AI Agents on the Phone: What Agentic AI Really Means

Somewhere around the start of 2026 the industry stopped saying chatbot and started saying agent, and the phone vendors were not far behind. Ask three providers what their AI agent does and you will get three answers that sound identical and mean completely different things. One means a recorded menu with speech recognition bolted on. One means a system that reads your website and answers questions about your opening hours. One means something that takes a booking, moves it, cancels it, and writes the result into the software your staff actually use. All three are called an agent, and only the third one takes work off anybody. The distinction matters more than the marketing does, because the first two get judged by your customers as a phone menu with delusions of grandeur, and the third gets judged as service. There is a single question that separates them, it takes one test call to answer, and almost nobody asks it: at the end of the call, did anything change in a system of record, or did the caller simply have a pleasant conversation with a computer and then have to ring back during business hours? This is what an AI agent is, what agentic actually means when it is not doing marketing work, which calls suit one and which will embarrass you, how the handover to a human has to be built, what the Australian rules require, and how to roll one out over six weeks without discovering the problems in front of your best customer.

AI Agents · Explainer · 2026

Everyone Sells an AI Agent Now. Very Few Are Selling One.

Agentic AI became the term of 2026, and within about a fortnight every phone vendor in Australia had one. Most of what is being sold is a menu with a better voice: it listens well, it talks well, and at the end of the call nothing in your business has changed. A real agent finishes the job. This is the difference, in plain terms, with the test that tells them apart in one call, the categories of call that genuinely suit one, and the six week way to put one in without frightening your customers.

📅 ⏱ 17 min read 🇦🇺 Australian owned, Australian hosted, Australian supported
TL;DR

An AI agent is not a better menu. The test is whether it writes, not whether it talks. Speech recognition, a natural voice and an accurate answer about your opening hours are all worth having, and none of them make something an agent. The line is crossed when the system can complete a task in a system of record: book the job, move the appointment, raise the ticket, log the note, send the confirmation. Everything above that line takes work off your team. Everything below it moves the work to a different point in the day. Ask one question of any vendor: show me a call where the outcome appears in the CRM or the calendar, without a person touching it. Automate calls that are repetitive, structured, high volume and low emotion: hours and location, booking and rescheduling, order and job status, simple triage, after hours capture, overflow at peak. Do not automate complaints, distressed callers, anything with a safety dimension, or the first call from a large new customer. Design the handover before the greeting, because the escalation path is what customers actually judge. And measure resolution, not containment. A call that never reached a human and never solved anything is not a saving.

Australian specifics: recording notification, a privacy policy that discloses automated handling, a genuine path to a person, and an emergency call that never enters an agent.

What an AI Agent Actually Is

Strip the marketing off and the definition is short. An AI agent is software that takes a goal, works out the steps, carries them out using tools it has been given, and reports what it did. On a phone call the goal arrives as a sentence from a customer, the tools are your calendar, your CRM, your job management system and your messaging, and the report is both a spoken confirmation to the caller and a record left behind for your team.

The word doing the work in that definition is tools. A system with no tools can only talk. It can talk extremely well, understand a broad Australian accent, handle an interruption gracefully, and give a genuinely correct answer about whether you open on the Saturday of a long weekend. That is useful. It is not an agent, and the difference shows up in your week rather than in the transcript.

The plain version

An assistant answers questions. An agent finishes jobs. If the caller still has to ring back, fill in a form, or wait for somebody to listen to a recording before anything happens, you bought an assistant. That may be exactly what you need and it is considerably cheaper, but it will not reduce anybody's workload, and you should not be paying agent prices for it.

Agentic is the adjective attached to the same idea. It has been stretched to cover almost anything with a language model behind it, which is why it has stopped meaning much in a sales conversation. Used properly it describes a system with three properties: it decides the sequence of steps itself rather than following a fixed script, it can use more than one tool in a single interaction, and it can recognise when it has reached the edge of what it should do and hand over. Take away the third property and you do not have an agent, you have a liability with a pleasant voice.

The Five Rungs, and Where the Line Sits

Everything currently sold as AI phone handling in Australia sits on one of five rungs. The rungs are cumulative and each is a legitimate product. The confusion is entirely about which one you are being quoted for.

RungWhat it doesWhat the caller experiencesWork it removes
1. Touch tone menuPress 1 for sales. Fixed tree, no understanding.Listening to options, guessing, pressing 0.Routing only. Adds caller effort.
2. Speech menuSame tree, spoken. "Say sales or accounts."Slightly faster. Still a tree.Routing, marginally better.
3. Scripted assistantNatural voice, fixed conversation flows written in advance.Sounds modern until the question is off script.The most common two or three questions.
4. Retrieval assistantAnswers from your real content: hours, services, pricing, policies.Genuinely helpful for questions. Cannot do anything.Answering repeat questions. Message taking.
5. Task completing agentAll of the above, plus it acts: books, reschedules, cancels, raises, updates, sends.The call ends and the thing is done.The work itself.

The line sits between four and five, and it is not a matter of degree. Rungs one to four all read from your business. Rung five writes to it. Reading is comparatively easy and carries little risk. Writing means the system can change a booking, create an obligation, and be wrong in a way that costs somebody a morning, which is why it demands integration, permissions, confirmation and audit that the other four do not need. When a vendor demonstration shows only questions being answered beautifully, you have watched a rung four product. Ask to see a write.

The One Question That Settles It

You do not need to understand model architectures to buy this well. You need one test call and one question.

Ask
Show me a call where the outcome appears in the calendar or the CRM, and nobody types it in afterwards.
Watch
The screen, not the transcript. The transcript always looks impressive. The record either exists or it does not.
Then ask
What happens when it gets that wrong, who finds out, and how fast can it be undone.

Three answers are common and only one is a rung five product. "It sends an email to your team who then enter it" is rung four with extra steps. "It creates a task for someone to action" is rung four with a queue attached. "It appears in the calendar, with the customer's details, and the customer has the confirmation" is an agent. The distinction is not pedantry: the first two leave the same amount of work in your business and simply move it to a quieter part of the day, which has some value, but it is not the value being priced.

A demonstration on the vendor's data proves very little

Every one of these systems performs beautifully on a fictional business with twelve services and no exceptions. Yours has a service that is only offered on Tuesdays, a technician who does not do the northern suburbs, a price that depends on whether the customer is on a maintenance agreement, and a rule about deposits that everybody knows and nobody has written down. Insist the trial runs on your real data, with your real exceptions, on your real calendar. The interesting failures only appear there, and they appear in the first afternoon.

What Happens on a Real Call

It helps to see the call broken into its phases, because each one is a place where a system either holds up or quietly falls apart, and because it tells you what to listen for when you test.

PhaseWhat the agent doesWhere it goes wrong
GreetingIdentifies the business, states plainly that it is an automated assistant, invites the caller to just say what they need.Pretending to be a person. Never do this. It is the fastest way to lose the caller's trust for the rest of the call.
IntentWorks out what the caller actually wants from an ordinary sentence, including when they give three things at once.Forcing the caller to rephrase into keywords. If it says "I did not understand that" twice, the design has failed.
IdentityFinds the existing customer from the calling number, or asks for the minimum needed to proceed.Asking for details it could have looked up, or worse, accepting an identity claim before doing anything sensitive.
RetrievalPulls the facts it needs: the booking, the job status, the balance, the next available slot.Answering from general knowledge instead of your records. This is the failure that produces confident nonsense.
ActionDoes the thing. Books, moves, cancels, raises, updates, orders the callback.Acting without reading back. Every write should be confirmed in the caller's own terms before it commits.
ConfirmationStates what was done, and sends it in writing by SMS or email while the caller is still on the line.Verbal only. If the caller has nothing in their hand, they will ring back to check, and you have gained nothing.
EscalationRecognises the limit and moves the call to a person, carrying the context with it.Escalating with nothing attached, so the customer starts again from the beginning. See the handover section.
RecordWrites the transcript, the summary, the outcome and the follow up into the CRM against the right contact.Writing it somewhere nobody looks. A record in a separate portal is a record that does not exist.

Notice how much of that list is not about the artificial intelligence at all. Identity, integration, confirmation and record keeping are ordinary systems work, and they are where deployments actually succeed or fail. The conversational part has been more or less solved for two years. The plumbing is what you are buying.

Which Calls Suit an Agent, and Which Do Not

The selection criteria are consistent across every business we have put one into. Good candidates are repetitive, structured, high in volume, and low in emotion. Bad candidates fail on the last one, and emotion is the criterion people underweight.

Automate first

Opening hours, address and parking. Booking, rescheduling and cancelling. Order, job or delivery status. Simple triage that sorts a caller to the right queue. After hours capture with a real outcome. Overflow when everyone is already on a call.

🤔

Automate carefully

Quoting anything variable. Taking payment. Account changes. Anything where being wrong creates an obligation. These work, and they need tighter confirmation, narrower permissions and a lower threshold for handing over.

🚫

Do not automate

Complaints. Distressed or vulnerable callers. Anything with a safety or medical dimension. Cancellations you would fight to keep. The first call from a large prospective customer. In each case the caller is judging whether you take them seriously.

The volume argument is the real one

Most Australian small businesses do not have a call quality problem. They have a call capacity problem: the phone rings while everybody is already on a job, and a share of those callers simply ring somebody else and are never counted. The value of an agent is rarely that it handles a call better than your best person. It is that it handles the fourteenth simultaneous call at 8:40am, which your best person cannot, because they are on the first one. That is also why the metric that matters is not how well it performs on the calls it takes, but how many enquiries stop disappearing.

The Handover Is the Product

Customers do not judge these systems on the calls that go perfectly. They judge them on the moment it becomes clear the machine cannot help, because that is the moment they find out what your business thinks of them. Get the handover right and a caller will happily use the agent again. Get it wrong once and they will start every future call by saying "operator" over the top of the greeting.

RuleWhy
A person is always reachable, and saying so is not a failureEvery agent should treat "let me talk to someone" as a valid first sentence and act on it immediately, without a negotiation. Hiding the exit is the single most resented pattern in phone automation, and it long predates AI.
Context travels with the callThe person who picks up sees who is calling, what was asked, what the agent already did, and what it could not do. A handover that makes the customer repeat themselves is worse than no automation, because now they have explained it twice.
Escalate on frustration, not just on failureRepetition, interruption, raised voice, a second attempt at the same request. These are signals to stop trying and pass it on, even when the agent believes it is capable.
Escalate on subject, alwaysKeep a list of subjects that never get handled automatically, regardless of confidence: complaints, legal, safety, anything involving a vulnerable person. This is a fixed rule, not a judgement call for the model.
Have an answer for nobody availableAfter hours, or when the queue is genuinely empty, escalation cannot mean an unanswered ring. It means a booked callback with a stated time, and it means keeping that time.

Guardrails, and Saying "I Do Not Know"

The failure mode people fear is the system inventing something. It is a real risk and it is largely a design problem rather than a model problem. Three controls remove most of it.

Answer from your material, not from the world. The agent should be restricted to your published content and your records for anything factual about your business. If the answer is not in there, the correct behaviour is to say so and offer to have somebody call back, which every reasonable customer accepts. General knowledge has no place in an answer about your pricing.

Write with permission, and narrowly. Booking into a defined calendar, within defined hours, for defined service types is a safe write. Amending an invoice is not, at least not on day one. Give the agent the smallest set of actions that covers the calls you chose, and expand it after you have seen a month of real behaviour.

Confirm before committing. Read the action back in the caller's own words and take an explicit yes before it commits. This single step converts most misunderstandings into a correction on the call rather than a problem discovered on the day of the job.

Prompt injection is now a phone problem too

An agent that reads incoming messages, emails, forms or documents as part of its work can be given instructions by whoever wrote that content. It is the same class of problem as an email link, and the answer is the same: treat anything that arrives from outside as data, never as instructions, keep the actions available to the agent narrow, and require confirmation for anything that moves money or changes an account. Ask your provider directly how they separate the two. A vendor who has not thought about the question is telling you something useful.

What It Has to Know, and Where That Lives

An agent is only as good as what it can see. This is the part of the project that takes real time, and the part that gets underestimated in every quote, including ours if we are not careful.

What it needsWhere it usually comes fromThe work involved
Services, prices, inclusions, exclusionsYour website, plus the things everyone knows and nobody wrote downWriting down the unwritten rules. This is the genuinely hard part and it is worth doing anyway.
Hours, holidays, locations, coverage areaYour phone system and your own calendarAn hour, and then a habit of keeping it current.
Availability and booking rulesCalendar or job management systemConnecting it, and deciding what the agent may and may not book.
Customer records and historyCRMMatching on calling number, and deciding what may be read out to a caller who has not been verified.
Job or order statusJob management, ecommerce platform, or a spreadsheetUsually an integration. Sometimes a decision to stop using the spreadsheet.
Escalation mapYour own head, at presentWriting down who handles what, and when. Useful even if the project stops here.

Businesses that already have their systems connected get an agent working in days. Businesses with a CRM nobody updates, a calendar that lives in one person's phone and pricing that depends on who answers will spend most of the project fixing that, and will get most of the benefit from having fixed it. Our note on phone system integrations and open APIs covers the connection side, and the Australian platforms we see most often covers what typically already fits.

Measuring It Honestly

Containment rate, the share of calls that never reached a human, is the number every vendor reports and the number that most easily flatters a bad deployment. A caller who gave up is contained. A caller who was told to ring back tomorrow is contained. Measure it, and never measure it alone.

MeasureWhat it tells youWatch for
Resolution rateCalls where the caller's need was actually met, verified against the record that was written.This is the real number. If your provider cannot produce it, ask why.
Escalation qualityWhat share of handovers arrived with usable context, and how long the human then spent.A handover that takes longer than the original call would have is a net loss.
Repeat contact within 48 hoursWhether the call actually finished, from the customer's point of view.Rising repeats with rising containment is the classic false positive.
Abandonment inside the agentPeople hanging up mid conversation. Almost always a design fault, not a model fault.Cluster them by intent. The pattern is usually one flow.
Answered enquiries, totalThe count that matters commercially: enquiries that reached your business at all.Compare to the same month last year, not to last week.
Time to first humanFor the calls that should reach a person, whether automation made that faster or slower.If it went up, your triage is too eager.

Read a sample of transcripts every week for the first two months. Twenty calls, chosen at random rather than by the system's own confidence score. It takes half an hour and it will teach you more about your customers than the dashboard does. Our guide to the contact centre metrics that actually matter sets out the wider measurement picture.

The Australian Rules That Apply

None of this is exotic, and all of it is easier to build in than to retrofit.

ObligationWhat it means in practice
Tell people it is automatedNot currently a specific telecommunications rule, and it is the right thing to do and increasingly expected. A short, plain statement in the greeting. Pretending to be human is a trust problem that outlives the call.
Recording and monitoring notificationRecording law in Australia is state based and inconsistent. One organisation wide standard of clear notification at the start of every call, including the automated path, is simpler and errs in the right direction. See our guide to call recording law and setup.
Privacy and disclosureReforms to the Privacy Act require organisations to disclose in their privacy policy where automated systems make or substantially influence decisions affecting individuals. If your agent decides priority, eligibility or access, that belongs in the policy. Our note on automated decisions and the Privacy Act covers the detail.
Where the data goesKnow which country the transcripts and recordings are stored in, and who else can see them. For health, legal, government and NDIS work this is usually a hard requirement rather than a preference.
Emergency callsA call to 000 must never enter an automated flow, on any device, in any circumstance. Check this rather than assume it. Our note on Triple Zero and cloud phone systems explains the surrounding rules.
Outbound and the Do Not Call RegisterAn AI agent making outbound calls is subject to exactly the same rules as a person making them. Washing lists, consent, calling hours and identification all apply unchanged. See the outbound calling rules.
Voice cloning and impersonationSynthetic voice is now also an attack tool. Staff should know that a familiar voice on the phone is not identification, and approvals should not depend on recognising one. Our note on voice cloning and vishing covers the defence.

What It Costs, and What It Replaces

Pricing in this market is unsettled and comes in three shapes: per minute, per resolved interaction, or included in a platform subscription with fair use. Per minute is predictable and rewards brevity, which is not always what you want. Per resolution aligns the incentives well and requires you to agree what resolution means, in writing, before you sign. Included pricing is simplest and is worth checking against your real call volumes rather than your average ones.

Compare it to the right thing

The comparison people reach for is a receptionist's salary, and it is the wrong one, because most small businesses were never going to hire a receptionist. The honest comparison is with the calls you are currently losing. If you take forty enquiries a week and miss six of them, and one in three of those would have converted at an average job value you can name, the arithmetic takes about two minutes and it usually settles the question without any reference to headcount. Our AI voice agent cost and ROI guide works through it properly, and the cost calculator covers the surrounding system.

Two costs are routinely left out of quotes. The first is the setup work described in the data section, which is real and is usually measured in days rather than hours. The second is ongoing curation: somebody has to read transcripts, notice the three questions it keeps getting wrong, and fix them. Budget an hour a week for the first two months and an hour a month after that. Deployments that fail almost always failed here, quietly, about six weeks in.

A Six Week Rollout

The order matters more than the speed. Every step exists because skipping it produces a specific, predictable failure.

WeekWhat happensWhy this order
1Listen. Pull a month of call records, and sort the last two hundred calls into intents by hand. Count them.Everybody thinks they know what customers ring about. Almost everybody is wrong about the proportions, and the proportions decide what you automate.
2Choose two intents, no more. Write down the exceptions, including the ones that live in people's heads.Two intents can be made genuinely good. Eight will all be mediocre, and mediocre is what customers remember.
3Connect the systems and build the escalation path first, before the conversation.If the handover works, a rough agent is still safe to put in front of customers. The reverse is not true.
4Run it after hours only. Read every transcript.After hours is the safest possible trial: the alternative was a voicemail nobody returns, so the floor is low and the learning is real.
5Add overflow during business hours, when everybody is already on a call.Second safest. Again the alternative is an unanswered ring, not a person.
6Review the numbers from the measurement section. Then either widen to a third intent or fix what is not working. Not both.The habit of reviewing before widening is the difference between a system that improves and one that quietly decays.

Tell your team what is happening before week four, not after. Phone automation lands badly when it arrives unannounced, and staff who understand that it is taking the fourteenth simultaneous call rather than their job will help you fix it. They are also the people who know the unwritten rules from week two.

How We Build It

Uniden Voice runs the AI on the same Australian platform that carries the calls, which matters for three practical reasons. Latency is short because the audio is not making a round trip offshore, and on a voice call a delay of half a second is the difference between a conversation and an interrogation. The transcripts, recordings and customer records stay on Australian infrastructure we operate. And when something needs fixing, the phone system, the AI and the integration are one product with one support number rather than three vendors pointing at each other.

We build to the rungs described above and we will tell you plainly which one you need. Plenty of businesses are best served at rung four, and we would rather sell that than have you pay for actions you never turn on. When you do want rung five, the setup work is done with you rather than to you, because the unwritten rules only exist in your head and getting them out is most of the job.

See it answer one of your calls

Bring us the two questions your phone gets most, and your calendar. We will show you the same call at rung four and rung five, on your own data, and you can decide which one is worth paying for.

Get Started Or call 1300 881 662

Frequently Asked Questions

What is an AI agent, in plain English?
An AI agent is software that takes a goal, works out the steps itself, carries them out using tools it has been given, and reports what it did. On a business phone call the goal arrives as an ordinary sentence from a customer, the tools are your calendar, CRM, job management system and messaging, and the report is both a spoken confirmation to the caller and a record left behind for your team. The word doing the real work in that definition is tools. A system with no tools can only talk, and it can talk extremely well: understand a broad Australian accent, handle interruptions, and give a correct answer about whether you open on the Saturday of a long weekend. That is genuinely useful and it is not an agent. The distinction that matters commercially is simple. An assistant answers questions. An agent finishes jobs. If the caller still has to ring back, fill in a form, or wait for somebody to listen to a recording before anything happens in your business, what you have is an assistant. That may be exactly what you need, it costs considerably less, and you should not be paying agent prices for it.
What is the difference between agentic AI and a chatbot or an IVR?
Think of five rungs. A touch tone menu is rung one: press 1 for sales, a fixed tree with no understanding. A speech menu is rung two: the same tree, spoken. A scripted assistant is rung three: a natural voice following conversation flows written in advance, which sounds modern right up until the caller asks something off script. A retrieval assistant is rung four: it answers from your real content, so hours, services, pricing and policies are all correct, but it cannot do anything. A task completing agent is rung five: it does all of that and then acts, booking, rescheduling, cancelling, raising a ticket, updating a record and sending a confirmation. The line sits between four and five and it is not a matter of degree. Rungs one to four read from your business. Rung five writes to it. Writing means the system can change a booking, create an obligation and be wrong in a way that costs somebody a morning, which is why it requires integration, permissions, confirmation and audit that the lower rungs do not need. Agentic, used properly, describes rung five: it decides the sequence of steps itself rather than following a fixed script, it can use more than one tool in a single call, and it knows when to hand over.
How do I tell whether a vendor is really selling an AI agent?
One test call and one question. Ask them to show you a call where the outcome appears in the calendar or the CRM without a person typing it in afterwards, and watch the screen rather than the transcript. The transcript always looks impressive. The record either exists or it does not. Three answers are common. If it sends an email to your team who then enter it, that is a retrieval assistant with extra steps. If it creates a task for somebody to action, that is a retrieval assistant with a queue attached. If the booking appears in the calendar with the customer's details and the customer already has the confirmation, that is an agent. The first two leave the same amount of work in your business and simply move it to a quieter part of the day, which has some value but is not the value being priced. Then ask the follow up question, which is what happens when it gets that wrong, who finds out, and how quickly it can be undone. And insist the trial runs on your real data with your real exceptions, because every one of these systems performs beautifully on a fictional business with twelve services and no exceptions, and the interesting failures only appear on yours.
Which calls should an AI agent handle, and which should it never handle?
Good candidates are repetitive, structured, high in volume and low in emotion, and the last criterion is the one people underweight. Automate first: opening hours, address and parking, booking and rescheduling and cancelling, order or job or delivery status, simple triage that sorts a caller to the right queue, after hours capture with a real outcome rather than a voicemail, and overflow when everybody is already on a call. Automate carefully, with tighter confirmation and narrower permissions: anything variable that has to be quoted, taking payment, account changes, and anything where being wrong creates an obligation. Do not automate complaints, distressed or vulnerable callers, anything with a safety or medical dimension, cancellations you would fight to keep, or the first call from a large prospective customer. In each of those the caller is judging whether your business takes them seriously, and an automated answer settles that question the wrong way. Worth naming the real argument for automation: most Australian small businesses do not have a call quality problem, they have a call capacity problem. The value is rarely that the agent handles a call better than your best person. It is that it handles the fourteenth simultaneous call at 8:40am, which your best person cannot, because they are on the first one.
Will an AI agent make things up about my business?
It can, and it is largely a design problem rather than a model problem. Three controls remove most of the risk. First, answer from your material rather than from the world: restrict the agent to your published content and your records for anything factual about your business, and make the correct behaviour for an unknown question be to say so plainly and offer a callback. General knowledge has no place in an answer about your pricing. Second, write with permission and narrowly: booking into a defined calendar within defined hours for defined service types is a safe write, amending an invoice is not, at least not on day one. Give it the smallest set of actions that covers the calls you chose and widen after a month of real behaviour. Third, confirm before committing: read the action back in the caller's own words and take an explicit yes, which turns most misunderstandings into a correction on the call rather than a problem discovered on the day of the job. One further risk worth naming is prompt injection. An agent that reads incoming messages, emails or documents can be given instructions by whoever wrote them, so anything arriving from outside must be treated as data and never as instructions.
How should the handover to a human work?
Customers do not judge these systems on the calls that go perfectly. They judge them on the moment it becomes clear the machine cannot help, because that is when they find out what your business thinks of them. Five rules. A person is always reachable, and every agent should treat let me talk to someone as a valid first sentence and act on it immediately without a negotiation, because hiding the exit is the most resented pattern in phone automation and long predates AI. Context travels with the call, so the person who picks up sees who is calling, what was asked, what the agent already did and what it could not do; a handover that makes the customer repeat themselves is worse than no automation because now they have explained it twice. Escalate on frustration rather than only on failure, treating repetition, interruption, a raised voice or a second attempt at the same request as signals to stop. Escalate on subject always, keeping a fixed list of things that are never handled automatically regardless of confidence: complaints, legal, safety, anything involving a vulnerable person. And have an answer for when nobody is available, which after hours means a booked callback with a stated time, and keeping that time.
What does an AI phone agent cost in Australia?
Pricing is unsettled and comes in three shapes. Per minute is predictable and rewards brevity, which is not always what you want from a conversation. Per resolved interaction aligns incentives well and requires you to agree in writing what resolution means before you sign. Included in a platform subscription with fair use is simplest, and is worth checking against your real call volumes rather than your average ones. More important than the shape is the comparison you make. People reach for a receptionist's salary, and that is usually the wrong benchmark because most small businesses were never going to hire a receptionist. The honest comparison is with the calls you are currently losing: if you take forty enquiries a week and miss six, and one in three of those would have converted at an average job value you can name, the arithmetic takes two minutes and usually settles the question without reference to headcount. Two costs are routinely left out of quotes. Setup work, which means writing down the rules that currently live in people's heads and connecting the systems, measured in days rather than hours. And ongoing curation, because somebody has to read transcripts and fix the three questions it keeps getting wrong. Budget an hour a week for two months, then an hour a month.
How do I measure whether the AI agent is actually working?
Do not rely on containment rate, the share of calls that never reached a human. It is the number every vendor reports and the one that most easily flatters a bad deployment, because a caller who gave up is contained and a caller told to ring back tomorrow is contained. Measure it alongside five others. Resolution rate, meaning calls where the need was actually met, verified against the record that was written; this is the real number and a provider who cannot produce it is worth questioning. Escalation quality, meaning what share of handovers arrived with usable context and how long the human then spent, since a handover that takes longer than the original call is a net loss. Repeat contact within 48 hours, which tells you whether the call finished from the customer's point of view; rising repeats alongside rising containment is the classic false positive. Abandonment inside the agent, which is almost always a design fault rather than a model fault and usually clusters in one flow. Total answered enquiries, compared to the same month last year. And time to first human for the calls that should reach a person. Beyond the dashboard, read twenty randomly chosen transcripts every week for the first two months.

What to Read Next

Your next reads

Uniden Voice Over Cloud logo

Australia’s smartest AI-powered cloud phone system — Australian owned, Australian hosted, Australian supported. unidenvoice.com | 1300 881 662