Quick answer: A fully-loaded human talk-minute costs roughly $0.60–$1.50 once salary, benefits, shrinkage, and supervision are counted, while an all-in AI voice-agent minute — speech-to-text, model, text-to-speech, and telephony summed — runs roughly $0.05–$0.30. That 5–10× gap closes only if the AI agent never finishes a call without help, so the real TCO question is not "AI or human" but "what share of each call does the AI resolve, and at what blended cost?" This post prices both sides from published rate cards and shows the break-even every buyer should run before signing anything.
1. The fully-loaded human talk-minute
A seat in a contact center is not a voice for an hour; it is an on-the-clock hour of which only a fraction reaches a caller's ear. Building the fully-loaded number means stacking every line a CFO actually funds, then dividing by the minutes that were genuinely spent talking:
- Wages and payroll burden — base salary plus employer-side taxes and benefits, typically 25–40% on top of gross pay in a support market.
- Shrinkage — breaks, training, meetings, absence, and idle queue time. Industry contact-center shrinkage runs 30–35%, so every paid hour delivers about 65–70 minutes of work before the next deduction.
- QA and coaching overhead — supervisors, quality analysts, and calibration sessions exist to keep human handling consistent, and their cost lands on the same talk-minute denominator.
- Seat licence and telephony — contact-center seat licences and per-seat telephony both scale per head, which is why traditional UCaaS consolidation prices per agent rather than per call.
Run the stack and a typical support operation lands in the $0.50–$1.50 per talk-minute range that the AI voice agent pricing guide already uses as its ROI baseline; this post works the middle of that band at $0.60–$1.50 so the example stays conservative. The number is a range because wage markets and handling policies differ — the lesson is the same either way: the per-minute cost of a human is a five-to-fifteen-fold multiple of the AI number below, before any AI call is even routed.
2. The AI voice-agent stack, per talk-minute
The AI side is a sum of four metered components, exactly the four-part split published in the AI voice agent pricing guide: speech-to-text (~$0.005–$0.02/min), the language model (per-token, highly variable), text-to-speech (~$0.01–$0.30/min depending on the voice), and telephony (~$0.004–$0.02/min on US destinations). Sum them on a typical support or booking configuration and the all-in figure lands in $0.05–$0.30 per talk-minute, a range this post uses end to end.
Two cautions carry over from that pricing guide and belong in any TCO spreadsheet. First, the model line is the least predictable — prompt caching and tiering small models for routing bring it down, long prompts and tool calls pull it up. Second, a per-minute headline without your voice, your model, and your destinations is marketing, not accounting; the table below uses the published range so a buyer can substitute their own quote.
3. Break-even table: hybrid cost by volume and complexity
The honest comparison is a hybrid escalation model: route every interaction to the AI agent first, let it resolve a share of calls on its own, and hand the remainder to a human at the human rate. The blended cost per talk-minute is a weighted average with the resolution share r:
> Blended $/min = r × (AI all-in rate) + (1 − r) × (human fully-loaded rate)
Three complexity tiers anchor the resolution share used below — a FAQ flow resolves ~85% of interactions, a triage flow ~70%, and a genuinely complex flow ~40% — figures that track the mid-range deflection numbers production deployments report and the 60% worked example in the pricing guide. At four minutes per average interaction (the same assumption the pricing guide's worked example uses), the monthly blended cost comes out as follows:
| Monthly interactions | FAQ tier (85% resolved by AI) | Triage tier (70%) | Complex tier (40%) |
|---|---|---|---|
| 100 | $53 – $192 | $86 – $264 | $152 – $408 |
| 1,000 | $530 – $1,920 | $860 – $2,640 | $1,520 – $4,080 |
| 10,000 | $5,300 – $19,200 | $8,600 – $26,400 | $15,200 – $40,800 |
Each cell is the formula above evaluated against the full band — low end of the AI and human ranges for the left number, high end for the right — so the table never depends on a single point estimate. For reference, the all-human floor for those volumes is $240–$600, $2,400–$6,000, and $24,000–$60,000 per month, and the all-AI ceiling is $20–$120, $200–$1,200, and $2,000–$12,000.
The pattern the table exposes is the whole argument. Even at the complex tier — where a human handles three of every five escalated calls — the blended line undercuts the all-human line across the volume ladder, and the saving compounds with volume rather than saturating. Break-even is therefore rarely "does AI pay off?" and almost always "how much of the FAQ and triage traffic can we shift into the first two tiers?"
4. Hybrid escalation: AI-first with a human handoff
The hybrid model above is a deployment pattern, not a spreadsheet trick, and it is the same entry-point decision the AI-agent-first vs UCaaS incumbent guide frames structurally: an AI-first platform routes the call to the agent runtime first and escalates with full transcript context, instead of bolting an AI attach onto a per-seat phone system. Escalation design — what triggers the handoff, what context travels with it, and which queue catches it — is where resolution share is won, and it is tunable per flow rather than per vendor contract.
The handoff matters beyond cost. Consistency, policy-sensitive responses, and moments of customer vulnerability all belong on a human's desk, and a deployed hybrid model defines those triggers up front. Tenant-owned controls govern compliance-adjacent behavior here: the escalation rules, the quiet-hours and consent state on record, and the queues a handoff targets are all customer configuration, not a vendor default.
5. Where Orbit's unified wallet makes the math concrete
The reason this arithmetic is runnable on Orbit rather than theoretical is that Orbit prices on one wallet across every channel — voice agents, SMS, WhatsApp, RCS, email, video, and a built-in contact center on a single account, billed pay-as-you-go with no platform fee, no seat fee, and no monthly minimum. The published rates the pricing guide cites — $0.014 per AI voice-agent minute, $0.007 per AI message, and an optional $0.005 per resolved outcome — come straight off the pricing page, so the AI side of the break-even table is a number you can read rather than a quote you have to beg a sales team for. Both sides of the hybrid also draw from the same wallet: an AI-resolved minute and the human agent's minute in the built-in contact center settle on one bill.
Per-message pricing matters the same way on the messaging side of a flow. When a voice interaction's follow-up shifts to SMS or WhatsApp, that traffic meters per message inside the same account instead of a second contract — the blended formula above stays one equation, not two vendor rate cards reconciled in a spreadsheet. Buyers who want the deeper per-component split can cross-check against the AI voice agent pricing guide, and the true-cost calculator runs the four-stage arithmetic with your own rates interactively.
If you are pricing the telephony leg itself rather than the agent layer, the programmable voice surface published per-destination voice rates, and the AI agents documentation covers the pipeline stages and escalation hooks precise enough to mirror the break-even formula in a build plan.
Frequently asked questions
What is the break-even between an AI voice agent and a human agent?
Compare the blended hybrid cost, not the two extremes. Weight the AI resolution share against the remainder at the human rate — blended $/min = r × AI + (1 − r) × human — and the AI-first deployment beats the all-human line at nearly every resolution share, because the AI minute is priced at a five-to-fifteen-fold discount to the fully-loaded human minute.
Why is the human talk-minute number a range?
Because fully loaded means more than wages. Benefits, shrinkage, QA and coaching, and the seat licence all stack onto the paid hour before it reaches a caller, and each varies by market and policy. The $0.50–$1.50 band used here is deliberately conservative; buyers should substitute their own payroll and shrinkage figures and re-run the same formula.
What share of calls can an AI agent resolve without escalation?
It depends on the flow's complexity tier. A bounded FAQ flow typically resolves the overwhelming majority of interactions (the table assumes ~85%), a triage flow ~70%, and a genuinely complex flow around 40% — figures in line with production deflection data. Measure the actual share on a pilot before locking a TCO projection.
Does the hybrid blended-cost model apply to messaging as well as voice?
Yes — the formula weights resolution share against each leg's per-unit rate, so a voice interaction that follows up over SMS or WhatsApp meters per message on the same account inside Orbit. The arithmetic is identical; only the per-minute and per-message unit changes.
Published 18 September 2026.