The Last Upgrade Article You Will Need to Read

This is the article I wish someone had handed me fifteen years ago, and it is deliberately the last one in the series. If you have read the other pieces I have written about answering machine detection, the money leak, the latency tradeoffs, the compliance angles, you already have the pieces. This is where the pieces become one picture, and where you get to stop thinking about this category, possibly forever.

The claim in the title is not marketing. It is a design statement. Answer handling, the split-second decision about whether your outbound call reached a person or a machine, has reached the stage that electricity reached a century ago. It can be utility-grade. Installed once by one person, invisible to everyone else, measured in one weekly number, billed monthly like any other line item. When a technology reaches that stage, the reading list stops and the meter reading starts.

What follows is the owner's view, not the engineer's. No jargon, no architecture. Just what you buy, what it costs, what you watch, and how you know it is working. If you never read another word about this category after this piece, you will still be better equipped than most operators in the market.

What Utility-Grade Means For An Owner

Utilities share four properties, and answer handling now shares all four. Let me be precise about each one, because the word gets used loosely and I want it nailed down.

Installed Once, By One Person

Electricity needed a wire run once. Water needed a pipe. Detection needs one command, run once, by whoever administers your dialer. It takes about five minutes. I have written the install story separately, but the short version is that there is no project, no integration sprint, no consultant. If your vendor proposes a services engagement to get detection running, you are talking to the wrong kind of vendor.

The reason this matters to you as an owner is not the five minutes. It is what the five minutes implies about the product's entire design philosophy. Things that install instantly are things that were built to be operated, not sold. Things that require an implementation team were built to be sold, not operated. You can tell almost everything about a vendor's future behavior from what their install feels like.

Invisible To Agents

Your agents should never know detection exists. No new screens, no buttons, no workflow changes, no training session. A call arrives that their dialer has already sorted, they take it, they talk. The sorting happened upstream in the time it takes a phone to ring twice.

This invisibility is a feature you pay for, not a missing feature. Every hour of agent attention is revenue-producing or it is not. Technology that consumes agent attention to justify its existence is billing you twice: once in fees, once in focus. The detection layer touches your agents zero times per day, and that number belongs on your vendor scorecard right next to accuracy.

Measured In One Weekly Number

A utility gives you a meter. Your electric bill is one number times one rate. Detection gives you the same shape: the percentage of live humans correctly connected, which you read off a weekly report, plus the detection count, which drives the bill. That is the whole dashboard.

I am not saying richer data does not exist. Our logs record every call with a verdict, a latency, and a confidence, and I have written at length about why per-call logs matter for compliance and about honeypot detection for catching networks that fake answers. But that depth is for the audit and the forensic afternoon, not for running the business. Running the business is one number, read weekly, compared to last week. If it moves, you ask why. If it does not, you go run the rest of your company.

The number worth knowing from day one: the default AMD that ships with most dialers drops an estimated 10 to 20% of live humans, a range we measured, not a guess. Every one of those drops is a person who said hello and was hung up on by a machine, plus the acquisition cost you spent to make that phone ring. I did the payroll math on wasted agent minutes in an earlier piece about the voicemail tax, and the short version is that misrouted voicemails are a salary leak hiding inside a scheduling line item.

There is one property of the meter worth stressing to anyone who has run a center through a bad quarter. A meter judges nothing. Your detection count falls when volume falls, the bill falls with it, and nobody at the vendor calls to ask why. Under older pricing models, a down quarter triggered the awkward conversation about minimum commitments, or license true-ups, or whether you were in breach of some seat count you signed for in better times. Utility pricing has no opinion about your quarter. The meter reads what you used, and you pay for what the meter reads. The first time you experience that during a contraction, you understand why the comparison to electricity is not decoration.

Built For The Person Who Signs, Not The Person Who Configures

One more property that separates utility-grade from everything else, and it is the one owners feel most. The person who signs the check and the person who runs the command are different people, and a utility-grade product lets each of them interact with it at their own altitude. Your dialer admin sees a config file and a log stream. You see a monthly line item and a weekly percentage. Neither of you has to attend the other's meetings.

Most procurement pain in this industry comes from products that force altitude mixing. The CEO gets pulled into a technical decision because the pricing is per-something-technical. The admin gets pulled into a commercial negotiation because the contract references infrastructure counts. Per-detection pricing with included volume dissolves the whole problem. The commercial unit is the call. The technical unit is the call. Everyone is looking at the same object, so nobody has to translate.

Priced Monthly, Like A Line Item

Utilities bill monthly, in arrears, based on the meter. Detection should do the same. No annual contract you have to lawyer, no per-seat true-up, no professional services retainer. Monthly, cancel anytime, and the price tracks the meter.

Here is the full meter schedule, from our own pricing page, because this article is the one place the whole ladder should be visible in one table.

Plan	Monthly base	Included detections	Overage per detection	Servers	Commitment
Sandbox	$0	50,000/mo, hard cap	None, cap enforced	Unlimited	None, no card
Starter	$79	500,000/mo	$0.00025	Unlimited	Monthly, cancel anytime
Growth	$299	5,000,000/mo	$0.00015	Unlimited	Monthly, cancel anytime
Scale	$999	25,000,000/mo	$0.00010	Unlimited	Monthly, cancel anytime
Partner	From $2,000/mo	Custom	Custom	Unlimited	Custom terms

Read the servers column again. Unlimited on every plan, including the free one. Adding dialer servers never adds fees, a point I argued at length elsewhere, but the one-line version is that servers are inputs and you should never pay for inputs twice. For how to read the detection count as a unit cost over a number's working life, see cost per number, the life metric.

The Total Cost Of Ownership Picture

CFOs taught me this framing and it is the right one. The price of the thing is not the cost of the thing.

The Free End: Sandbox

The true entry cost is zero dollars, a five-minute install, and no card. 50,000 detections a month, hard capped so there is no surprise bill, which makes it safe to point at real traffic on day one. Use it to establish your baseline drop rate and your true live-answer percentage with your own numbers, not a vendor's. The signup takes minutes at amdy.io/auth/signup.

The reason a serious company gives away the entry tier is that this category has been sold on claims for twenty years. Demos, reference calls, slideware. The only evidence that matters is your drop rate on your traffic against your current setup, and the only honest way to get it is to run it free. A vendor who will not let you test on real calls before paying is telling you what they think will happen if you do.

The Middle: Starter Through Scale

Most centers live between 500,000 and 25 million detections a month, which is exactly the span from Starter to Scale. The math is transparent enough to put in a formula rather than a pitch deck: monthly cost = base + max(0, detections - included) x overage rate. At Starter that means a center doing 700,000 detections pays $79 plus 200,000 times $0.00025, which is $50, totaling $129. Run your own volume through it. The answer will not require a sales call, which is the point.

The step between tiers is worth one comment. When your volume crosses a tier boundary, the overage math self-corrects: at some volume it becomes cheaper to move up a plan and take the bigger allowance, and the crossover point is simple arithmetic, visible on the pricing page. A good vendor makes tier selection your math problem, not their sales team's quota.

The Top: Partner

From $2,000 a month, custom terms, for resellers and multi-site operators with specific volume shapes. Listed here for completeness, not aspiration. If you need it, you already know, and the conversation is about your numbers, not about features.

How To Read The Meter Without Becoming An Engineer

Once you are running, the operating rhythm is deliberately boring. Pick a day, Friday morning works, and read two numbers: the live-connect percentage for the week and the detection count. Compare both to the prior week. That is the entire recurring management overhead of a utility-grade detection layer.

Set your thresholds before you need them, and make them boring too. If the connect percentage drops more than a couple of points week over week, ask why before you panic. The usual culprits are ordinary and diagnosable: a new list source with different answer patterns, a carrier change, seasonal shifts in when people pick up, or genuine model drift, which is the one failure mode that deserves its own reading and has one at AMD model drift. If the detection count jumps unexpectedly, that is a campaign or pacing question, not a detection question, and it belongs to whoever owns the dialer.

Notice what is absent from this rhythm. There is no upgrade planning meeting, because a service that improves server-side ships without a maintenance window on your side. There is no capacity review, because the allowance is volume, not hardware. There is no version matrix to check against your dialer build. The absence of these rituals is the delivered product. Most of what vendors call an upgrade cycle is really a scheduled opportunity to invoice you, and the entire point of utility grade is that the opportunity never arises.

One caution before moving on, because a manifesto that only cheers is a brochure. Set-and-forget applies to the infrastructure, never to the measurement. The week you stop reading the meter is the week it starts lying to you quietly. Utilities are trustworthy, not telepathic, and the three minutes on Friday is the full price of that trust.

The Costs That Are Not On The Invoice

Three costs never appear on any vendor's bill but always appear somewhere in your business if the detection is poor.

The first is the dropped human. Every live person your AMD classifies as a machine is wasted acquisition spend plus a burned number. At a 10% drop rate on a modest 500,000-detection month, that is 50,000 humans hung up on by software, roughly, adjusted by your true live-answer rate. The formula for the leak: dropped humans = detections x live-answer rate x drop rate.

The second is the payroll tax. Every voicemail routed to an agent is agent minutes spent listening to a greeting. The formula: wasted hours per month = misrouted voicemails x average minutes each / 60 x agent hourly cost. Plug in your own numbers and try not to wince.

The third is the compliance shield cost. The FTC's amended Telemarketing Sales Rule caps abandonment at 3% of answered calls per day. Detection quality and abandonment are mechanically linked, and if a regulator ever asks, your defense is per-call logs, not testimony. Poor detection is not just a revenue issue. It is an unpriced risk sitting on your balance sheet, and unlike the other two costs, this one arrives all at once.

A Closing Manifesto For The Non-Technical Reader

Here is the whole category in four sentences, compressed from everything above and everything in the earlier pieces.

The Money Leak

Default detection drops 10 to 20% of live humans. Those are paid-for humans. Measure the leak with the formula, price it at your cost per lead, and you will find it is larger than any vendor's invoice.

The Payroll Tax

Misrouted voicemails consume agent salaries. The fix costs less per month than one agent's shift in most centers. The tax is optional. Keep paying it only if you enjoy it.

The Compliance Shield

Per-call logs plus accurate verdicts plus the 3% abandonment cap means your detection layer is also your evidence layer. A vendor without logs is selling you horsepower without a speedometer. If that vendor is ever in the room when a regulator asks questions, you will wish you had chosen differently. The drift risk deserves respect too, which is why I keep a standing piece on AMD model drift.

The Trial-First Buying Rule

Never buy this category on a demo. Demand a free tier or a paid pilot on your own traffic, read your own drop rate, and decide from the meter. We built the free tier because we want that rule to be universal, even when it costs us deals to vendors who prefer the demo route. The latency and accuracy tradeoffs you should check during any trial are written up in AMD latency versus accuracy, and the insurance vertical's version of the story is in AMD for insurance outbound. The feature list behind all of it lives on the features page.

That is the reading list done. The category is understood. The meter is waiting at the signup page, 50,000 detections a month, no card, five minutes. The only remaining question in this entire field is the one your dialer answers a hundred thousand times a day without you watching: of the humans who say hello, how many actually reach one of yours?