What running an AI operations system actually costs, line by line

Almost nobody publishes the cost stack behind an AI operations system, which is convenient for vendors and useless for buyers. Here is ours, with real ranges, so you can tell whether a quote you are given is reasonable.

First, the thing everyone gets wrong about 24/7

Running around the clock does not cost extra by itself. Nothing in the stack charges by the hour of availability. Voice bills per minute of actual conversation. The model bills per token actually processed. Texts bill per message. An idle system at 3am costs the hosting fee and nothing else.

What 24/7 actually costs is reliability engineering. If the system is down at 2am, a no-heat call goes unanswered after someone promised it would not. That is not a vendor bill, it is a design requirement: health checks plus an automatic fallback that routes calls to the existing voicemail or on-call number the moment the system stops responding. Ten to thirty dollars a month in uptime monitoring, and it should be written into the contract.

The five lines that make up the bill

1. Voice minutes

The largest and most variable line. Advertised platform rates sit around $0.05 to $0.12 a minute, but the all-in figure after transcription, the model, voice synthesis and telephony lands closer to $0.10 to $0.31. The gap between advertised and actual is commonly three to six times, because the pricing page shows the platform fee and not the four things underneath it.

For a mid-size property manager taking roughly 450 after-hours and overflow calls a month at three minutes each, that is about 1,350 minutes, or $135 to $245.

2. The model

Cheaper than most people expect, and easy to underestimate. Current API pricing runs about $1.25 per million input tokens and $4.25 per million output, with cached input far cheaper.

The trap is assuming one pass per message. Agentic work re-sends the playbook and the customer context on every tool call, so a single email might be five to ten round trips. Daily digests read a whole day of data. Budget $75 to $180 a month for a busy account, not the $30 a naive calculation suggests.

3. Telephony and messaging

Phone numbers, per-message costs, and in the US an A2P 10DLC registration with a one-time brand and campaign vetting fee plus carrier surcharges. Small but not zero: $40 to $80 a month for a few numbers and a couple of thousand texts.

4. Hosting, monitoring and storage

The always-on listener, a queue, a database, error monitoring, and call recordings held for the retention period. $50 to $130 a month.

5. Spam filtering, which nobody budgets for

After-hours business lines attract a lot of robocalls, and you pay per minute even while the agent politely ends a call with a robot. A spam-score lookup on every inbound call costs $10 to $40 a month and saves several times that in wasted minutes.

What this means when you are buying

Three things to take into any vendor conversation:

• Ask what happens when volume spikes. A flat monthly price with unlimited minutes is either capped somewhere in the contract or priced for the worst case. Both are fine; not knowing which is not.
• Ask about a fair-use line. A reasonable structure is a generous included allowance with a stated per-minute rate beyond it. That protects both sides and makes the bill predictable.
• Be suspicious of very low prices on high-volume accounts. If someone quotes $99 a month for something handling two thousand call minutes, the arithmetic does not work, and the usual resolution is a tool that answers quickly and does nothing else.

How we handle this

The two weeks of supervised running in our Workflow Package are also a measurement window. We meter minutes and tokens per property, then set the fair-use line from real numbers rather than benchmarks. If your usage turns out to be materially higher than assumed, you hear it in week three rather than in a surprise invoice in month four.

Sources

• AI Voice Agent Pricing 2026: per-minute cost of 10 platforms and AI Voice Agent Cost per Minute, 2026 for voice stack economics.
• Meta Model API pricing and rate limits for token costs.
• AI Automation Agency Cost, 2026 for agency-level tooling cost benchmarks.