What a Voice AI Agent Actually Costs to Run Per Month

I did a demo call with a roofing company owner last month, walked him through the whole flow — inbound call, agent qualifies the lead, books the inspection straight into his calendar. He liked it. Then he asked the question everyone asks eventually, usually right after they stop being impressed and start doing math in their head.

"Okay but what's this actually going to cost me a month?"

Most people in this space don't want to answer that honestly. You'll see a flat "$997/mo, all in" on a sales page, and that number is doing a lot of work to avoid telling you what's underneath it. Or you'll get quoted per-minute, and the person quoting conveniently forgets that four other things also bill per minute and they all stack.

So I'm just going to open the whole thing up. If you run a clinic, a dealership, a contracting business, whatever — you should be able to read this and rough out your own number by the end of it.

The short version

A production voice agent runs somewhere around $0.10 to $0.20 a minute in raw platform cost. That's the floor. Everything above that is either an agency markup or one of the hidden costs further down this page.

Say you take 300 calls a month at roughly 3.5 minutes each. That's about 1,050 minutes, which lands you around $110 to $210 a month just in infrastructure. Add a monthly management fee on top and a well-run setup usually ends up somewhere in the $500–$1,500/mo range, all in.

Now, where does that per-minute number actually come from? This is where a lot of the sleight of hand happens, so let's go line by line.

1. Telephony

The actual phone line. Twilio, Telnyx, Vonage — or more likely, whatever your platform is reselling under the hood.

There are two things almost nobody catches until it bites them. First — if the agent transfers a call to a human, you're paying for two legs running at once: the inbound leg to your agent, and a separate outbound leg to whoever it's transferring to. So a call that's 6 minutes with a 3-minute warm transfer tacked on effectively bills like 9 minutes, not 6.

Second, if you're sending SMS confirmations (you should be, it's genuinely one of the higher-value things a voice agent can do), you need 10DLC registration in the US first. Skip that step and your texts just get silently filtered by carriers. You won't find out until a client calls asking why their customers never got a confirmation text, and by then it looks like your system is broken rather than unregistered.

Telephony is cheap. It's also the line item most likely to get you in trouble if you don't set it up properly.

2. Speech-to-text

This is what turns the caller's voice into text the model can read. Deepgram covers most of the market here, and the price is basically a rounding error:

• Streaming STT: $0.004–$0.006 per minute

That's not a typo. Six-tenths of a cent a minute.

But don't let the price fool you into thinking it doesn't matter — it's arguably the single biggest factor in whether an agent sounds sharp or sounds broken. If the transcription is off, the LLM ends up confidently answering a question the caller never asked. I had a roofing client where callers kept saying things like "there's a leak by the soffit," and one early STT setup kept hearing "soft it." The model would then try to be helpful about something that doesn't exist. That's not a model problem, that's a bad transcript problem, and it looks identical to the caller either way.

Don't cheap out here. There's nothing to cheap out on, it's already almost free.

3. The LLM

The part everyone assumes is the expensive one. It's not, usually — but it's the part where cost actually moves depending on decisions you make.

Voice is a strange use case for a model because it gets called on every single turn of the conversation, and every call re-sends the whole conversation history up to that point. Turn one might be 500 tokens. Turn fifteen could be 4,000. The cost inside a call isn't flat, it climbs as the conversation goes on.

Rough numbers per minute:

Most of what I ship runs on a mid-tier model. Frontier models handle a caller who changes their mind halfway through a sentence noticeably better, no argument there — but they're also slower, and in a phone call, latency isn't a minor detail, it's basically a feature of the product. Wait 1.8 seconds too long before responding and the caller says "hello? you there?" and now the whole exchange feels off, even if the eventual answer was perfect.

Here's the actual lever most people miss: it's not which model you pick, it's how long your system prompt is. I've seen prompts pushing 4,000 tokens because someone crammed in every FAQ the business has ever gotten. That entire block gets resent on every single turn, twenty turns a call, times however many calls a month. You're paying rent on words the caller will never once hear. Trim it down, move the reference material behind a tool call the agent only pulls when it's actually needed, and the bill drops by something like 40% without touching the model at all.

4. Text-to-speech

This is where most of the money actually goes.

One thing that trips people up: TTS bills on characters spoken, not on call length. So if your agent is only actually talking for 40% of a 4-minute call, you're being billed for roughly 1.6 minutes of speech, not 4. Which means an agent that listens more than it talks is quite literally cheaper to run — not just a nicer experience, an actual cost saving.

Which works out well, because a good agent probably should talk less anyway. Nobody wants a receptionist that monologues at them.

Is ElevenLabs worth paying three times what Deepgram Aura costs? Depends what the call is for, honestly. If you're a high-ticket roofing outfit and one booked inspection is worth $8,000, sure, spend the extra few cents a minute, it's nothing next to the outcome. If you're running reminder calls to 4,000 people about a dental cleaning, use the cheap voice — nobody's evaluating your brand on a reminder call.

5. The platform layer

Vapi, Retell, Bland, or something custom built on LiveKit or Pipecat. This is the layer that glues STT, the LLM, and TTS together — handles interruptions, turn-taking, fires tool calls when the agent needs to check a calendar or a CRM.

Vapi runs around $0.05/min for its own layer, on top of whatever the underlying models already cost you. Others land in a similar range, and some just bundle everything into one blended per-minute number, usually somewhere around $0.09–$0.15.

You can absolutely build this yourself and skip the fee — I've done it. But unless you're doing serious volume (I'd say north of 50,000 minutes a month before it's worth it), you'll burn more engineering time on interruption handling and reconnect logic than you'd ever save. Just pay the five cents.

Three scenarios, worked out

A — Solo contractor or small clinic

~120 calls/mo, avg 3 min, one phone number, budget TTS

B — Established roofing company

This is honestly the most common profile I see, so it's worth the detail.

~350 inbound calls/mo, avg 4 min, premium voice, SMS confirmations, 15% of calls transferred to a human

C — Outbound lead qualification, at volume

3,000 dials/mo, roughly 35% connect rate, avg 2.5 min on the calls that do connect

Worth noticing: going from Scenario B to C is roughly 10x the dial volume, but nowhere near 10x the cost. Most outbound dials never connect — voicemail, no answer, a disconnected number — and those cost you almost nothing beyond a few seconds of telephony. No LLM spend, no TTS spend. Outbound is cheaper than people assume going in. The hard part was never the infrastructure bill, it's building a list worth dialing.

What nobody puts in the proposal

This is the part I genuinely wish someone had sat me down and explained before my first build, because I learned it the expensive way.

Testing minutes are real money. Getting an agent tuned means calling it, over and over, adjusting, calling it again. I'll usually burn 300 to 600 minutes on a build before a single paying customer ever dials the number. That's not free, and if your quote doesn't account for it, guess who's eating the cost.

Dead air adds up too, in a small way. Someone picks up, says nothing, hangs up twenty seconds in — you still paid for the connected call, the greeting TTS, a couple of LLM turns that went nowhere. It's a few dollars a month in the grand scheme, but if you're not filtering it out of your analytics, it also makes your numbers look worse than they are.

Then there's the automation layer underneath all of it — n8n or Make, pushing data into a CRM, a spreadsheet, a calendar. Self-hosted n8n on a cheap VPS runs maybe $6 to $12 a month. Cloud plans are more like $20 to $50. Not expensive, but it's a real line item that tends to get left off the estimate entirely.

Storage matters more than people think if you're keeping call recordings and transcripts for compliance — usually just a few dollars a month, sometimes more if you're in a regulated industry. And when something fails downstream, like a calendar API timing out, the agent retries, and retries cost tokens too. Small, but it's not zero.

The one that actually matters long-term is prompt maintenance, and it's the one almost nobody budgets for. A voice agent isn't a website you build once and leave alone. It drifts — a new service area gets added, pricing changes, a competitor runs a promo, business gets seasonal. The agent that sounded great in March can sound mediocre by August if nobody's touched it. Somebody needs to actually listen to call recordings every couple of weeks and fix what's drifted.

Where a fee fits into this, and what's fair

From what I've seen across the market, pricing tends to fall into a few buckets:

Personally I lean toward a setup fee plus a flat monthly, with infrastructure passed through at cost. The client can see exactly what the models cost, and I get paid for the part that's actually worth paying for — judgment, not tokens.

The thing I'd push back on is someone charging a flat $2,000/mo with zero visibility into what's actually being spent underneath. That's not automatically a ripoff — it could be a genuine bargain if the agent's booking six jobs a month — the problem is you have no way to check. If a vendor won't show you a per-minute breakdown, that's worth asking about directly.

What actually moves the number

If you want to bring the bill down, roughly in order of how much it actually helps:

Shortening the system prompt is the biggest and easiest win, and almost nobody does it because it feels like busywork instead of a real fix. Getting the agent to talk less is close behind — shorter responses cut TTS cost and latency at the same time and usually make the call feel better too, which is a rare case where the cheap option and the good option are the same option. Downgrading the TTS voice on low-stakes calls (reminders, confirmations) while keeping the good voice for first-touch sales calls is an easy split most people don't bother making. Using a smaller model for genuinely simple flows — an after-hours "take a message, text the owner" agent doesn't need a frontier model — saves real money for zero downside. Ending calls cleanly and quickly matters more than it sounds like it should; agents that let callers ramble past the point of usefulness cost actual dollars. And caching reference data behind a tool call instead of stuffing the whole price list into the prompt every single turn is one of those changes that feels small and isn't.

The comparison that actually matters

None of these numbers mean much floating on their own, so here's what they're up against.

A human answering service runs $1.00 to $1.50 a minute plus per-call fees. The same 1,400 minutes from Scenario B would cost $1,400 to $2,100 a month that way, and it still can't book directly into your calendar. A full-time receptionist in the US runs $3,000 to $4,500 a month once you count payroll taxes, and works about 40 hours a week — meaning nobody's answering the phone at 9pm on a Sunday when someone's roof is actively leaking. Voicemail is free, and depending who you ask, it loses somewhere between 60% and 80% of callers to whoever picks up next.

Against that, roughly $200 in infrastructure plus a management fee isn't really a close call. And to be clear, it's not because the AI is smarter than a good receptionist. It isn't — a sharp human beats it in plenty of ways. It's that the AI answers on the first ring at 2am, and the competitor down the street doesn't.

What I'd actually tell you across a table

If there's one thing worth taking from all this: the infrastructure cost is not the number worth obsessing over. Two hundred dollars a month is noise for most businesses that would even consider this. The number that actually matters is whether the thing books jobs, and that comes down to how well it's written and whether anyone bothers improving it after launch.

I've watched a $40/mo agent quietly outperform a $900/mo one, purely because one person actually sat down and listened to the call recordings and fixed what was broken instead of assuming it was fine.

So ask your vendor for the per-minute breakdown. Ask what happens when a caller talks over the agent mid-sentence. Ask who's actually listening to the recordings by month four. Those three questions will tell you more than any pricing sheet ever will.

---

A note on the numbers: these are roughly current as of mid-2026, and pricing in this space moves fast — mostly downward, model costs especially. Check current rates before committing to anything long-term. And if you're billing in INR, remember the underlying costs are USD on someone's card, so leave yourself some FX room.

If you want help running these numbers for your own call volume, I do this for contractors and service businesses fairly often. Happy to just tell you the number, no pitch attached.