Code three hundred calls before you pick a single intent
Everyone has a story about a phone system that would not let them reach a person. Those stories are not about voice technology; they are about design decisions somebody made to protect staffing at the caller's expense. Modern speech systems are good enough that a well-built one is genuinely faster than waiting on hold. They are also good enough to hide a badly designed flow behind a pleasant voice, which is how a company ends up with a lower answer time and a worse reputation.
Start where you would start with any operational problem: measure. Pull three hundred recorded calls at random across a full week, including the Monday morning surge and the Friday afternoon lull, and have two people code each one for the reason the caller phoned. Not the reason they said first — the reason they hung up satisfied or not.
The pattern is remarkably consistent across businesses that answer a lot of phones. In a multi-site medical or dental group, scheduling and rescheduling is usually thirty to sixty percent of inbound volume, with prescription and billing questions well behind. In a distributor, order status and product availability dominate. In a home services contractor, it is booking, appointment windows, and “is the technician coming today.” In a restaurant group it is hours, reservations, and large-party questions, with a heavy after-hours share nobody is currently answering at all.
You are probably here because
- The phones ring more than your staff can answer and abandonment is climbing
- Calls that arrive after five o'clock go to a voicemail box nobody empties
- A vendor quoted a containment number and you cannot tell what it counts
- Your front desk is doing scheduling instead of the work in front of them
The first two are the strongest cases for automation. The third is the section on measurement. The fourth is often solved better by online self-service than by a phone bot.
What is automatable, and what is not
The line is clearer than most vendors admit. Automate calls where the answer is bounded, the caller is calm, and a system of record holds the truth. Do not automate calls where the caller is upset, where the request is ambiguous, or where money moves in a direction that is hard to reverse.
Good candidates. Confirming or moving an appointment. Order and delivery status against a real record. Hours, location, directions, and what to bring. Refill requests that get queued for a human to approve. Taking a message with the right fields filled in. Routing to the right department based on what the caller actually said instead of a nine-option menu.
Poor candidates. Anything with a complaint in it. Anything where the caller does not know the vocabulary — a patient describing a symptom, a homeowner describing a noise. Collections. Cancellations, which are almost always a retention conversation. And any first call about an emergency; if your line takes “no heat in February” or “water in the basement,” the automation's only job on that call is to recognize it in one turn and get a human on immediately.
Latency is the experience
Everything about whether a voice system feels natural comes down to timing. In ordinary conversation the gap between one person finishing and the other starting is roughly two hundred milliseconds. People tolerate more from a machine, but not much: once the silence after they stop talking passes about eight hundred milliseconds, callers assume they were not heard and start talking again, and now both parties are talking at once.
So there is a budget, and it is tight. Detecting that the caller finished speaking is usually two hundred to four hundred milliseconds. Deciding what to say, including any lookup against your systems, needs to land in a few hundred milliseconds. Starting to produce audio takes another hundred or so. That whole chain has to fit inside the tolerance, which means the database query that takes 1.4 seconds is not a performance nit — it is the difference between a conversation and an argument.
Three engineering choices matter more than model selection. Stream everything, so audio begins before the full response is composed. Support interruption — if the caller talks, the system stops talking immediately, and what it had already said counts as said. And when a lookup will genuinely take two seconds, say something first: “let me pull that up” costs nothing and converts dead air into normal human behavior.
| Stage | Reasonable budget | What blows it |
|---|---|---|
| Detecting the caller stopped | 200–400 ms | Aggressive silence thresholds that cut people off mid-sentence |
| Understanding and deciding | A few hundred ms | A synchronous lookup against a slow system of record |
| First audio out | 100–300 ms | Generating the whole reply before speaking any of it |
| Handling interruption | Immediate | No barge-in support, so the caller talks over a monologue |
| Transfer to a person | As fast as your queue allows | Re-asking everything the caller already told the bot |
Names, addresses and numbers are the hard part
Speech recognition on ordinary conversational English over a decent connection is very good. Speech recognition on a surname, a street name, an email address or a policy number, spoken by someone in a truck with the window down, is a different problem, and it is where these systems most often fail in a way that costs real money.
Three practical rules. First, do not collect by voice what you can collect another way — look the caller up by the number they are calling from, and confirm rather than ask. Second, always read back anything consequential, in the caller's own terms: “that is Tuesday the twelfth at two-thirty with Dr. Alvarez, is that right?” Read-back catches the error while it is still free. Third, for anything alphanumeric, offer the keypad. Asking someone to spell a policy number aloud will be wrong often enough to matter; typing it will not.
And know your population. A system that works beautifully for a general consumer base can fail badly for an older patient population, for callers in a noisy plant, or for a customer base with names outside the recognizer's comfort zone. Test on real calls from your real callers, not on the vendor's demo set.
What decides whether callers hate it — our weighting
Weights sum to 100. Our judgment, not a measurement. The last row is where most demos spend their time.
The escape hatch has to be real
One rule above all others: a caller who wants a person gets one, quickly, without needing to guess a magic word. “Representative,” “operator,” “agent,” “person,” an exasperated sigh followed by silence, and pressing zero should all work. So should saying it while the system is still talking.
The temptation to make this harder is real, because containment is the number the project gets judged on and every escape is a lost containment. Resist it completely. Every business that has made the exit hard has traded a metric for goodwill, and goodwill is the more expensive of the two. If the automation is good, most callers will not want the exit; if it is not, hiding the exit does not make it good, it makes the eventual call angrier.
Two other exits deserve the same treatment. If the caller says something the system does not understand twice in a row, transfer — do not try a third time. And if the caller's tone or words indicate distress, transfer immediately, even mid-flow. A person on hold is a person waiting; a person stuck is a person leaving.
The transfer packet is not optional
When a call escalates, the person picking it up should already have the caller's identity, the reason as the system understood it, everything already collected, and a transcript on screen. Without that, the automation did not save the caller time — it charged them a toll before the real conversation. This one piece of engineering is what separates callers who say “that was fine” from callers who say “I had to explain it all twice.”
After hours is where most of the value is
Look at the hourly distribution of your call volume next to your staffing, and the case usually writes itself. A home services company gets its emergency calls at night and its booking calls at seven in the morning. A restaurant group gets its reservation calls during dinner service, when nobody can answer. A dental group loses Monday morning appointments because eleven people call in the first twenty minutes and four of them hang up.
Those calls are currently going to voicemail, to an answering service that fills in a form, or nowhere. Automating them is not taking work away from staff; it is capturing revenue that currently leaks. It is also much lower risk than automating the busiest daytime queue, because the alternative the caller is comparing against is not a person — it is a beep.
Start there. A first deployment scoped to after-hours booking and message-taking, with clean transfer to an on-call phone for anything urgent, is a small project with an unusually clear payback and almost no downside if it underperforms.
Outbound is a different animal, and often should not be voice
Appointment reminders and confirmations have the best return of anything in this category, and for most populations a text message with a link outperforms a call by a wide margin. It is cheaper, it is asynchronous, and people answer it. The exceptions matter though: older patient populations often prefer and respond better to a call, and field crews with poor data coverage may not see the message at all. Ask, rather than assuming, and let people choose.
Wherever you do place automated outbound calls, the compliance picture is genuinely complicated and the penalties are assessed per call. Consent requirements, the rules around automated dialing, and calling-hour restrictions all apply, and they vary by state and by what the call is about. Recording rules vary too — some states require every party to consent. This is one of the few places in a software project where the right first meeting is with your counsel, not your vendor, and doing it in the wrong order is expensive.
Disclose that the caller is speaking with an automated system. It is required in some places, it is good practice everywhere, and in our observation it reduces rather than increases frustration, because callers calibrate their expectations and stop testing it.
Send us your call reasons and we will tell you what is worth automating.
A week of call volume by hour, a coded sample of reasons, and the name of the system that holds your appointments or orders. Email contact@precisionfederal.com and you get a written read on which reasons are safe to automate, which are not, and roughly what share of volume that covers. One business day, no charge.
contact@precisionfederal.comMeasure the things that reveal whether callers were helped
Containment — the share of calls that finish without a human — is the number vendors quote and it is easy to inflate. A system that traps people has excellent containment. A system that handles only the four easiest intents and routes everything else can quote a high containment rate on the subset it chose to count. Ask exactly what the denominator is.
Four numbers together tell the truth. Repeat-call rate within twenty-four hours, which is the honest measure: if the caller phoned back, they were not helped, whatever the transcript says. Abandonment during the automated portion, which is people giving up. Handle time on transferred calls, which tells you whether the transfer packet is working — if transferred calls take longer than they used to, the automation is costing your agents time. And completion of the specific task: did the appointment actually get booked, in the actual calendar.
One consequence to plan for that surprises people. If the automation absorbs the easy calls, what remains for your staff is only the hard ones. Average handle time goes up, first-call resolution goes down, and both of those look like a regression on the dashboard while being exactly what should happen. It also means the emotional load on the team increases, because every call is now a difficult one. Tell the team in advance, adjust the targets, and expect to need fewer people but better-supported ones.
What it costs and how long it takes
A first production line covering four to eight intents, integrated with one system of record, with real escalation, a transfer packet and disclosure handled properly, is typically eight to fourteen weeks. Build cost usually lands in the mid-five to low-six figures depending on integration difficulty. Per-minute platform and model costs are real but rarely the deciding factor at ordinary volumes; the integration and the flow design are where the money goes.
Then it needs tending. Someone listens to a sample of calls every week — twenty is enough — and fixes what they hear. New intents arrive as the business changes. Prices, hours, policies and staff names change, and a system that confidently states last quarter's hours is worse than no system. Budget a few hours a week, indefinitely, and give it to someone who knows the business rather than the software.
When you do not need this
Under roughly a hundred and fifty calls a day, spread across many different reasons, a voice system will cost more to build and maintain than it returns. If most calls are the same three questions, publishing the answers clearly and putting the two most common on a short, well-designed greeting will absorb a surprising share for almost nothing. And if the underlying problem is that your online booking is bad, fix the booking — people are calling because the website failed them, and a phone bot papers over that at higher cost.
The mistakes we get called in to fix
- No reliable path to a person, because containment was the only metric on the dashboard
- The caller re-asked everything by the agent who picked up the transfer
- Silence after the caller stops talking, so both sides speak at once for the whole call
- No barge-in, forcing callers to sit through a monologue they already answered
- Names and numbers collected by voice with no read-back and no keypad option
- An emergency call handled as a routing question for four turns before escalating
- Containment quoted on a hand-picked intent set, with repeat calls never measured
- Nobody listens to calls after launch, so the system quotes hours that changed in March
Before you approve the build
- Three hundred real calls coded by reason, by two people
- Automatable and not-automatable reasons written down separately
- A latency budget with a number for each stage
- Barge-in supported and tested on real calls
- Every consequential value read back to the caller
- Keypad entry offered for anything alphanumeric
- Escalation on any of a dozen phrasings, plus zero, plus repeated confusion
- Transfer packet: identity, intent, collected fields, transcript
- Consent, disclosure and calling rules cleared with counsel first
- Repeat-call rate and transferred handle time on the dashboard from day one
Bottom line
Phone automation earns its place when it answers calls nobody was answering and finishes tasks people were waiting on hold to complete. It burns goodwill when it becomes a wall. The engineering that separates the two is unglamorous: a tight latency budget, interruption handling, read-back, an escape hatch that always works, and a transfer that carries everything the caller already said. Start with the hours you are not covering, measure repeat calls rather than containment, and have somebody listen to twenty calls a week forever.
Frequently asked questions
It depends entirely on what your callers phone about, which is why coding three hundred real calls comes first. In businesses where one or two bounded intents dominate — scheduling, order status, hours — taking fifteen to thirty-five percent of total inbound volume in the first year is a reasonable target. Vendor containment figures well above that are usually measured on a chosen subset of intents rather than on all calls.
Three things, in order: not being able to reach a person, being asked something they already answered, and having to repeat everything to the agent after a transfer. Voice quality and how human the system sounds matter far less than any of those. A plain, slightly robotic voice that never traps anyone will outperform a beautiful one that does.
For most populations a text with a link performs better and costs less. Calls remain the better channel for some older populations and for people working where messages go unread. Offer both, record the preference, and honor it. Whichever you choose, the consent and calling-hour rules apply and vary by state, so clear the approach with counsel before building.
Usually it changes the mix before it changes the count. The automation absorbs the simplest calls, so what reaches your team is the harder residue — average handle time rises, first-call resolution falls, and the emotional load goes up even as total volume drops. Plan targets and support for that shift, or it will look like a failure on the dashboard.
Typically eight to fourteen weeks and a build cost in the mid-five to low-six figures for four to eight intents against one system of record, with proper escalation and transfer. Per-minute platform costs are usually a small share at ordinary volumes. Budget a few hours a week afterward for someone who knows the business to listen to calls and keep the flows current.