Why AI Is Specialising
The first wave of generative AI sold a simple idea: one very large model that can do everything. For writing and research that idea still works well. For running parts of a business it has limits, because a model trained on the public internet knows nothing about your prices, your hours or the job a caller is ringing about.
So the industry has started to specialise. Gartner predicts that by 2027 more than 50% of the generative AI models used by enterprises will be specific to their industry or business function, up from about 1% in 2023. In a press release on 9 April 2025 it also predicted that by 2027 organisations will use small, task-specific models at least three times more than general purpose large language models. The reasons it gave are the ones every business owner cares about: general models become less accurate when a task needs business context, and smaller specialist models answer faster and cost less to run.
Those are predictions about big organisations. But the same logic shows up plainly on a small business phone line, where a wrong answer or a two second pause is heard straight away. On the phone, "domain-specific" is not one feature. It is a stack.
The Five Layers of a Phone AI
Every AI phone agent starts with a general language model. What sits on top of it decides whether it sounds like a clever stranger or like someone who works for you. Here are the five layers, from the bottom up.
| Layer | What it adds | Where it comes from |
|---|---|---|
| 1. General model | Language: understanding requests and forming sensible sentences | A foundation model from a large AI company |
| 2. Voice | Hearing speech over noise and accents, speaking naturally, taking turns, coping with interruptions | Speech models tuned for phone audio |
| 3. Telephony skills | Transferring, holding, sending SMS, taking messages, knowing when a call has ended | The phone system itself |
| 4. Industry knowledge | How a clinic, trade business or agency usually handles calls, and the words callers use | Tuning and templates for that type of business |
| 5. Your business | Your hours, prices, service area, calendar, CRM and rules for urgent calls | You, connected through retrieval and integrations |
Layers two and three are what make an AI a phone specialist. Layers four and five are what make it your specialist. Gartner describes the two main ways businesses customise models as fine-tuning and retrieval augmented generation (RAG). Fine-tuning mostly shapes layers two to four: it teaches the model how a type of job is done. Retrieval powers layer five: the AI looks up your own information at the moment it answers, instead of relying on memory.
Symptoms of a Missing Layer
You do not need to see inside a product to know which layer is weak. Callers will tell you, through the mistakes they notice. This is the most practical way to use the stack: listen to a few calls and match the symptom to the layer.
The most common failure in 2026 is not layer one. General models are very good at language now. The weak points are usually the voice layer (long pauses and misheard names), telephony (an AI that cannot actually transfer or book, so every call ends with "someone will call you back") and the top layer, where the AI was never properly given the business's facts. For a closer look at what an AI can do beyond talking, see our guide what is an AI agent.
The Top Layer Is Yours
Here is the part most AI marketing skips. The top layer, your business, is the one that makes the biggest difference to callers, and it is the one only you can supply. No supplier knows that you do not go past Campbelltown, that Tuesday is theatre day, or that anyone who says "gas" goes straight to the owner's mobile.
Good specialist AI makes that layer easy to build and easy to change. Bad AI buries it in a setup project you cannot touch afterwards. Before you buy, write down what the top layer should contain:
Facts
Hours including public holidays, locations, service area, prices or price ranges, and what you do not do.
Systems
The calendar it can book into, the CRM or job system it should check and write to, and who gets the summaries.
Rules
What counts as urgent, who it transfers to and when, what it must never promise, and what it should say when it does not know.
"I don't know" is a feature
A well built top layer includes the edges of what the AI knows. An agent that says "I'll get someone to confirm that price for you" is doing its job. An agent that invents a price to sound helpful is a general model showing through the stack.
If you want the wider picture of how the phone system connects to the CRM, calendar and other tools in that top layer, our article on the phone system as the brain of your AI ecosystem covers the integrations.
Smaller Models, Faster Calls
Gartner's point about smaller, task-specific models matters more on the phone than almost anywhere. In a text chat, a three second wait is fine. On a call it feels like the line has dropped, and callers start saying "hello?" over the top of the answer.
Every step in the stack adds a little delay: turning speech into text, working out the answer, looking up your information, turning the answer back into speech, and moving audio between servers. A giant general model for every step is slower and dearer than a model sized for the job. The same goes for distance. If the audio travels overseas to be understood and back again, you hear it in the pauses.
The practical question for a vendor is not "which model do you use", but "how long does a caller wait, and where does the audio go". Ask them to show you on a live call from a mobile, not a recording.
Who Owns Each Layer
The five layers are rarely built by one company. A typical AI answering product might use a foundation model from one company, speech from another, run on a third party's cloud, and connect to your phone number through your phone provider, while a fourth company sells you the result. Each join adds delay, another privacy policy and another place for a fault to hide.
That is why the ownership question matters as much as the model question. Ask a vendor to name the company behind each layer, where each one processes and stores call audio and transcripts, and whether any of it is used to train models. You are looking for short, specific answers. We go deeper on this in how many companies sit between you and the model. Also note that from 10 December 2026, businesses covered by the Privacy Act must explain certain automated decisions in their privacy policy, which our automated decisions guide explains.
A Ten Minute Layer Test
You can test each layer yourself in about ten minutes of calls, as long as the vendor has loaded your real information first. Do it from a mobile, somewhere a little noisy.
- 1
Layer 2: talk over it
Say your name and a tricky street name, then interrupt it mid sentence. Does it stop, listen and get the name right? How long are the pauses?
- 2
Layer 3: ask it to do something
"Can you put me through to accounts?" and "Can you text me the address?" Does the transfer happen, and does the text arrive?
- 3
Layer 4: use your industry's words
Ask the way your customers ask. "Is it bulk billed?", "Can you do a rough quote?", "Do you do strata?" Does it understand?
- 4
Layer 5: ask what only you know
A price, a public holiday opening, a suburb you do not cover. Right answer, or an honest "let me check"?
- 5
The edges: push it
Say it is urgent. Ask for something you do not sell. Get cross. Does it follow your rules and hand over to a person?
Then read the transcript and summary. They should match the call and land where your team works. If you are comparing several providers, our ten call test turns this into a scored trial.
Where to Start
The shift to specialist AI sounds like a big company topic, but the first steps for a small business are simple and cost nothing.
| This week | This month | This quarter |
|---|---|---|
| Write your top layer down: facts, systems and rules, in one document | Pick one job for AI, such as after hours or overflow, and trial it on your real calls | Review transcripts, fix the gaps in your top layer, and decide whether to widen the job |
| List the ten questions your phone gets most, in customers' words | Run the layer test on two or three providers | Update your privacy policy for automated decisions before 10 December 2026 |
If you are not sure which calls to hand to AI first, our guide on which calls to automate sorts them by risk and reward.
How VOCPhone Builds the Stack
VOCPhone owns and operates its own network in Australia, and our AI agents are built into the phone system rather than bolted on through call forwarding. That means layers two and three, voice and telephony, live in the same place as your calls: the audio does not leave the country to be understood, transfers and texts are native to the platform, and there is no extra number in the chain. Our agents answer with a natural Australian accent, book appointments, qualify enquiries and route calls, and every call gets a summary in the app.
We do not claim to have built our own foundation model, and we do not think you should care who has. What we build with you is the top of the stack: your facts, your calendar and CRM, and your rules about who gets put through. Your call content is not used to train anybody's model, and when something needs fixing you ring one Australian team on 1300 663 222 rather than working out which of four companies owns the problem.
Frequently Asked Questions
What does domain-specific AI mean?
Domain-specific AI means an AI model or system built, trained or tuned for one industry or one type of task, rather than a general model used for everything. It can mean a model trained on a field such as finance or medicine, a model fine-tuned on examples of one job such as phone calls, or a general model connected to a business's own information so it answers from those facts. Most business AI products combine these. The goal is more accurate answers on the work that matters, with faster responses and lower running costs.
Is domain-specific AI replacing large language models?
Not replacing, building on. Most specialist systems still start with a general language model and add tuning, industry knowledge and business data on top. What is changing is how much work the giant general model does on its own. Gartner predicts that by 2027 more than half of the generative AI models used by enterprises will be industry or function specific, up from about 1% in 2023, and that small task-specific models will be used at least three times more than general purpose models, because general models lose accuracy where business context is needed.
What are the layers of an AI phone agent?
A useful way to think about it is five layers. At the bottom is a general language model that understands and writes language. Above it is a voice layer that hears speech over noise and accents and speaks naturally. Next are telephony skills such as transferring, sending SMS and taking messages. Then industry knowledge, meaning how a clinic, trade or agency usually handles calls. At the top is your own business: hours, prices, service area, calendar, CRM and rules for urgent calls. The bottom layers make it a phone specialist and the top layers make it your specialist.
How can I tell which layer of an AI receptionist is weak?
Listen for symptoms. Long pauses, misheard names and trouble with interruptions point to the voice layer. An AI that cannot transfer calls, send texts or book, so every call ends with a promised callback, has weak telephony skills. Not understanding your customers' usual phrasing points to missing industry knowledge. Wrong prices, hours or service areas, or confident guesses instead of an honest answer that it will check, mean your own business information was not properly loaded or connected. Testing each in turn with a few calls from a mobile shows where the problem is.
Why do small AI models matter for phone calls?
Because callers notice delay. Every step of an AI call, from turning speech into text to looking up your information and speaking the answer, adds time, and a pause of more than a second or two feels like a dropped line. Smaller models sized for the task respond faster and cost less to run than using a very large general model for every step. Distance matters too: if call audio is sent overseas to be processed, the round trip shows up as pauses. Ask any vendor how long callers wait and where the audio is processed.
What should a small business do about domain-specific AI now?
Start with the layer only you can supply. Write down your business facts, the systems an AI would need to use such as your calendar and CRM, and your rules for urgent calls and hand-overs. List the ten questions your phone gets most in your customers' own words. Pick one job for AI, such as after hours calls or overflow, and trial it on real calls with two or three providers, testing each layer. Review transcripts to fix gaps, and update your privacy policy for automated decisions before the Privacy Act change on 10 December 2026.
Does VOCPhone use domain-specific AI?
VOCPhone's AI agents are built into its phone system on its own Australian network, so the voice and telephony layers sit where your calls already are. Audio is not sent overseas to be understood, transfers and SMS are native to the platform, and the agents answer with a natural Australian accent, book appointments, qualify enquiries and route calls, with a summary for every call. VOCPhone does not claim to have built its own foundation model. The specialism comes from building the phone layers into the network and working with each business on its own facts, systems and rules. Call content is not used to train anybody's model.
