Callease AI

Callease sells a subscription against a cost its buyers already pay and cannot see: the enquiry that rang out. The commercial question is therefore not whether an AI agent can hold a conversation — several vendors have settled that — but whether the recovered call is worth more than the subscription, and whether one customer can be onboarded without an engineer. Everything below is an attempt to answer those two questions with figures that can be checked.

Calls answered live

37.8%

Of inbound calls to 85 small businesses across 58 industries. The remaining 62% reached voicemail or nothing at all.

Source · 411 Locals inbound-call study, as reported by Zadarma

Employing businesses in Australia

994,178

Of 2,729,648 actively trading businesses at 30 June 2025. The employing subset is the population that staffs a phone.

Source · Australian Bureau of Statistics, Counts of Australian Businesses

Forecast CAGR, AI voice agents

39.0%

2026 to 2033, on a market valued at USD 2.5bn in 2025.

Source · Grand View Research, AI Voice Agents Market Report

Reply latency target

< 500ms

The product's own design budget for the gap between a caller finishing and the agent starting.

Model · Design target set during the build, not a measured production average.

The market it sells into

The category Callease competes in is young enough that its size is still being argued about, and the forecasts differ by an order of magnitude depending on whether a research house counts voice agents as their own market or as a slice of conversational AI. Taking the narrowest credible definition — software whose product is an autonomous voice agent on a telephone line — Grand View Research values the global market at USD 2.5bn in 2025 and forecasts USD 35.2bn by 2033, a 39.0% compound annual growth rate, with North America holding 38.1% of 2025 revenue 1.

• $2.5bn
• $3.5bn
• $35.2bn

• 2025
• 2026
• 2033

Hollow columns are the published forecast, not a measurement.

Source · Grand View Research, AI Voice Agents Market Report

The Australian Bureau of Statistics counted 2,729,648 actively trading businesses at 30 June 2025, of which 994,178 were employing 2; 97.3% of all Australian businesses had fewer than twenty staff 3. Those two facts set the shape of the product. The addressable buyer is small, has no IT department, and the person who would configure an AI agent is the same person answering the phone — which is why "the customer configures it, not us" is a commercial requirement rather than a nicety.

All actively trading businesses

2.73m

ABS, 30 June 2025. Includes sole traders with no staff and no inbound call volume.

Employing businesses

994k

ABS. Someone is paid to be reachable, so an unanswered call has a cost attached to it.

Booking- or enquiry-led sectors

~300k (assumed 30%)

Health, trades, real estate, hospitality, professional services. Assumption, not a published figure.

Plausible near-term buyers

~30k (assumed 10%)

Willing to route a live line to software today. Assumption.

Model · ABS business counts, then two stated assumptions (30% sector fit, 10% near-term willingness). The first two rows are published data; the last two are ours and are labelled as such.

The buyer's problem, in numbers

The best-documented study of what actually happens to small-business phone calls tracked 85 businesses across 58 industries: 37.8% of inbound calls were answered by a person, 37.8% went to voicemail, and 24.3% received no response at all 4. The important part is the last figure. A voicemail is a weak recovery path; no response is none.

Answered by a person38%

The only outcome that can convert today.

Went to voicemail38%

Recoverable in principle, rarely recovered in practice.

No response at all24%

Not even a voicemail option. Unrecoverable without automation.

Source · 411 Locals inbound-call study, as reported by Zadarma

The alternative to software here is people, and people are the reason the problem persists. Median weekly earnings for a receptionist in Australia are AUD 1,175 5, or roughly AUD 61,100 a year before superannuation, penalty rates or cover for leave. Answering a phone across all 168 hours of a week rather than a 38-hour ordinary week takes about 4.4 of those people.

That number is why the pricing conversation is short. A subscription does not have to beat a receptionist on quality of conversation to be worth buying. It has to beat the marginal cost of the hours a receptionist is not there, which is where most of the missed calls are.

Where the revenue comes from

Three revenue mechanics sit inside the build, and they are not equally valuable. Ranking them is the difference between a product that grows and an agency that bills.

41%on making agents configurable rather than bespoke

Agent configuration system41%

Persona, knowledge base and business rules as structures a customer edits.

Voice orchestration and latency27%

Keeping audio, transcription, reasoning and speech in step.

CRM and calendar automation18%

The part that turns a conversation into a booked, recorded outcome.

Transcripts and analytics14%

Evidence for tuning scripts, and the first place a failure shows up.

Model · Our own apportionment of build effort across the four workstreams, reconstructed after delivery. Directional, and rounded.

The constraint the whole product is built around

Everything difficult about this system comes from one number. A caller will tolerate roughly half a second of silence before assuming the line is dead or talking over the reply. That budget has to contain the full round trip — audio in, transcription, model reasoning against the customer's knowledge base, speech synthesis, audio out — and any component that blows it takes the conversation with it.

1. L1

   Configuration surface

   Where a customer builds and tunes their own agent, and reads what it did. The commercial moat lives here, not in the voice.

   • Next.js
   • TypeScript
   • Tailwind CSS
   • Call analytics
   • Agent setup
2. L2

   Orchestration

   Coordinates the live call: telephony events in, reasoning out, integrations fired without blocking the conversation.

   • Node.js handlers
   • Voice webhooks
   • Synthflow call flows
3. L3

   Transient state

   Holds what a call needs while it is happening, and makes automation jobs safe to retry when a downstream system is slow.

   • Redis
   • Session state
   • Retry-safe jobs
4. L4

   System of record

   Everything that must survive the call: transcripts, logs, per-customer configuration and agent settings.

   • MongoDB
   • Transcripts
   • Call logs
   • Customer config
5. L5

   Business systems

   Where a successful call becomes a commercial fact — a calendar entry, a CRM record, a follow-up.

   • CRM
   • Calendar
   • Payments

1. 01Call lands on a local Australian number

   Numbers are provisioned per customer. An unfamiliar interstate caller ID is answered far less often, so this is a conversion decision disguised as telephony configuration.

   ▾

2. 02Session opens, agent context loaded

   Persona, knowledge base and business rules for that customer are pulled into the session. Held in Redis, because none of it needs to outlive the call.

   ▾

3. 03Speech to text, then grounded reasoning

   The model answers against that customer's knowledge base rather than from general knowledge. Grounding is what keeps it from inventing a price.

   ▾

4. 04Can the agent close this itself?decision

   Intent is matched against what the agent is allowed to do — book, quote, qualify, or none of the above.

   • ↳ Out of scope → hand to a human
   • ↳ Caller asks for a person → transfer
   • ↳ Low confidence → capture details and escalate
   ▾

5. 05Live availability checked, meeting booked

   Writing into a real calendar rather than a queue. A booking that a human has to re-key is not an automated booking.

   ▾

6. 06CRM record written, follow-up triggered

   Fired through retry-safe jobs, so a slow CRM delays the record and never the conversation.

   ▾

7. 07Transcript stored and scored

   The call becomes evidence: the script is tuned on it, and a flow that has started failing is visible the same day rather than at month end.

What we would watch

An honest analysis names the ways the model breaks. These are the three that matter for this one, and none of them are engineering problems.

References

1. 1.AI Voice Agents Market Size And Share Report, 2026–2033 · Grand View Research
2. 2.Counts of Australian Businesses, including Entries and Exits · Australian Bureau of Statistics
3. 3.Number of small businesses in Australia · Australian Small Business and Family Enterprise Ombudsman
4. 4.The Hidden Cost of Missed Calls (411 Locals inbound-call study) · Zadarma
5. 5.Receptionist — median weekly earnings · Victorian Government careers data