Free for 50,000 Calls: The Trial CFOs Like

Proof on your own traffic before a single dollar moves.

That sentence is the whole pitch, and if you are the person who signs off on software line items, it is the only sentence that should matter. Most detection vendors want a contract before they show you anything. We do the opposite: 50,000 detections a month, free, no credit card, running on your real campaigns while your current setup keeps working. You get a report from your own traffic. Then you decide.

I have sat on both sides of this table. I have been the operator begging for budget and the executive asking why the number in column three moved. Every pitch that starts with "trust us" dies in procurement. Every pitch that starts with "here is last Tuesday's data from your own dialer" gets a fair hearing. This article is about why a capped free tier is the cleanest procurement instrument for answering machine detection specifically, and what you as a CFO or owner should demand from the trial before you approve a dollar.

The problem the trial has to prove out

If you run outbound with a Vicidial-based dialer, your current answering machine detection is the Asterisk default: fixed timers listening for silence and word counts. It works well enough that nobody questions it, and it fails quietly enough that nobody measures it. Our measured data puts the failure at an estimated 10 to 20 percent of live humans misclassified as voicemail and dropped. Read that again. Somewhere between one in ten and one in five people who actually said "hello" never reach your agents.

Nobody sees this happen. The person says hello, the dialer hears a machine, the call ends, and your reporting shows nothing unusual. It looks like a slightly worse contact rate. It is actually a hole in the bucket.

That is the claim the trial has to test. Not our claim about the industry, our claim about your traffic. Your lead sources, your carriers, your call windows, your answer patterns. A demo cannot do that. A sandbox with recorded audio cannot do that. Only your production calls can.

I wrote about which errors matter and in which direction in a separate piece on error direction in AMD. The short version: a false machine (hanging up on a human) costs you revenue directly. A false human costs you agent seconds. They are not symmetric, and your trial report should count them separately.

Why "free trial" usually fails the CFO test

The card-on-file trap

Most "free trials" in this space are free the way a gym membership is free. Card on file, auto-convert, and the burden is on you to cancel before day 30. The vendor is betting on inertia. You know this, which is why finance teams treat trial signups as procurement debt: a small future liability, undocumented, that turns into a surprise line item when nobody cancels.

Our Sandbox plan takes no card at all. There is nothing to cancel and nothing that can auto-convert. If you do nothing at the end of the trial, the trial simply stops at 50,000 detections in a month. That is a different risk profile, and it changes who can start the trial. With no card and no contract, the trial can be started by a dialer admin or an ops manager without a purchasing process. The CFO gets involved only when there is data to review.

The uncapped-usage trap

The second failure mode is worse: a free tier with no cap, billed on overage. You install it, forget about it, and a busy month of dialing produces a five-figure invoice. This is the single most common way "free" becomes expensive.

The Sandbox plan is a hard 50,000 detections per month. Not a soft limit, not a courtesy threshold. At 50,000 the service stops until the next month. There is no overage charge, ever, on the free tier. The worst possible financial outcome of the trial is zero dollars. I want that in writing-equivalent form here because it is the first question every CFO asks and the answer should be unambiguous: can this trial generate a bill? No.

The demo-data trap

The third failure mode is subtler: trials that run on vendor-supplied audio. A curated set of clean recordings proves nothing about your traffic. Every call center's answer patterns differ. Debt collection calls get guarded, clipped "hello?" pickups. Solar leads get long household greetings. Insurance renewals get a lot of seniors with slow greetings and background noise. A classifier tuned or cherry-picked on one distribution can fall apart on another, and the fall is exactly the drop-rate problem you are trying to measure. We have watched model performance drift when traffic mix shifts, and I wrote about that mechanism in AMD model drift. The trial must run on your distribution, or it is theater.

What the trial looks like in practice

The install is one command, run by whoever administers your dialer. It takes about five minutes. AMDY runs alongside your current setup during the trial; you are not ripping anything out to evaluate a replacement, which is the objection that kills most evaluations before they start.

Here is the shape of it:

1. Sign up for the free Sandbox at the signup page. No card.
2. Your dialer admin runs the install command on the existing Vicidial server. About five minutes.
3. AMDY classifies calls and the verdict stream starts at 125 milliseconds. Your current detection keeps doing its job in parallel, so nothing degrades while you measure.
4. Every classification is logged and exportable, per call.
5. At the end of the month, or at 50,000 detections, you pull the log and compare.

That last step is the whole point. The per-call logging means the trial produces evidence, not vibes. If you want to sample 200 calls where the two systems disagreed and have a human listen, the log gives you the call records to do it. I would insist on exactly that, vendor regardless. Any detection vendor that cannot export per-call verdicts for independent review is asking you to grade their homework with their own pen.

The plans, and what the trial is deciding

The trial is not deciding "is AMDY nice." It is deciding which line item, if any, to add. Here is the full plan table so you can see the decision before you start the trial rather than after:

Plan	Monthly price	Detections included	Overage per detection	Card required	Cancel
Sandbox	$0	50,000 (hard cap)	None. No overage, ever	No	Nothing to cancel
Starter	$79	500,000	$0.00025	Yes	Monthly, anytime
Growth	$299	5,000,000	$0.00015	Yes	Monthly, anytime
Scale	$999	25,000,000	$0.00010	Yes	Monthly, anytime

Every plan, free or paid, includes unlimited servers, the real-time dashboard, and exportable per-call detection logs. The upgrade path is not a feature gate; it is purely volume. That matters for the CFO conversation because the trial-to-paid decision reduces to one variable: your monthly detection volume. 50,000 detections a month of trial data tells you your real run rate, which tells you your exact plan, which tells you your exact cost. No negotiation, no annual commitment, no "contact sales" tier until you are at partner scale.

Monthly cancel-anytime on the paid plans is the other half of the procurement story. If the accuracy regresses or your dialing volume drops, you are not trapped in an annual contract defending a sunk cost. The vendor carries the retention risk, not you. That is how it should be when the product's value is measurable monthly.

How a CFO validates the spend with trial data

Step one: establish the baseline

Before comparing anything, get your current numbers: calls per month, human answer rate, and contact rate. If your dialer already tracks these, pull the last three months so you can see normal variance. The trial comparison only means something against your own baseline, not against an industry average we quote at you.

Step two: count the disagreements

From the trial log, count the calls where your current detection said machine and AMDY said human, then sample them and listen. This is the number that becomes money later. If sampled humans confirm at the rate the log claims, you have measured your own drop rate for the first time. If it is 2 percent, the product may not be worth it for you, and the trial just saved you from a bad purchase. If it is 15 percent, the paid plan pays for itself almost immediately. Both outcomes are wins for the process.

Step three: convert to a monthly figure

The arithmetic is simple and I will not fake it with invented numbers. Use your own:

• Recovered conversations per month = calls per month × your measured human-drop rate
• Value per month = recovered conversations × your close rate × your average value per closed deal
• Payback = monthly plan cost vs that value per month

We built an ROI page that walks this with your inputs, and I wrote the longer version of the method in cost per number, the life metric. The trial exists precisely so the middle variable, your measured drop rate, is a measurement instead of a guess.

Step four: kill it or scale it

If the math clears, the paid plan is a monthly decision with a one-digit or two-digit cost relative to the value at stake. If it does not clear, you spent zero dollars and five minutes of admin time to avoid a recurring expense. Either way you now know your drop rate, which you did not know before, and which will make you a harder negotiator with every vendor in this category from now on.

What the trial report should contain

If a vendor hands you a trial summary, it should have at least these items, and ours does:

• Total detections, by week, so you can see consistency rather than one good week.
• Verdict distribution: human, machine, and uncertain, counted separately. Uncertainty hidden inside the other two buckets is how accuracy numbers get inflated.
• The disagreement set: calls where your current system and AMDY differed, exportable, so your team can audit by listening.
• Per-carrier breakdown, including false answer supervision. Some carriers signal answers that never happened, which corrupts any detection comparison that does not account for it.
• Latency figures, because detection that takes a second to decide is compliance risk, not just UX. The FTC caps predictive dialer abandonment at 3 percent of live answers per campaign over 30 days under the Telemarketing Sales Rule (16 CFR 310.4(b)(4), readable at ecfr.gov). A verdict starting at 125ms keeps you far from the zone where slow classification itself generates abandoned calls. I cover the interaction in more detail in AMD latency vs accuracy.

If a trial report is a single accuracy percentage with no export, it is a brochure.

Objections I have heard from finance teams

"Free means the vendor is desperate"

Or it means the vendor does not need a sales process to close a measurable product. Both readings exist in the market, so judge by structure. A desperate vendor gives free trials with no cap and a card on file, because the trial is a billing trap. A capped, cardless trial that ends in an exportable log is a vendor who expects to win on the data and is willing to lose on it too. The same trial that can produce your business case can produce a reason to walk, and a vendor who would rather you walk than buy blind is one you can renew next year without a negotiation.

"We do not have anyone to run an evaluation"

The install is one command by whoever already administers your dialer, about five minutes, and it runs alongside the existing setup. There is no parallel environment to build, no traffic to mirror, no change freeze to schedule. The evaluation effort after install is listening to a sample of calls where the two systems disagreed, which is work your QA team can absorb in an afternoon. If a detection evaluation requires a project plan, the vendor has made it your problem on purpose.

"Procurement will take longer than the trial"

That is fine, because nothing expires. The Sandbox tier does not shut off on day 14 to manufacture urgency. Run the trial, take the report through your internal process, and the data keeps. A purchasing timeline should never be the reason a vendor gets to reset your evidence.

The part nobody puts in the brochure

A free trial with a hard cap costs us something. 50,000 detections a month is real compute, given away with no card and no commitment, and some percentage of accounts will use it forever without paying. We accept that because the alternative is asking executives to buy detection on trust, and trust is a terrible procurement instrument. The cap also keeps the free tier honest on our side: we cannot afford to run a free tier poorly, so the free tier gets the same classifier and the same logging as the paid plans. The trial you run is the product you would buy.

That is also why the trial ends in a report rather than a countdown. Nothing expires on day 14 to pressure you. The natural end is the data itself: once you have 50,000 detections of your own traffic, you have either a business case or a reason to walk. The decision makes itself, which is the only kind of decision that survives contact with finance.

Where this fits in the broader risk picture

Answering machine detection sits at an awkward intersection of revenue and compliance. Drop humans and you lose money. Classify too slowly or too aggressively and you drift toward the abandonment cap. Keep no logs and you cannot defend either side of it. The trial's per-call logging does double duty here: it measures the revenue question and documents the compliance question from day one. We keep an exportable audit trail on every plan for that reason, and I went deeper on it in AMD logs as compliance defense. If you are in insurance outbound, the sector-specific framing is in AMD for insurance outbound.

The one-line version: the trial is free, capped, and on your own traffic, so the only way to lose money on it is to not run it.

Start it

Fifty thousand detections on a dialer doing a few hundred thousand attempts a month is a full month of evidence. On a smaller dialer it can be two. Either way you come out the other side knowing your own drop rate, your carrier quality, and your exact plan pricing, and you will have spent nothing but one admin command.

Start the free trial at amdy.io/auth/signup. Full plan details are on the pricing page, and the feature set is on the features page. When your trial report lands on your desk, the only question left will be whether the recovered conversations were worth $79 a month. Your own numbers answer that better than I can.

What would your board say if you presented a tool purchase where the vendor had already run the pilot, for free, on production traffic, before the first invoice existed?