Interactive voice response
Interactive voice response earned its reputation honestly: a tree of menus built around your org chart, which the caller has to solve before anybody helps them. The tree exists because a machine could not understand a sentence. That constraint is gone, and most of the trees are still up.
Why the tree was built, and why it stayed
A menu is a routing table read aloud. It was the only way to get a caller to the right queue when the only thing a phone system could recognise was a keypress. Every branch you added made the routing better and the call worse, and the trade was worth it because there was no alternative.
What keeps them up now is not technology. It is that the tree encodes decisions nobody has written down anywhere else — who handles a warranty claim, what counts as urgent, which office covers which postcode. Rip out the menu and those decisions have nowhere to live. That is the actual project.
Read your own menu back as a list of decisions
Before anybody builds anything, write each option a caller hears today as a sentence about who handles what. That list is the real specification, and it tends to surprise the person who owns it.
- The caller who wants both. Press one for sales, two for support — so where does somebody with a question about an invoice for a thing that never arrived belong? The tree decided that years ago and nobody wrote it down.
- The option nobody chooses. Every menu has one, and it is usually a queue built for a problem that stopped happening two years ago. It still costs every caller the seconds it takes to say it.
- The option everybody chooses. If most callers press zero or simply wait, the tree is not routing anything. It is a delay with a recording on it, and you are paying for the recording.
- What it says at nine at night. The out-of-hours branch gets written once and never re-read, and it is where the expensive misroutes live — the ones nobody is awake to catch.
- The route that exists because of one person. They leave, the branch stays. That is how a tree grows a limb nobody can explain, and why nobody dares prune it.
What replaces it
- One question instead of nine. The caller says why they are calling, in their own words, and the system routes on the meaning rather than on a digit. No list to memorise, no going back a level.
- The routing rules written down. Every branch of the old tree becomes an explicit rule with an owner. This is the part that takes the time, and it is worth doing even if you never build the voice layer.
- Identification before the queue. Account number, order number, policy number — captured and verified while the caller is still talking, so the person who picks up already knows who it is.
- A real handover. When it goes to a human it goes with the transcript and the reason attached, so the caller does not start again. Repeating yourself to the second person is what people actually hate.
- A recorded reason for every route. Why this call went where it went, kept and readable, so a bad routing rule can be found instead of argued about.
The four it must refuse
- Refuse to reach a person. Any caller who asks for a human gets one, on the first ask, with no loop back into the system. A machine that traps people is worse than the menu it replaced.
- Guess an identity. If the account number does not verify, it says so and hands over. An assumed identity in a phone system is a data-protection incident waiting for a bad day.
- Decide an emergency on its own. What counts as urgent in your business is written down by you, and anything matching goes straight to a person — not into a queue, not into a callback list.
- Quote a price, agree a refund or confirm a delivery date. It can read back what is already written in your system. Anything that creates a new obligation stops and waits for somebody who can make it.
Not worth building yet if
- Your call volume is under a few dozen a day and one person answers all of them. A person is better and cheaper at that volume, and there is nothing to route.
- The routing rules are genuinely undecided. If two managers would send the same call to different places, that is a decision to make, not a thing to automate. Writing them down is the first job and it does not need a machine.
- Your systems cannot be read from. If the account data lives somewhere that has no way in, the system can be polite but it cannot be useful, and a polite machine that knows nothing is a slower menu.
Three pieces of it, in production now
- Customer-chat assistant — the same routing and the same refusal rules, in text, in production: it answers in the business's voice and a second model blocks anything it should not promise.
- Personal AI assistant — speech turned into structured text and filed where it can be found — the half of this that has nothing to do with the phone network.
- Systems watchdog — the monitor that notices when something stops running. A phone line that silently stops answering is the worst failure this can have, so watching it is part of the build.
What people ask before replacing a menu
Is this the same as an IVR system?
It is what an IVR is for, without the tree. The job of interactive voice response has always been to get a caller to the right place with the right context. The keypad menu was the only way to do that when a machine could not follow a sentence. The job stays, the menu goes.
Do we have to replace our phone system?
No, and you should not want to. This sits in front of what you already run and hands calls to it. If your provider has a way in — most do — the work is in the routing rules and the handover, not in ripping out the switchboard.
What happens to callers who just want a person?
They get one on the first ask. That is a rule, not a setting, and it is tested before anything goes live. A system that makes people fight for a human generates more angry calls than it deflects.
How long before it is answering real calls?
The build is short; the writing-down is not. Expect the routing rules and the never-do list to take longer than the voice layer, because they are decisions rather than code. A first working version on a narrow slice of calls is a matter of weeks, not months.
How does the bill behave on a busy day?
It follows the minutes rather than the roster, so a day with four times the calls costs four times more and needs nobody extra — which is exactly the day a queue used to cost you customers instead of money. Whether that trade is worth making is a sum on your own call log, and the AI voice agent page works it through.
Count the calls that gave up
Pull last month's call log and count how many callers hung up inside the menu. Then count how many of those called back. That pair of numbers decides whether this is worth anything to you, and you do not need me to work it out. If it points at building something, the first conversation is an hour and it is free.
Book that hour →