AI Voice Agent Pricing: Compare Quotes and Total Cost
AI voice agent pricing for Australian buyers: compare billable units, currencies, telephony, latency, GST, integrations, and total cost of ownership.


Google Cloud Speech-to-Text V2 standard recognition rate for the first 500,000 audio minutes per account each month
Source: Google CloudGoogle Cloud Chirp 3 HD text-to-speech rate after its first one million free characters
Source: Google Cloudpage-listed GPT-Live 1 voice-session rate, with backend model and tool usage billed separately
Source: OpenAIIn this article
- 01How much does an AI voice agent cost?
- 02Why “per minute” is not one unit
- 03What infrastructure prices reveal
- 04Normalise every quote before comparing
- 05Worked example with assumptions
- 06Add total cost of ownership
- 07Treat latency as a cost and quality variable
- 08Pharmacy buyers need a privacy cost line
- 09Procurement questions worth sending in writing
- 10A simple decision rule
- 11Where Krepko and Kat fit
The safest way to compare AI voice agent prices is to rebuild every quote against one workload. A headline rate is meaningless until its unit, currency, inclusions, and rounding are clear.
Start with your own call data. Record connected calls, caller conversation time, transfer time, simultaneous peaks, outcomes, and staff follow-up. Then apply each quote to that same month.
This guide explains the arithmetic. It does not present market price ranges or rank providers.
How much does an AI voice agent cost?
There is no reliable universal figure. The total depends on call volume, duration, transfers, included services, currency, implementation, and the definition of a billable unit.
Compare named options separately in our Australian AI voice-agent guide. For a pharmacy decision, use the buyer scorecard beside the price model.
Why “per minute” is not one unit
One caller can create several billable meters. The phone network carries the call, speech recognition processes audio, a model handles the conversation, and speech synthesis creates the reply.
Recording, storage, analytics, tool calls, and a transferred human leg can add more charges. A vendor may bundle some components and pass through others.
| Pricing unit | What may be measured | Question to ask |
|---|---|---|
| Connected call | Each answered call | Are failed, short, or transferred calls charged? |
| Conversation minute | Time the caller interacts with the agent | Does silence, hold, or ringing count? |
| Telephony minute | Time on each inbound or outbound phone leg | Does a transfer create a second billed leg? |
| Model session minute | Time a real-time model session remains open | Is billing per second or rounded? |
| Audio minute | Audio processed by speech recognition | Are channels billed separately? |
| Character or token | Text or audio processed or generated | Are prompts, system text, and tool outputs included? |
| Concurrent call | Simultaneous capacity | Is capacity included, capped, or reserved? |
| Completed outcome | A defined booking, resolution, or handoff | What qualifies, and how are disputes audited? |
Never convert these units into one “effective minute” without the workload assumptions. They do not necessarily move together.
What infrastructure prices reveal
Public infrastructure prices are not finished voice-agent prices. They reveal why an all-in quote needs a component map.
The provider prices below were checked on 15 September 2026. Verify current rates, currency, and contract terms before purchase.
Google Cloud lists Speech-to-Text V2 standard recognition at USD $0.016 per audio minute for the first 500,000 minutes monthly. Requests round to one-second increments.
Channels matter. Google says a 30-second file with four channels is billed as 120 seconds. A vendor’s audio-minute count may therefore exceed call duration.
Speech output can use another unit. Google lists Chirp 3 HD at US$30 per million characters after the first million free characters. Spaces and most markup count.
Real-time models can have another clock. OpenAI lists GPT-Live 1 at $0.05 per session minute, billed per second. Backend model and tool usage is separate.
OpenAI’s page does not name the dollar currency. Do not label that rate USD or AUD without written confirmation.
Australian telephony pages can have the same ambiguity. Twilio’s Australia page displays local inbound calls at $0.0100 per minute and local numbers at $3.00 monthly.
The Twilio page does not identify a currency. Treat those as page-displayed dollar amounts, not verified Australian-dollar amounts.
The same page lists recording at $0.0025 per minute and storage at $0.0005 per stored minute monthly. Currency remains unspecified there.
These examples are cost components, not benchmarks for a complete service. A managed product also needs orchestration, integrations, testing, support, monitoring, and accountability.
Normalise every quote before comparing
Use one worksheet row for each charge. Preserve the vendor’s original unit before calculating a monthly amount.
| Field | Record exactly |
|---|---|
| Currency | AUD, USD, or another named currency |
| Tax | GST-inclusive, GST-exclusive, or not stated |
| Base fee | Monthly amount and included units |
| Usage rate | Price, unit, tier, and overage rule |
| Billing clock | Per second, partial-minute rounding, or another increment |
| Telephony | Inbound leg, outbound leg, number, porting, and transfer charges |
| AI stack | Model, speech recognition, speech generation, and tool charges |
| Storage | Recording, transcript, logs, retention period, and retrieval |
| Capacity | Included concurrent calls and burst behaviour |
| Services | Setup, integration, support, change requests, and training |
| Contract | Minimum spend, term, indexation, renewal, and exit fees |
For each quote, calculate the normalised monthly cost.
Normalised monthly cost = fixed fees + included-usage adjustment + usage charges + telephony + storage + support + other recurring charges.
Keep currencies separate until you select one documented exchange-rate source and date. State whether the final comparison includes GST.
Worked example with assumptions
The following figures are invented assumptions for arithmetic only. They are not observed quotes, recommended prices, or a market range.
Workload assumptions: 800 connected calls monthly, averaging 2.5 agent conversation minutes. Twelve percent transfer to a person, with each transferred leg averaging three minutes.
That creates 2,000 agent conversation minutes and 288 transfer-leg minutes. Assume all 2,000 conversation minutes are recorded.
Quote A assumptions: AUD $600 monthly includes 1,500 conversation minutes. Excess usage is AUD $0.42 per conversation minute.
Transfer legs cost AUD $0.08 per minute. Recording costs AUD $0.01 per conversation minute.
Quote A becomes A$600 + (500 × A$0.42) + (288 × A$0.08) + (2,000 × A$0.01) = A$853.04 monthly.
Quote B assumptions: no base fee and AUD $0.32 per conversation minute. Transfer legs cost AUD $0.14 per minute, while recording is included.
Integration and support cost AUD $180 monthly. Quote B becomes (2,000 × A$0.32) + (288 × A$0.14) + A$180 = A$860.32 monthly.
Assume every amount above is GST-inclusive and no other charge applies. Under these assumptions, Quote A is A$7.28 cheaper for that month.
Change the call mix, transfer rate, retention, or included minutes and the result changes. That sensitivity is the reason to use your own call records.
Add total cost of ownership
Normalised invoice cost is only the first layer. Total cost of ownership adds implementation, internal work, risk, and exit.
The Australian Government Department of Finance lists whole-of-life costs such as purchase, operations, licensing, added features, transition, and decommissioning. The framework is useful for private buyers too.
| TCO category | Costs to test |
|---|---|
| Discovery | Call analysis, workflow design, policy review, and acceptance criteria |
| Implementation | Configuration, integration, number porting, testing, and launch support |
| Operations | Subscription, usage, telephony, monitoring, support, and change requests |
| Human work | Escalation handling, quality review, corrections, and incident response |
| Data | Recording, storage, retrieval, deletion, backups, and overseas handling |
| Reliability | Fallback routing, outages, capacity, and recovery procedures |
| Exit | Data export, number release, integration removal, and replacement setup |
Calculate staff time using a consistent loaded hourly cost. Include wages, superannuation, and any other costs your finance method normally applies.
Avoid assumed savings from “replacing an employee.” Measure changed tasks, avoided external spend, recovered staff time, and completed caller outcomes instead.
Treat latency as a cost and quality variable
Voice latency is the accumulated delay across the phone network, audio transport, speech processing, model or tool work, and speech generation.
Location affects part of that chain. Twilio says geographic distance directly affects network latency. Its documentation describes noticeable silence when Australian workloads use the default US region.
Ask vendors for measured end-to-end performance on Australian calls. Request median and tail response times, test conditions, interruption behaviour, and transfer delay.
Do not accept a model inference figure as the full voice response time. It excludes other components and may exclude tool calls.
Test ordinary conversations and difficult ones. Include background noise, interruptions, long account lookups, upstream failures, and escalation to a person.
Track repeats, caller abandonment, mistaken interruptions, and transfers alongside latency. A low-cost call that creates staff recovery work can have a higher TCO.
Pharmacy buyers need a privacy cost line
The OAIC says a pharmacist is a health service provider. Privacy Act coverage can apply even when the organisation is a small business.
Include time for data mapping, privacy review, access controls, retention, incident planning, deletion testing, and contract work. Our pharmacy AI privacy guide covers the legal context.
Ask where audio, transcripts, logs, and backups are processed. Also ask about vendor access, subprocessors, retention, deletion, breach support, and data export.
Procurement questions worth sending in writing
- What exactly starts and stops each billable unit?
- Which currency applies, and what exchange-rate or indexation rule can change it?
- Are prices GST-inclusive or GST-exclusive?
- Which usage is included, and how do tiers and overages work?
- Does forwarding or transfer create another billed phone leg?
- How are short calls, silence, hold time, retries, and failed calls billed?
- Are model, speech, tool, recording, storage, and number charges included?
- How many concurrent calls are included, and what happens above that limit?
- What implementation, integration, support, and change fees apply?
- What end-to-end latency was measured on Australian phone calls?
- Where is every data type processed, stored, backed up, and accessed?
- What data export, deletion, number release, and exit assistance are included?
Ask for a sample invoice using your workload. Then reconcile the first production invoices against that model and investigate every new line item.
A simple decision rule
Compare three views: normalised monthly invoice, whole-of-life TCO, and cost per successful outcome. Keep service quality and risk measures beside all three.
A quote is comparison-ready only when units, currency, tax, inclusions, rounding, capacity, and contract terms are documented. If any field is missing, the price remains unverified.
Where Krepko and Kat fit
A Kat quote covers a pilot-stage administrative call-management scope for an Australian pharmacy. Within that scope, Kat owns calls, requests, consent and identity steps, context, ticketing, summaries, assignment and follow-up. Price phone, email, SMS and dispense-software connections only after separately agreeing compatibility, permissions and third-party dependencies. The pilot plan should identify its call scope, staff owner, escalation path and follow-up process.
For Krepko, request a tailored discounted quote based on selected modules, functions, integrations and pilot scope. Pharmacy staff handle clinical, complex, uncertain or human-requested calls and keep professional decision authority.
Frequently asked questions
- What does per-minute AI voice agent pricing usually measure?
- It can mean conversation time, telephony leg time, model session time, or rounded billing time, so the quote must define its unit.
- How should two AI voice agent quotes be compared?
- Apply both quotes to the same monthly workload, currency, tax treatment, rounding rule, transfer pattern, inclusions, and contract term.
- What costs sit outside the headline voice rate?
- Possible additions include phone numbers, inbound and outbound legs, transcription, speech generation, recording, integrations, support, implementation, monitoring, and exit work.
- Why does latency matter when comparing pricing?
- A cheaper architecture can still cost more operationally if delays increase repetition, abandonment, transfers, or staff recovery work.
- What should an Australian pharmacy ask about privacy?
- Ask where audio, transcripts, logs, and backups are processed, who can access them, how long they remain, and how deletion works.
Sources
- Programmable Voice Pricing in Australia · Twilio
- Twilio Regions · Twilio
- Speech-to-Text pricing · Google Cloud
- Text-to-Speech pricing · Google Cloud
- GPT-Live 1 Model · OpenAI
- Value for money · Australian Government Department of Finance
- What is a health service provider? · Office of the Australian Information Commissioner


