shift schedule. If you have multiple campaigns with comparable offers, you can instead split by campaign for the full week, which gives you bigger samples per arm.

Resist the urge to randomize by agent. Agents differ enormously in skill, and a strong agent landing in one arm will swamp the detection effect. Days and campaigns are the fairer coin.

The One Measurement That Settles It

If you track a single number for the whole week, track this: live conversations per agent-hour. Not dials per hour, which pacing changes can inflate. Not raw answered calls, which include voicemails. Conversations per agent-hour is the number that connects detection quality directly to payroll efficiency, because agent-hours are what you are actually buying with wages.

Compute it daily per arm: count of calls where a live human was connected to an agent, divided by total logged agent hours that day. Keep the arithmetic this simple. A test the manager cannot reproduce in a spreadsheet is a test nobody trusts after the meeting.

What to Watch Beyond the Headline Number

Three secondary checks make the result honest:

1. Close rate per conversation. If conversations rise while close rate holds, you have found real demand. If close rate drops sharply, investigate, because more marginal conversations can dilute the average. In our experience the drop is small or absent, since a recovered live answer is exactly as qualified as the answered calls you were already converting, but verify on your traffic, that is the point of the week.
2. The machine bucket ratio. Compare the size of the answering-machine bucket between arms. A meaningfully smaller bucket under the new classifier, with contact rate up, is the fingerprint of recovered humans.
3. Agent complaints. Zero is the target. Agents noticing nothing means the change is clean.

How Much Traffic Makes the Week Meaningful

You need enough live answers in each arm that a 10-15% difference cannot be luck. As a working rule, if each arm produces at least a few hundred live conversations over the week, the comparison is worth acting on. Below that, run two weeks. If your floor cannot produce a few hundred live conversations per arm in a week, your volume is small enough that the free tier will cover the entire test anyway, which conveniently brings us to cost.

A note on noisy days

One week of outbound data contains noise nobody controls: a holiday, a list batch that underperforms, a Monday with staffing gaps. Do not let a single odd day decide anything. Look at the per-day series across the whole week and ask whether the pattern holds on most days, not whether the totals flatter one arm. A difference that only exists on one day of seven is weather. A difference that shows up on five or six days of seven, in the same direction, is a floor.

The Decision Math: When Free Ends and Whether It Matters

The free Sandbox plan is 50,000 detections per month, hard capped, with no credit card required at signup (https://amdy.io/auth/signup). A hard cap means the failure mode is the service simply stopping at the limit, not an invoice arriving. There is no overage billing on Sandbox, ever. That design exists precisely so that an owner can run this exact experiment without a procurement conversation.

If your volume exceeds the free tier and the week of data supports an upgrade, here is the full pricing structure, which is the real schedule and not a teaser:

Plan	Monthly price	Included detections	Overage per detection	Servers
Sandbox	Free	50,000/month, hard cap	None, service stops at cap	Unlimited
Starter	$79	500,000	$0.00025	Unlimited
Growth	$299	5,000,000	$0.00015	Unlimited
Scale	$999	25,000,000	$0.00010	Unlimited

Every plan is month-to-month and cancellable anytime. Two structural facts in that table deserve an owner's attention. First, unlimited servers on every tier means the price scales with detections, not with your infrastructure footprint, so you never pay a per-box tax for growing. Second, the overage rate declines as volume rises, which is how a usage-priced service should behave when your volume is the customer's volume. The numbers behind choosing a tier, including the per-campaign ROI arithmetic, are laid out at https://amdy.io/pricing.

A Worked Decision Rule

Here is the rule I would use after the test week, and it is conservative on purpose:

• If conversations per agent-hour rose by less than 3%, stay where you are or retest with more volume. Either the leak is small on your traffic or the test was underpowered.
• If it rose 3-10%, compute what those recovered conversations are worth using your close rate and average revenue per close, then compare against the tier your volume puts you in. The ROI framing and formula are laid out at https://amdy.io/roi.
• If it rose more than 10%, the decision has already made itself, and every week you wait is the leak running at full flow.

Notice that in all three branches you made the call from your own week of data rather than my claims. That is the entire design of the test. Claims are how vendors get pilots. Agent-hour data is how vendors keep customers.

The No-Worst-Case Ledger

Executives weigh downside before upside, so let me write the worst-case ledger explicitly, because it is unusually short.

• Financial worst case: $0. The test runs inside the 50,000-detection free cap with no card on file. If you exceed it, the service stops. There is no bill to dispute.
• Operational worst case: your dialer admin reverts one command. There is nothing to uninstall beyond that, no residue in your stack, no carrier to call.
• Data worst case: you learn your current detection is better than expected. That is worth knowing and costs nothing to learn.
• Reputational worst case: none in the test direction. If anything, fewer dropped live answers reduces the dead-air experiences regulators measure. The FTC caps abandoned live answers at 3% per campaign over 30 days (16 CFR 310.4(b)(4); the rule text is public at https://www.ecfr.gov/current/title-16/part-310), and every classification is logged and exportable, so your audit position strengthens rather than weakens. We treat the record-keeping side of detection as a first-class feature, not an afterthought, and the reasoning is here: https://amdy.io/blog/amd-logs-compliance-defense.

Compare that ledger against the one you mentally drafted the last time a vendor asked for a pilot. The asymmetry is the pitch.

Three Mistakes That Ruin the Test Week

I have watched this experiment run on enough floors to know where it goes wrong. None of these mistakes are exotic, and all of them are avoidable with one sentence of instruction to whoever owns the test.

Mistake one: changing two things at once. The week you test detection is not the week to also launch the new script, the new list vendor, or the new incentive plan. If anything else on the floor changes mid-test, you no longer know which change produced the movement, and you will have to run the week again. Freeze everything else. This is the discipline that separates a measurement from an anecdote.

Mistake two: telling agents which arm is which. Agents who believe the new system is being evaluated will unconsciously work the "new" days harder, or brace for disappointment on the "old" days. The test only stays blind if the floor stays uniform. Tell them nothing beyond "the dialer team is doing routine checks this week." If your floor is small enough that secrecy is impossible, split by campaign instead of by day, and accept slightly more noise between campaigns.

Mistake three: judging on day one. Detection changes produce their effect instantly, but your conversion counts lag by the length of your sales cycle. If deals take ten days from first conversation to close, the conversations recovered on Tuesday are still open on Friday. Judge conversations per agent-hour at the end of the week, and judge close rate on the full cycle afterward. Owners who declare victory or failure on day one are reading a number that has not finished forming.

What Week Two and Beyond Look Like

Suppose the week came back positive, you picked a tier, and you are rolling out. What does steady state actually feel like on an outbound floor?

For the first few days, managers typically report mild disbelief at the contact rate column, followed by a scramble for list depth, since agents finish their lists faster when fewer live answers are thrown away. That scramble is the operational proof of the leak, arriving from the opposite direction. You are now short on dials for the same headcount, which means either more list spend or a pacing conversation, both of which are normal management problems rather than silent losses.

Within a month, the pattern I hear about settles into three durable changes. Conversations per agent-hour run above the old baseline. The machine bucket stabilizes at its true size, which makes list quality judgments cleaner, since a "bad list" now means actually bad rather than badly read. And close rate per conversation sits where it always did, because the added conversations were real people all along.

The audit trail quietly earns its keep in this period too. Every classification is logged and exportable, so when a client or a compliance reviewer asks how you handle answered calls, the answer is a file rather than a paragraph. On the days a regulator or a carrier asks harder questions about dead air, that file is the difference between a defensible operation and an assertion. Detection quality also is not a set-and-forth property, since carrier voicemail behavior shifts over time and a classifier must be watched, which is a subject we go into here: https://amdy.io/blog/amd-model-drift.

A Word on Who Should Run This

The test needs three roles and no consultants. The dialer admin runs the one command and confirms it took, which is a five-minute task. A manager owns the daily spreadsheet of conversations per agent-hour, roughly ten minutes a day. An owner or VP looks at the end-of-week number and makes the tier decision from the table above.

That last role is the one that should not be delegated, and not because the arithmetic is hard. The arithmetic is division. The reason to keep the decision is that the number it produces, recovered revenue from contacts already paid for, is a strategic fact about the business, in the same category as knowing your true cost per acquisition. Managers optimize what they are handed. Only the person who reads the P&L treats a leak like this as a priority, and in my experience the act of personally signing off on the test week is what makes the result real to the organization rather than another dashboard row.

If you want the longer-form framing of detection quality as a board-level topic, including how it interacts with what you pay per contact over a number's full life, the features overview is at https://amdy.io/features and the cost-per-number piece is at https://amdy.io/blog/cost-per-number-life-metric.

Two Questions I Get From Owners at This Point

"If the old detection was this bad, why did it take a decade to fix?"

Because the error was free to ignore. The cost lands on the prospect's end of the line, in a bucket your reports never break out, and the people being dropped cannot complain to you. Old detection also predated classifiers that could return a verdict in 125ms, so the industry built its habits around a tool that had to guess fast and guess rough. Habits outlive the constraints that created them. We spend a lot of time on why detection quality drifts over time, and how to notice it before your contact rate tells you, here: https://amdy.io/blog/amd-model-drift.

"What does this do to my cost per number?"

It lowers it, which is why that metric is the one I would put on the wall. Cost per number in your CRM measures the full weight of a contact across dialing, marketing, and handling, and it improves when a contact you already paid for converts instead of getting dropped. We wrote a full piece on using cost per number as a life metric for the floor here: https://amdy.io/blog/cost-per-number-life-metric. Between that metric and conversations per agent-hour, you will have the two numbers that make this decision self-evident, in the direction your own traffic points.

The Takeaway

Your next campaign will spend the same budget on lists, wages, and minutes whether your detection is good or bad. The only variable you control before it launches is how many of the people who answer actually reach your agents. One command, five minutes, one week of your own data, and a hard cap of zero dollars on the downside. Recover first, spend later.

Which day this week does your dialer admin have five minutes?