How to Replace Your IVR With an AI Voice Agent
A migration plan for moving off phone trees — what to keep, what to cut, how to run both in parallel, and the mistakes that send callers back to zero.
On this page
“Press 1 for sales, press 2 for support” has survived for thirty years not because it works but because there was no alternative. There is now, and the migration is far less frightening than most teams expect — provided you do it in the right order.
This is the plan we give customers moving off phone trees: what to find out first, what to keep, how to run both systems side by side without risking your main line, and the mistakes that undo the whole thing.
Why phone menus fail
The core problem is that an IVR asks the caller to translate their problem into your internal org chart before you will help them. Callers do not know whether a billing question about a delayed order is “billing” or “orders”. So they guess, guess wrong, and get transferred — or they press zero repeatedly until they reach a human, which defeats the point of the menu entirely.
| What the menu does | What the caller experiences |
|---|---|
| Lists six options | Forgets option two by the time option five is read |
| Asks for an account number by keypad | Does not have it to hand, and cannot ask a question about it |
| Routes on a single choice | Their problem does not fit one category |
| Handles one caller per line | Hold music |
| Cannot answer anything | A menu, then a queue, then a person — for a question that took one sentence |
If a large share of your callers press zero — or say “operator”, or mash the keypad — your menu is a toll booth rather than a router. That single number is the strongest argument for the project, and you can usually get it from your existing phone system in an afternoon.
What actually changes
| IVR | AI voice agent | |
|---|---|---|
| How the caller states their need | Chooses from a list you wrote | Says it in their own words |
| Wrong-category calls | Common, and require a transfer | Largely disappear — there are no categories |
| Can it answer anything? | No. It routes. | Yes, if you gave it the information |
| Can it do anything? | Occasionally, via keypad entry | Books, looks up, updates, transfers |
| Concurrency | Limited by lines | Effectively unlimited |
| Changing it | Re-record prompts, re-map the tree | Edit a paragraph of plain English |
| Out of hours | Voicemail | The same service as during the day |
That second-to-last row is quietly the biggest operational change. Adjusting an IVR is a project; adjusting an agent is editing a sentence and taking a test call. Teams that migrate end up improving their phone experience continuously rather than annually, simply because it stopped being expensive to do.
Step 1: audit what your IVR really does
Do not start by rebuilding your menu tree. Start by finding out what callers actually want, which is almost never the same shape as the menu.
- Pull call volumes per menu option for the last three months. You will usually find two options carry most of the traffic and three exist for organisational reasons.
- Get the zero-press rate — callers escaping to a human.
- Sit with whoever answers the phone for two hours and write down every question. This is the highest-value hour of the whole project.
- List the top twenty questions by frequency. This becomes your agent’s knowledge, and it is usually shorter than people expect.
- Mark which ones need a real system lookup — a calendar, an order status, an account balance. Those decide which integrations you need.
Most teams discover that 70–80% of calls are four or five recurring questions, none of which the IVR answers. The menu was never routing to answers — it was routing to people who answer the same five things all day.
Step 2: decide what to keep
Not everything should move. Three categories, and being honest about the third saves a lot of trouble.
| Category | Examples | Decision |
|---|---|---|
| The agent should handle it | Opening hours, prices, booking, rescheduling, order status, balance enquiries, directions | Move. This is the deflection that pays for the project. |
| The agent should collect and pass on | Complaints, custom quotes, anything needing judgement | Let the agent gather the details, then transfer or take a message. Half the work still gets done. |
| Keep it as-is | Emergency lines, regulated advice, anything with a legal script, safeguarding routes | Route these straight to a human, immediately, with no attempt to help first. |
If you serve a market where a caller might be in distress — health, utilities, finance, anything involving arrears — the rule to escalate instantly must be in the agent’s instructions before it takes a single real call. “If the caller is distressed, in pain, or describes an emergency, transfer immediately without asking further questions.”
Step 3: build the agent
The build is now the short part. What matters is that the twenty questions from your audit are answerable and that escalation works.
- Instructions — the twenty questions become the agent’s knowledge and rules. Structure and worked examples are in the prompt guide.
- Knowledge base — anything longer than a paragraph of facts. Your price list and policy documents go here, not in the prompt: knowledge bases.
- Integrations — whatever the lookups from your audit need. A calendar turns “take a message about booking” into “book it”: integrations.
- Transfer tool — non-negotiable, pointed at a number somebody answers: tools.
End to end, this is the AI receptionist playbook, and a first working agent is about ten minutes from the quickstart.
The greeting is not the menu
The instinct is to replace “press 1 for sales” with “you can ask me about sales, support, billing or opening hours”. Do not. That is a menu read aloud, with all the same problems.
Brightsmile Dental, this is Amara — how can I help?
Open-ended, short, and it lets the caller state their actual problem. If disclosure is required in your jurisdiction, it goes here too: “…this is Amara, Brightsmile’s virtual assistant — how can I help?”
Step 4: run both in parallel
This is the step that de-risks everything, and the one people skip because they are impatient. Nothing customer-facing changes yet.
| Approach | What happens | Risk |
|---|---|---|
| After-hours only | The IVR runs as normal by day; the agent takes evenings and weekends. | Almost none. Those calls currently reach voicemail, so anything is an improvement. |
| Overflow | The agent picks up when the queue is full or nobody answers. | Low. It only handles calls that were being abandoned. |
| One menu option | A single low-risk route — “opening hours and directions” — goes to the agent. | Low and contained. |
| A parallel number | A second number runs the agent while the main line is untouched. | Zero. Best for staff testing before customers ever reach it. |
After-hours is the strongest starting point for most businesses. Those calls are being lost today, so the comparison is against voicemail rather than against your team — and within a fortnight you have real transcripts from real customers to tune against.
Not a sample — all of them, if volume allows. This is where you find the questions your audit missed, and the two-week transcript review is what separates deployments that work from deployments that get switched off. Call history has the recording and full text for every call.
Step 5: cut over gradually
Once the parallel period is producing calls you are happy with, widen it in stages. Each stage is one setting at your telephony provider, and each is reversible in a minute.
- Forward on no answer. Your team keeps first refusal; the agent catches everything they miss. The most popular permanent configuration.
- Forward on busy. The agent becomes the overflow, and no caller ever gets an engaged tone.
- Forward always. The agent is the front door and transfers what it cannot handle.
Because you are forwarding rather than porting, your existing number never moves and rolling back is a single change at your current provider. The mechanics are in phone numbers.
Staff who discover an AI is answering the phones by overhearing a transferred call will assume it is coming for their jobs. Frame it accurately and early: it is taking the calls nobody was answering, and the repetitive five questions, so the team handles the conversations that actually need a person.
What to measure
Compare against your pre-migration baseline, which is why the audit in step one matters.
| Metric | What good looks like | What it tells you |
|---|---|---|
| Calls answered | Up, often sharply | The clearest win. Calls that used to hit voicemail or an engaged tone now get an answer. |
| Deflection rate | Rising over the first month | Share resolved without a human. This is where the money is. |
| Transfer rate | Falling, then flat | Should drop as you close knowledge gaps. It should never reach zero — some calls need a person. |
| Average handling time | Down for routine calls | No menu to sit through, and the answer comes immediately. |
| Abandon rate | Down | Nobody is on hold. |
| Very short calls | Low | A cluster of ten-second calls means people are hanging up on the greeting. Rewrite it. |
Mistakes that send callers back
| Mistake | What happens | Fix |
|---|---|---|
| Rebuilding the menu in words | You have an IVR that talks. Same problems, new voice. | One open question. Let callers say what they want. |
| No transfer path | Callers who need a person get trapped, and tell everyone. | A transfer tool pointed at a staffed number, with clear triggers. |
| Transfer to an unstaffed line | Worse than no transfer — you promised a person and delivered ringing. | A hunt group or mobile. Out of hours, take a message instead. |
| Launching without the knowledge | It cannot answer the five things everybody asks, so it transfers everything. | Load the top twenty questions before going live. |
| Big bang cutover | Every problem arrives at once, on your main line, in front of customers. | Parallel run. After-hours first. |
| Nobody reads the transcripts | The same failure repeats for months. | Twenty minutes a week. It is the whole improvement loop. |
| Hiding that it is AI | Callers work it out and feel deceived. In places it is unlawful. | One word in the greeting. |
| A calendar that lies | It books over lunch, holidays and staff leave. | Block non-bookable time in the calendar itself. The agent believes what it reads. |
When to keep the IVR
Sometimes the honest answer is not to migrate, or not entirely.
- Regulated scripts that must be read verbatim. If a legal disclosure has to be delivered word for word before anything else, a recording is the more defensible mechanism.
- Emergency and safeguarding routes. These should reach a human as fast as physically possible. No conversation first.
- Pure switchboard with no questions. If callers genuinely only ever ask for a named person by extension, a directory is simpler and cheaper.
- Very low volume. Ten calls a week answered comfortably by the person at the desk does not need automating.
Hybrids are common and entirely reasonable: an agent as the front door with an immediate, unconditional route to a human for the categories that need one.
It is the lowest-risk version of this entire article. Those calls are already being lost, so there is nothing to break — and two weeks of real transcripts will tell you more about your callers than any amount of planning.
For the commercial framing rather than the operational one, see replacing IVR with AI, and the real cost of a voice agent for budgeting the minutes.
Frequently asked questions
How long does it take to replace an IVR with an AI voice agent?
A working agent takes an afternoon. A confident production cutover takes two to six weeks, and almost all of that is parallel running and tuning rather than building. The build is genuinely the easy part now — the time goes into reading transcripts, closing gaps in what the agent knows, and getting comfortable that escalation works.
Do I have to replace the whole phone menu at once?
No, and you should not. The safest migration puts the AI agent in front of one route first — after-hours, or overflow when the queue is full — while everything else stays exactly as it is. You get real calls and real transcripts with no risk to your main line, and you can widen it once the evidence is in.
Will I need to change my phone number?
No. The usual approach is to buy a number for the agent, test on it privately, and then forward your existing business number to it once you are satisfied. Your customers keep dialling exactly what they always have, and reverting is one setting at your current provider.
What happens if the AI cannot help the caller?
It transfers them to a human, provided you configured that — which every production agent should. An agent with no escape route will keep trying to handle a caller it cannot help, and that is the interaction people complain about. Point the transfer at a number that is genuinely staffed during the hours the agent runs.
Is an AI voice agent cheaper than an IVR?
The licence cost is often comparable; the saving is in what happens after the menu. An IVR routes a call to a queue, so somebody still has to answer it. An agent that resolves a booking or a balance enquiry without escalating removes the human minute entirely. The economics turn on your deflection rate rather than on the platform fee.
Can an AI agent handle several callers at once?
Yes. An agent is not a phone line — a dozen people can be in separate conversations with it simultaneously. This is usually the point at which teams realise they are replacing the queue rather than replacing a person.