How to Route Calls by Intent with AI: Lessons from 12M+ Calls a Year
AI call routing lessons from a telecom handling 12M+ calls a year: why misrouting multiplies cost, and how to handle the caller who asks for a human.
Nico Silva, Lead Forward Deployed Engineer · · 4 min read
On this page
- The problem was misrouting, not volume
- Lesson 1: the failure is the moment callers ask for a human
- Lesson 2: 120+ intents need disambiguation, not menus
- Lesson 3: accuracy cannot cost call time
- Lesson 4: keep your telephony
- Lesson 5: scale changes the design
- What to do before you start
- What AI routing does not solve
AI call routing means an agent asks why the caller is phoning, understands the answer, and sends the call to the place where it gets solved. Done well, it removes calls that never needed a person and stops the chain of wrong transfers. Here is what we learned running it for a telecom operator in an operation of 12M+ calls a year across 120+ intent categories.
The problem was misrouting, not volume
The operator’s legacy IVR misrouted about 25% of calls. That number is the heart of the story, because misrouting is not a user-experience annoyance. It is a cost multiplier.
A wrongly transferred call often triggers a second transfer. About 250K calls a month were reaching human agents, and a large share of those did not need to. Repeat contacts compounded the cost. Most of the money wasted came from routing failures, not from the number of calls.
The fix was to change the routing decision itself. After the redesign, misrouting fell to under 7%, human-bound calls fell from about 250K to about 70K a month, and that is the 180K fewer human-handled calls a month. We estimate that at about $3.2M a year, based on roughly $1.5 per human-handled call. That is the client’s number on the client’s operation.
Lesson 1: the failure is the moment callers ask for a human
We expected the hard part to be intent classification. It was not. The failure was losing control at the exact moment callers asked for a person.
Many callers say “agent” or “representative” in the first seconds. A traditional IVR either obeys, sending the call to a queue with no information, or ignores them and repeats the menu. Both are bad.
Our answer was to treat that moment as a conversation. The agent acknowledges the request, asks what the caller needs, and often discovers that the caller can be sent to the right place directly, or that the question is one the agent can answer. When the caller truly needs a person, the agent transfers with the reason attached.
The lesson for any routing project: design the “I want a human” moment first. If you get it right, the rest becomes easier.
Lesson 2: 120+ intents need disambiguation, not menus
With 120+ intent categories you cannot present a menu, and you cannot rely on one-shot classification either. Callers describe problems vaguely: “it’s about my bill,” “my service isn’t working.”
The agent asks again when the request is unclear. A short follow-up is better than a confident wrong guess. This is where conversational design matters more than the model.
Lesson 3: accuracy cannot cost call time
Routing accuracy that adds a minute to every call would have been a bad trade. The average interaction stayed under 28 seconds, with no increase in call duration. If you measure only misrouting, you can make a system that is accurate and slow. Measure both.
Lesson 4: keep your telephony
The system works without replacing existing telephony infrastructure. That kept the project out of a long migration. If a routing project begins with “first we must change the phone system,” it will probably not finish.
Lesson 5: scale changes the design
At 12M+ calls a year, small percentages are large numbers. A 1% error is 120K calls. The system needed sub-second decisions and an architecture built for load, not for demos.
What to do before you start
Map your intents. Pull a few weeks of call reasons and group them. You will probably find a long tail. The head of the list is your first release.
Measure the baseline. Misrouting rate, transfers per call, human-handled calls per month, average call length. Without a baseline you cannot tell if the agent helped.
Define the destinations. Routing needs somewhere to send the call. If two queues overlap or no queue owns a problem, fix that first.
Write the handover. What does the person see when a call arrives? The reason, what the caller said, what the agent already asked. If the answer is “nothing,” you have only moved the problem.
Roll out in steps. We begin with 30–100 interactions a day, review daily, and scale only after the system is stable. See our post on moving from pilot to production.
What AI routing does not solve
It does not fix a back office where the right team does not exist. It does not decide your policies. And it does not remove the need for people. The goal is that the calls reaching your team are the ones that need them.
If you run a high-volume inbound line and suspect misrouting is costing you more than it should, bring us a sample of your call reasons and we will show you where an agent would start.