Why Buying More Leads Won't Fix a Leaky Dialer
Buying more data into a dialer that drops live answers is filling a cracked glass. The funnel math that shows why lead spend compounds the leak instead of covering it.
Why Buying More Leads Won't Fix a Leaky Dialer
When connect rates sag, the reflex in most outbound shops is identical: buy more data. The list must be tired. The broker must have sold recycled numbers. The lead vendor gets blamed, the budget for July data doubles, and for one week the connect rate ticks up. Then it sags again, because the new data poured into the same hole the old data was pouring into.
I run answering machine detection for ViciDial, Asterisk and FreeSWITCH shops, and I have watched this cycle from the inside more times than I can count. The expensive belief underneath it is that lead volume is the input that fixes output. It is not. If your dialer misclassifies a slice of live humans as machines, every additional lead you buy adds customers to the pile of humans you hang up on. More water, same cracked glass.
Here is the funnel math that makes the point impossible to argue with, and how to tell whether your problem is data or detection before the next data invoice goes out.
The funnel is a chain of multiplications
Outbound conversion is not a pipeline you can patch at the end. It is a chain of fractions multiplied together.
The four fractions
Take a simplified month: 100,000 dialing attempts. A fraction connects at all, say 45%. A fraction of connects reaches a live human rather than a machine, say 27% of attempts. A fraction of those humans is correctly delivered to an agent. A fraction of delivered conversations converts.
Each fraction multiplies the ones before it. Improve the last fraction, conversion, by 20% and revenue rises 20%. Improve the delivery fraction by 20% and revenue rises 20% too, but lead spend stays the same. Now watch what happens when the delivery fraction is broken and you attack the top of the funnel instead.
The leaky-glass arithmetic
Suppose stock dialer AMD is dropping 10% of live humans as machines, which is the conservative end of what we measure on real traffic. Your funnel:
| Scenario | Attempts | Live humans reached | Delivered to agent (AMD quality) | Conversations at 25% close of delivered |
|---|---|---|---|---|
| Baseline, stock AMD | 100,000 | 27,000 | 24,300 (90%) | 6,075 |
| Double the data, stock AMD | 200,000 | 54,000 | 48,600 (90%) | 12,150 |
| Same data, AMD at 98% | 100,000 | 27,000 | 26,460 (98%) | 6,615 |
Read the second and third rows carefully. Doubling data spend doubles everything, including the absolute number of dropped humans: 5,400 people hung up on instead of 2,700. You paid twice for the privilege. The third row spends nothing on data, recovers 2,160 humans a month that were already paid for, and gets you 89% of the way to the doubled-data outcome at a fraction of the cost.
The data doubling also doubles telephony, doubles compliance surface, and burns your best numbers faster. The detection fix doubles none of those. Nothing in the doubled-data row is free except the dropped humans.
How to tell which problem you actually have
Before the next lead order, run these three checks. They take an afternoon and they settle the question.
Check the answer mix by list source
Split last month's answered calls by data source and look at the human answer rate per source. Tired data shows its real signature here: answer rates fall, but the ratio of humans to machines among the answers stays roughly stable. A detection problem shows the opposite signature: answer rates stable across sources, but the human fraction of answers varies wildly for no list-related reason. The failure follows the audio, not the list.
Pull a sample and listen
This is the unglamorous ground truth. Take 50 calls your dialer marked as answering machines, at random, across campaigns. Listen to the first six seconds of each. If you hear a human greeting in five or more of them, you have found your leak with your own ears, and no broker conversation is going to fix it. The Asterisk project documents its stock detection behavior at https://docs.asterisk.org/ and the settings have not fundamentally changed in years. Timing heuristics from that era were tuned for a phone network that no longer exists.
Compute the ratio that matters
Of all calls answered by a live human last month, how many reached an agent? We describe this number in detail in the one KPI your dialer report is missing. If nobody in the building can produce it, that absence is itself the diagnosis: the dialer's classification layer is trusted blindly, and it is the one layer never audited.
Why the reflex points at data anyway
Three reasons, all human.
First, data spend is a lever a manager can pull today. Detection quality is a system property nobody owns. Organizations reach for the lever they have authority over.
Second, lead vendors get measured, detection does not. Every broker has a replacement policy and a reputation to protect, so complaints have a destination. The AMD layer has no vendor to blame, because it shipped free inside the dialer. Free things never get audited.
Third, the failure is silent in every report. A dropped human looks identical to a machine in the logs. The waste hides inside a category that is supposed to be waste. We walked through exactly where this money goes in the revenue leak no dashboard shows, and the same blindness is what keeps the data-buying reflex alive.
The order of operations that actually works
Fix detection first, then buy data. The sequence matters because of how the fractions multiply.
Step one: measure the classification layer
Run AI answer detection alongside stock AMD for a week on live traffic. Count divergent verdicts. Every divergence where the AI says human and stock says machine is a recovered conversation waiting to happen. The method for a clean before and after comparison is in five minutes to install, a week to prove itself. No production risk, no migration, five minutes of dialer admin time.
Step two: reprice your data with the fixed funnel
Once delivery is honest, your per-lead economics change. The same $10,000 of monthly data now yields its real output, and you can judge sources on delivered conversations instead of raw connects. This is when buying decisions get sharp: some sources look better under honest delivery, some worse. The metric to steer by is cost per delivered conversation, not cost per lead, a distinction we drew hard in why CPL is the wrong KPI.
Step three: then scale spend
Only now does doubling data double output cleanly, because the glass no longer has a crack. Every extra dollar buys a human with a fair chance of reaching an agent.
A worked example: the insurance shop spending its way sideways
A Medicare supplement shop, 30 seats, spending $26,000 a month on aged leads from three brokers. The owner's instinct when conversions sagged was identical to the reflex we opened with: the data is tired, buy fresher data, buy exclusive data, buy data from the new vendor who called last week. He ran that experiment for five months before anyone looked at the dialer.
What five months of extra data bought
Each new source lifted the connect rate for roughly two weeks, then decayed back to trend. The owner read that as evidence of recycling and kept rotating brokers. The pattern was real but the diagnosis was wrong. What actually happened: fresher lists answered more, and more answers meant more humans reaching a classifier that was mishandling 15% of them. The extra data did produce extra humans. The dialer then dropped roughly one in seven of the extra humans it had just paid to reach, so the visible gain was the six that got through, and the invisible loss was the one who heard silence and formed an opinion about the brand.
You can see the trap in the numbers. The gain shows up in reports. The loss shows up nowhere. Five months of data rotation bought the shop a treadmill: constantly paying for a seventh human to lose the seventh human.
What the audit found
Eight days of parallel detection on live traffic, plus the listen-to-50-drops exercise described above. Eleven of 50 sampled machine-classified calls opened with a human greeting. At their volume that computed to 380 dropped humans a day. Their true per-campaign abandonment math, recomputed with those events included, had been riding above 3% on their biggest campaign most afternoons, a fact their dashboard had never surfaced because misreads were logged as machines on both sides of the fraction.
The reversal
Detection went in first, data spend stayed flat for 60 days by agreement. Delivered conversations per dollar of data rose 24% in the first month, without a single new lead purchased. The owner's own summary on the follow-up call: the shop had been paying three brokers to fill a bucket with a hole in it, and arguing with the brokers about the water.
Only after the funnel was honest did data decisions get sharp. Under clean delivery, one broker's numbers proved worth 30% more per dollar than the other two, invisible until then because that broker's lists skewed older with more voicemails, which the broken classifier had been punishing disproportionately. The detection fix revealed the best vendor, rather than replacing the need for one. That is the correct order of operations doing its job.
Three objections, answered
"Our data vendor has a replacement policy"
Good, use it, for the numbers that never answer. A replacement policy does nothing for the human who answered and got dropped, because that number looks worked in your system. You will never file the claim. The policy covers dead data. The leak lives in live data.
"Our IT person says AMD is already configured"
Configured is not measured. Every dialer ships with AMD settings that somebody once tuned, often to the vendor's defaults, often years ago against a phone network that has since changed underneath them. Voicemail platforms shortened their greetings, carriers added announcements, call screening arrived. The question is not whether settings exist, it is what fraction of live humans those settings drop today. That is a measurement, it takes an afternoon, and we described exactly how to run it in five minutes to install, a week to prove itself.
"We'll fix detection when we scale"
The leak is a percentage, so it scales with you automatically. The shop dialing 100,000 attempts a month loses humans at the same rate as the shop dialing two million. The small shop loses fewer dollars in absolute terms, but it also has the least slack to waste, and every month of unfixed percentage is a month of compounding brand damage on numbers you will want to dial again. There is no scale at which waiting becomes the economical choice, which is why the free tier exists at 50,000 detections a month: the test costs nothing but an afternoon.
Why the list is never entirely innocent
A fair article has to say the other half out loud: data does go bad, vendors do recycle, and some lists genuinely are tired. The leaky-dialer argument is not that data quality is imaginary. It is that data quality and detection quality are two separate failure modes that produce identical symptoms from the manager's chair, and shops reliably treat 100% of those symptoms as data problems because data has vendors to blame and detection does not.
The symptoms overlap almost perfectly
Falling connect rate: tired data answers less, and a decaying detector also connects fewer humans. Campaign decay week over week: lists burn in, and drift worsens as answer-side audio shifts. Vendor performance varying: sources genuinely differ, and the classifier's error rate also varies by list demographics, since older populations produce more voicemail and more screening, the exact audio timing-based detection handles worst. Every symptom you would use to indict the data has a twin explanation in the detection layer. Symptom-reading cannot separate them. Sampling can, in an afternoon, which is why the audit precedes the invoice.
The interaction effect nobody budgets
Here is the part that makes the compounding worse than either failure alone. A decaying classifier punishes your best data hardest. Clean, fresh, human-heavy lists put the most live answers in front of the detector, so a 15% human-drop rate eats its percentage out of your most expensive numbers. The aged remainder, mostly machines, passes through untouched. The result is a silent rebalancing of your effective data spend toward your worst inventory, invisible in every report, felt only as that maddening sense that the good vendor's numbers are not performing like they used to. Sometimes the good vendor became the punished vendor, and the punishment came from your own dialer.
A note on the brokers
None of this excuses sloppy data practices. Dedupe against your own DNC and litigator lists, age your inventory, track per-source performance honestly. Do it on top of a clean funnel, not instead of one. The shops that run both disciplines well pay less for data every quarter, because they can measure what each number is actually worth once it reaches the classifier. The measurement is the whole advantage, and the measurement requires the classification layer to be trusted first. That is the whole argument of this article in one sentence: you cannot price your data correctly while your dialer is dropping the humans the data paid for.
The one-week plan
If you want this settled before the next data invoice, here is the week, day by day.
Monday: pull last month's answer data split by list source and compute the human-answer ratio per source, the first diagnostic from earlier in this article. Tuesday: sample 50 machine-classified calls and listen to the opening seconds of each. Write down the count of human greetings. That single number tells you whether the rest of the week is worth running, and in five months of doing this with shops I have never seen the count come back zero. Wednesday: install detection in parallel, Sandbox tier, five minutes of dialer admin time. Thursday and Friday: let it run alongside stock AMD and collect divergent verdicts. Weekend: ignore it, it is logging, not judging. Monday after: price the divergences with your own cost per attempt and your own conversion value, and the data-versus-detection question is answered with your numbers instead of my arguments.
Notice what the plan does not contain: no migration, no carrier change, no downtime, no committee. The whole week rides on one afternoon of listening and one five-minute install. Whatever you conclude, you will conclude it from evidence you generated yourself, on your own traffic, which is the only evidence that survives contact with a budget meeting.
What the fix costs, in plain numbers
The free Sandbox tier is 50,000 detections a month, no card, no day limit, enough to run the comparison test on a small campaign. Paid plans start at $79 a month with 500,000 detections included and $0.00025 per detection past that. For the 100,000-attempt shop in the table above, that entry tier covers the month with room to spare.
Compare that to the doubled data invoice. One buys new humans to hang up on. The other stops hanging up on the humans you already paid for. Fill the crack before you pour more water.