A voice agent's cost is the sum of five components: telephony, speech-to-text, text-to-speech, the language model, and orchestration. Observed 2026 bands put a self-assembled stack at $0.05 to $0.15 per minute all-in, and a premium managed service as high as $0.40, which works out at roughly $0.30 to $0.50 per completed two-to-three-minute call.
This is a bottom-up cost build. It sums five per-minute component rates, multiplies by the minutes you expect to consume, adds any fixed fee, and compares the total against a staffing cost you supply and against published self-serve list prices. It models the inputs you entered, not a quote.
Every voice agent, whoever sells it to you, is assembled from the same five parts, and every price you are ever quoted is those five parts plus a margin. Knowing the parts is what turns a quote into something you can evaluate. Telephony carries the call. Speech-to-text turns the caller's audio into words. A language model decides what to say. Text-to-speech turns that back into audio. Orchestration is everything that holds the loop together: call state, barge-in handling, function calls out to your calendar or CRM, logging, retries, and the vendor's margin. The observed 2026 bands for each are published in the table above. Add the floors and you land at $0.046 a minute; add the midpoints and you land near $0.13; add the tops and you reach $0.22. That spread, on the same functional product, is why per-minute pricing quoted without a breakdown tells you almost nothing.
Two different things are being sold under one label. At the bottom of the range you are buying components and assembling them yourself. At the top you are buying an operated service, and most of the difference is the orchestration line, which is the only component whose band has no ceiling. A self-assembled stack lands at $0.05 to $0.15 a minute all-in. That number is real, and it is also incomplete, because it prices none of the engineering time to build it, none of the evaluation work to keep it from booking the wrong appointment, and none of the on-call burden when a carrier degrades at 4pm on a Friday. A premium managed service runs as high as $0.40 a minute. You are paying for someone else to own the integrations, the prompt regressions, the voice quality complaints and the pager. Whether that is expensive depends entirely on whether you were realistically going to do that work yourself. For a normal two-to-three-minute booking call, both models land at roughly $0.30 to $0.50 per completed call once everything is counted. That is usually the number worth arguing about, because it is the one that compares cleanly to what the call is worth.
This calculator will not tell you what a receptionist costs, because we cannot publish a salary figure that is true across markets, and inventing one would be exactly the kind of number this site has been cleaned of. Instead you type in your own fully loaded hourly cost, and the model does the comparison on your figure. Fully loaded means more than the wage. Include payroll taxes, benefits, paid time off, recruitment amortised over expected tenure, the desk, the software seat, and the supervisor's time. A common mistake is comparing a per-minute agent rate to a bare hourly wage, which flatters the agent by a wide margin. The comparison also has to be honest about coverage. A person covers the hours you staff. An agent covers all of them, and the fair comparison for after-hours volume is usually against an answering service or against nothing at all, not against your existing front desk. The other half the arithmetic misses: a person handles the walk-in, calms the angry customer, and notices that the same caller has phoned three times this week. An agent does not. The realistic deployment for most businesses is overflow and after-hours rather than replacement.
The self-serve market publishes flat monthly entry prices: Rosie at $49, Goodcall at $59, myaifrontdesk at $65, and Smith.ai at $95. Those are real list prices and they are genuinely cheap for low volume. The thing to check before comparing them to a per-minute build is what the entry tier includes. Flat plans cap minutes or calls, and the overage rate is where the economics actually live. At low volume a flat plan is almost always cheaper than assembling anything. At high volume the per-minute maths reasserts itself, and the crossover point is worth calculating rather than assuming. The second thing to check is what the agent is allowed to do. Answering and taking a message is a different product from booking into your calendar, checking availability, taking a deposit and writing the job back to your CRM. Cost per minute is only comparable between agents doing comparable work.
Per-minute pricing is a usage rate. It is not the cost of the deployment, and treating it as one is the most common budgeting error in this category. The costs that sit outside the meter are the ones that decide whether the project succeeds: mapping what the agent is allowed to say and do, connecting it to the calendar and the system of record, building an evaluation set so you can tell whether a prompt change made things worse, recording and reviewing real calls, and the escalation path for when the agent should hand off to a person. For a production deployment, that build work is usually larger than the first year of usage cost. It is also the work that most distinguishes a deployment that quietly books jobs from one that gets switched off in month two.
Observed bands put a self-assembled stack at $0.05 to $0.15 per minute all-in, and a premium managed service at up to $0.40. For a normal two-to-three-minute call that is roughly $0.30 to $0.50 per completed call.
Text-to-speech and the language model are the largest fixed contributors, at $0.015 to $0.030 and $0.010 to $0.060 per minute respectively. But orchestration and margin has the widest band, $0.010 to $0.100 and upward, and it is what separates a cheap stack from an expensive one.
Usually yes on a pure per-minute basis, but the comparison only holds if you use a fully loaded staff cost including taxes, benefits, downtime and supervision, and only if you are honest that an agent does not do the non-phone half of the job.
Because they cap volume. Flat entry tiers are genuinely cheaper at low call counts. Find the included-minutes limit and the overage rate, then use this calculator at your actual volume to see where the crossover sits.