The hard part of AI call routing is the destination list

The short version

• AI call routing classifies what the caller says and sends the call to a destination, rather than asking the caller to map their problem to a keypad menu.
• The limiting factor is usually the destination list, because most phone systems have queues nobody can describe and rotas that change weekly.
• A misroute costs more than a transfer to a general queue, so routing needs a confidence threshold and a default destination that is always reachable.
• If the person who picks up gets no context, the caller explains themselves twice and the system has added a step rather than removed one.
• A vendor's per-minute price covers one line item; transcription, the model, the voice and the carrier minutes are billed on top.

What is AI call routing?

It is using what the caller says to decide where their call goes. The caller describes the problem in their own words, a speech model transcribes it, a classifier maps it to one of your destinations, and the call is transferred with that context attached. Some systems then go on to handle the call themselves. The routing decision is the part that has to be right first.

Three components, three distinct failure modes. Transcription fails on accents, background noise, speakerphone and crosstalk. Classification fails when the caller describes a symptom rather than a category, which is most of the time. Transfer fails when the destination is unstaffed, or when the person who picks up receives nothing and has to start the conversation again.

A call is synchronous, and that constrains every other decision. There is no typing indicator and no free way to say 'one moment'. Silence on a phone line reads as a dropped call, so the routing decision has to be forming while the caller is still speaking.

How is AI call routing different from a keypad IVR menu?

It moves the mapping work from the caller to the system. A keypad menu asks the caller to translate their problem into your org chart. Most callers cannot do that, so they pick the option that sounds closest, arrive in the wrong queue, and get transferred by somebody who now has to apologise for the menu.

What it does not change is the destination table. AI routing does not invent queues, rotas or owners. If the menu was wrong because the underlying destinations were wrong, the AI version is wrong more politely and more quickly.

Keep the keypad for what the keypad is better at: account numbers, reference numbers, card digits, and a hard option that reaches a person. Digits spoken aloud are easy to mishear, and a key press is exact. A caller who wants a human should never have to argue with a classifier to get one.

What breaks first in an AI call routing project?

The routing table, not the model. Every destination needs five things written down: who or what receives the call, when that is staffed, what happens outside those hours, who covers when the named person is away, and what the caller hears while they wait. Most organisations have never written any of it down.

'Just send it to the right person' is an awkward brief, because the right person is often defined by a rota that lives in somebody's head and changes weekly. A model cannot read that rota. It can only read a system you connect it to, so the rota has to exist somewhere a machine can query.

This work is unglamorous and it is the work. Enumerate the destinations, agree the hours, pick the default. The default matters most, because every call the classifier is unsure about lands there, and a bad default turns ordinary uncertainty into a lost caller.

What does a routing decision need before it can be made?

• •The caller's intent, at the level of detail your destinations actually distinguish between.
• •Whether the caller needs to be identified, and whether that identity has been verified rather than claimed.
• •Whether the destination is staffed right now, including the on-call rota and public holidays.
• •A confidence threshold, below which the call goes to a person instead of to a guess.
• •A default destination that is always reachable.
• •What the caller hears while the transfer is happening.

How fast does the routing decision have to be?

Fast enough that the caller never hears unexplained silence. In practice that means classifying on partial transcripts while the caller is still talking, not starting the decision once they stop. Any gap a caller would describe as 'it went quiet' is a defect, because that is the moment they hang up or say hello twice.

Budget the whole chain, not the model call. Transcription, the classifier and any lookup against a CRM, rota or ticketing system all sit inside the caller's patience. The lookup is usually the slowest part and the easiest to underestimate, because it is instant in a test against one record and slow against a production system on Monday morning.

Two mitigations do most of the work. Run the lookup in parallel with classification rather than after it, and speak before you are certain: 'let me get you to the right team' buys a second of legitimate audio. Then put a timeout and a fallback behind every external call, so another system's bad day never becomes the caller's silence.

What happens at the moment of transfer?

Either the person who picks up knows why the call is coming, or the caller explains themselves twice. A cold transfer with no context has added a step to the call rather than removing one, and callers notice. The transfer is where most of the value of routing is either delivered or thrown away.

Decide which mechanism you are using and test it from the receiving seat, not the caller's. A warm transfer keeps the system on the line until a human answers. A cold transfer hands the leg away and the system is gone. Context can travel as a screen pop, a note on the record, a SIP header or a spoken summary before the caller is connected, but it has to travel somehow.

Then define what happens when nobody answers. Ringing out into nothing is the worst available outcome, because the caller has already described their problem once and is now further from help than when they dialled. Set the ring timeout, choose where the call lands next, and make sure the context follows it there.

How do you tell whether AI call routing is working?

Measure misroutes, not containment. Containment counts the calls the system kept, which is the number a vendor prefers to show you. The misroute rate counts the callers who arrived somewhere that could not help them, which is what the caller actually experienced.

Four measures are usually enough: onward transfers after the first destination, repeat calls from the same number inside a day or two, abandonment during the routing turn itself, and time to the first human word on calls that need a human. Each of them moves when the routing table goes stale, which it will.

Then read the transcripts. Sampled, on a schedule, by somebody who knows the business. Automatic labelling reports the intent the classifier assigned, and that label is the thing under audit, so it cannot also be the evidence. Reading a handful of calls each week finds problems no dashboard will surface.

What does AI call routing cost to run?

More than the per-minute figure on the pricing page, because that figure is one line item in a stack. Vapi publishes $0.05 per minute as a hosting fee only, with transcription, the model, voice and telephony billed separately at provider cost, and its page lists Twilio inbound at $0.008 per minute. Retell AI's published pay-as-you-go pricing stacks voice infrastructure at $0.055 per minute with text to speech, the model, and US telephony at $0.015 per minute on top.

The arithmetic then depends on what you are buying. A routing turn is short, often well under a minute, so routing by itself is cheap per call. Minutes accrue when the system handles the whole conversation, which is a different product with a different risk profile. Compare prices after you have decided which of the two you want, not before.

The costs that never appear on a pricing page are the destination table, the integration with the rota or CRM, the weekly transcript review, and somebody owning the routes when the business reorganises. Those are the line items that decide whether the thing still works in six months.

Where does the per-minute cost come from?

When is AI call routing not worth doing?

When the calls already reach the right place. If two people answer everything and both can help with anything, routing is not your constraint; volume or hours of cover is. A classifier there adds a new way to be wrong without removing any work.

It is also not worth starting when the destination list cannot be agreed. If two managers give different answers about where a billing dispute should go, a model will not settle that argument. It will encode whichever answer was in the prompt that week, silently, until somebody complains.

The adoption data is consistent with the mechanism. Census Bureau survey data shows large firms with at least 20 employees are the biggest users of AI among US businesses, and routing pays off where there are enough distinct destinations, and enough volume, for a misroute to be expensive. Interest is not confined to large firms: Verizon's own survey of 600 US small businesses reported a surge in demand for AI. A vendor-run survey of 600 firms is weak evidence for anything beyond direction.

Does AI call routing replace the people who answer the phone?

No. It replaces the keypad menu and the 'who should I put you through to' question. Answering the phone remains a large job: the Bureau of Labor Statistics puts national employment for receptionists at 947,500 for its May 2025 reference period. The routing decision is the mechanical part of that role, not the whole of it.

The outlook for adjacent work is softer, and the numbers should be stated precisely. BLS projects customer service representative employment to fall by about 141,800 jobs from a 2025 base of 2,666,000 over the decade to 2035. That is a projection rather than an outcome, and it is not attributable to any single technology.

It is also worth noticing the shape of the deployments that have been measured properly. One of the few field studies of generative AI in support followed a staggered rollout of a conversational assistant across 5,172 customer support agents at a software firm, with the system sitting beside the human rather than taking the call. Routing is a smaller step than that, which is a point in its favour.