Lead Gen Agencies: Deliver More Live Transfers Without Buying More Leads

The cheapest way to raise transfer volume is to stop losing the answered ones. Not buying more data. Not another lead vendor. Not a bigger dialing budget. The humans who said hello and got classified as voicemail.

I have spent enough time around lead gen operations to know how the P&L is built. You buy leads, you dial them, a fraction of the answers are live humans, a fraction of those warm up enough to transfer, and the client pays per accepted transfer. Every improvement conversation I have ever been in was about the top of that funnel: better lead sources, more volume, tighter targeting. Almost nobody looks at the leak in the middle, because the leak is invisible in the reporting. The dialer does not label a dropped human "dropped human." It labels the call "answering machine," and the report aggregates it with every actual answering machine, and the number looks boring and normal.

Here is the measured reality. Default answering machine detection on Vicidial-class dialers drops an estimated 10 to 20% of live humans. We measured that across detection data (features). Ten to twenty percent of the people who picked up your phone call, heard the start of a pitch or dead air, and got hung up on by software that thought they were a mailbox. For an agency, that is the most expensive silent line item in the business.

The agency economics, stated cold

A lead gen agency sells transfers. Roughly:

Transfers sold per month = leads dialed × contact rate × live answer share × transfer conversion

Buying more leads scales the first term linearly, and the cost scales with it. Everything downstream, contact rate, live answer share, transfer conversion, is supposed to be your operating skill. When AMD drops a live human, the call never gets a chance to convert. The lead is spent, the dialer moves on, and your per-transfer cost rises in a way that never shows up as a line item.

Put it as unit economics:

Cost per accepted transfer = (lead spend + dialing cost + agent cost) ÷ accepted transfers

Any dropped human raises cost per accepted transfer without appearing anywhere in the reports. You cannot manage what your dialer mislabels. This is why I tell agency owners to think in cost per transferred conversation over the life of their numbers rather than cost per lead; the framing is written up here: cost per number, the life metric. Cost per lead flatters lead vendors. Cost per transferred conversation tells you what your operation actually converts.

Where the humans go

When the built-in classifier meets an ambiguous answer, a person who says "hello... hello?", someone answering in a noisy room, a quick "yeah, who's this?", the timer-based defaults in Asterisk's app_amd are what decide. The stock configuration gives the analysis five seconds and trips on silence thresholds; when it cannot decide, it returns NOTSURE and the dialplan decides what that means. Campaigns that route NOTSURE to hang-up or voicemail treatment are quietly composting live prospects. The default drop estimate above, 10 to 20% of live humans, is what that looks like in aggregate.

Why agencies get the worst of it

Two things make it worse for agencies specifically. First, lead gen verticals, solar, insurance, home services, debt, are exactly the verticals with messy answers: people answer mid-task, screens doors, carry phones into yards. Second, agency dialing is usually tuned hard for agent efficiency, which pushes every ambiguous case toward "machine," because machines routed to agents hurt the agent-availability stats everyone watches. The incentive gradient points at your revenue.

Corrected yield math, with your inputs

Do not take a yield number from me or from any vendor. Take the formula. Run it once with your current drop rate and once with the drop rate your trial measured. The gap is the whole story.

First, estimate your current silent loss:

Answered humans per month = dials × contact rate × live answer share

Currently dropped humans = answered humans × drop rate (use 0.10 to 0.20 if you have never measured; measure during the trial)

Then the recovery, assuming the humans you stop dropping behave like the humans you currently reach:

Additional conversations = currently dropped humans × recovery share (measured in your trial, not assumed)

Additional transfers = additional conversations × your transfer conversion

Additional margin = additional transfers × price per accepted transfer − added cost (agent minutes for the extra conversations plus AMD cost)

The honest way to fill in "recovery share" is the trial itself, not my placeholder. Audit a random daily sample of calls, listen blind, count the humans the classifier marked machine. We publish the exact procedure: how to test AMD accuracy. One week of auditing a few hundred calls a day gives you a drop rate with your name on it.

A worked frame, no invented numbers

I will not quote you a case study with a happy ending and a made-up percentage; I have never seen your books. Instead, the two-line version any ops manager can run at the next Monday meeting:

Quantity	How to get it	Where it comes from
Answered humans / month	Dials × contact rate × live answer share	Your dialer reports
Current drop rate	Audit 100 to 200 random answered calls/day for a week, listen blind, humans marked machine ÷ humans answered	Your audit; or start with the 10 to 20% measured default range
Recovery share	Re-run the audit after the fix; drop rate delta is the recovery	Your trial
Transfer conversion	Transfers accepted ÷ live conversations (from your CRM)	Your CRM
Price per accepted transfer	Your client contract	Your contract
AMD cost	Plan pricing, e.g. $79/month including 500K detections, then $0.00025 each	Vendor pricing page

Every row is yours. That is the point. An agency that can fill this table knows more about its own leak than any vendor pitch will ever tell it.

Selling answer quality to your clients

Here is the part most agencies miss. Fixing the leak is not just margin recovery. It is a sellable line in the offer.

Clients buy transfers and increasingly audit them. They ask about contact rates, about why a campaign delivered under forecast, about compliance. An agency that can say "we measure our misclassification rate weekly and here is the per-call log for every classification decision" is making a different sale than "we dial hard." The evidence base matters too: per-call detection logs are the artifact you hand over when a client disputes delivery or when a regulator asks questions; the compliance angle is laid out in AMD logs as compliance defense.

Three ways I have seen answer quality folded into an agency offer without renaming the whole company:

1. Guaranteed conversation depth

Bid on conversations that actually start, not raw connects. You can only safely guarantee what you can measure, and dropped humans are unmeasured by definition until you fix classification. Once your live-human drop rate is known and low, a "live conversation" guarantee stops being a bet and starts being arithmetic.

2. Quarterly answer-quality reporting

Put the misclassification audit in the client report: calls audited, humans found in the machine bucket, corrective action. No client expects this because almost no agency does it. It costs you a QA person a few hours a month once the tooling logs every call. It reads as operational maturity, because it is.

3. Pricing on transferred conversations

The natural end state: price per accepted transfer goes up in exchange for a floor on answer quality, and you win the comparison against the $1-cheaper competitor because their effective cost per real conversation is higher than their headline. The ROI math is what you put in front of the client to make that case, run on their campaign numbers.

One more external reality worth having in your back pocket: the FTC's Telemarketing Sales Rule caps abandonment at 3% of live answers per campaign over 30 days (16 CFR 310.4(b)(4), current text at ecfr.gov). Dropped humans are not just lost revenue; enough of them, poorly accounted, and they start looking like an abandonment problem to the people who count these things for a living. The FTC's own business guidance for the telemarketing sector is at ftc.gov. Agencies get asked about this in vendor questionnaires more every year. Answer quality is a compliance posture as much as a yield lever; the interaction is covered in detail in AMD error directions.

What the fix actually involves

If you are braced for an integration project, stand down. The fix in a Vicidial shop is one command run by whoever administers your dialer, about five minutes, no carrier change, no dialplan rewrite; the classifier runs server-side and takes over the AMD decision. Your leads, your agents, your carriers, your CRM: unchanged. What changes is the verdict quality and the fact that every verdict is logged per call.

Then pricing, stated plainly because agencies hate surprises as much as clients do. Sandbox is free, 50,000 detections a month, no card, hard cap, which is enough to run a real evaluation on one campaign for a month. Starter is $79 a month including 500,000 detections, then $0.00025 per detection. Growth is $299 including 5 million, then $0.00015. Scale is $999 including 25 million, then $0.00010. Unlimited servers on every plan, billed monthly, cancel whenever (pricing). Work out your monthly dials, pick the tier, and the AMD cost per transfer in the table above is a division you can do before you talk to anyone.

For the first month, the trial math is the whole pitch: start the free trial, keep dialing exactly as you dial today, audit a few hundred calls a week, and compare the machine bucket before and after. If the humans do not reappear in your numbers, you have spent zero dollars and one QA workload to learn something true about your operation. If they do reappear, every one of them is a transfer candidate you already paid the lead cost for.

Running the trial inside an agency week

Agencies do not have slack weeks, so run this inside normal operations. The shape that works:

Monday. Dialer admin runs the install on one server, about five minutes, and you pick one campaign with real volume. Keep routing as-is for the first 48 hours; the classifier logs verdicts in shadow while nothing changes for agents. You collect a same-week before picture without asking anyone's permission to change anything.

Wednesday. Flip routing on that campaign to the new verdicts. Tell the floor manager nothing beyond "AMD changed," because you want the agents' unprompted read. In our experience the first signal that humans are being recovered is not a dashboard, it is an agent saying the calls feel different today.

Every day. QA pulls the random audit sample, 100 to 200 answered calls, listens blind, counts humans in the machine bucket. Twenty minutes of listening per hundred calls once they get fluent. By the following Tuesday you have a week of both buckets and the delta is your recovery share, measured, not promised.

The Monday after. Fill in the table from earlier and look at one number: additional transfers at your current price. If it is smaller than the plan cost, you have spent $0 and learned your dialer was fine. If it is larger, the decision is arithmetic and the only remaining question is how fast you roll it across the other campaigns and servers.

Two failure modes to watch. First, do not let the audit sample come from the campaign's best list source; pull it across everything the campaign dialed that day. Second, do not credit the fix with transfers that came from a mid-trial lead batch change. Freeze media mix for the week if you can, and if you cannot, note the change next to the numbers so future-you knows what moved.

Selling it internally first

If you are an ops lead reading this rather than the owner, your problem is not the math, it is that the math is invisible. The dropped humans appear nowhere as a cost line, so fixing them looks like spending money to solve nothing. Two framings work in that meeting.

The first is the lead-vendor comparison. Ask the room what the company would do if a lead vendor silently delivered 15% fewer billable leads than invoiced. There would be emails, demand letters, a switch. Then show that the dialer does the equivalent to leads already paid for, every month, and nobody's inbox lights up because the report says "answering machine."

The second is the agent-efficiency reframe. Managers fear that keeping more ambiguous calls means dumping machines on agents and watching handle time climb. The error direction matters here: you are not trading agent time for humans at one-to-one. The machine pass-through rate and the live-human drop rate move on separate dials, and the tuning that fits a transfer shop leans toward keeping humans at the cost of a few extra quick hang-ups. Which errors to buy and which to avoid is laid out in AMD error directions. A 20-second agent hang-up has never cost anyone a client. A human who heard silence and got dropped has.

Capacity planning after the leak is fixed

One operational consequence people forget. If you recover a chunk of answered humans, your agents talk more. That is the point, but it changes the math the capacity planner uses. Before the fix, a share of your dialing capacity was quietly spent on calls that ended in software hang-ups. After the fix that capacity converts to conversations, which means either more transfers per agent-hour or the same transfers on less dialing. Both are wins, but only if someone re-derives the pacing. An agency that fixes AMD and then re-plans pacing captures the gain twice: once on yield, once on dialer spend per transfer.

Roughly:

Agent-hours needed after = agent-hours before × (1 + recovered conversation share × avg handle time effect)

Do not over-engineer it. Run a week at current pacing, look at abandonment signals and agent idle, and adjust once. The point is that the fix is not a bolt-on; it shifts the operating curve of the room, and the agencies that treat it that way pull the margin off the table instead of leaving it in the dialer.

The line I would leave you with

Every agency I have worked with was already paying for the transfers it drops. The lead spend, the dialer time, the agent idle, all of it spent, and then a threshold timer throws the person away at the last second.

You would never accept a 10 to 20% shrink rate on delivered leads from a lead vendor. Why accept it from your own dialer, invisibly, on leads you already bought?