AINovember 20, 20267 min read

Skills-based routing and how to get callers to the right human agent faster

Portrait of Joe Huffnagle

Joe Huffnagle

VP Solution Engineering & Delivery @Parloa

At 10 a.m., the billing specialists are between calls, and the general queue shows a nine-minute wait. You walk the floor and see why: callers who pressed “account questions” on the menu were calling about invoices, and the system transferred each one the moment a human agent picked up.

The skills-based routing engine read the tag on each call and matched it to a human agent with that skill. The tag was wrong, so the match was wrong, and the specialists who could have closed those calls sat idle. So how could the routing engine know what the caller never told it?

What is skills-based routing?

Skills-based routing is a contact center method that assigns each incoming call to the available human agent whose skills best match what the call needs. Instead of routing calls to whoever frees up next, it reads call attributes, such as the dialed number, the caller’s identity (opens in a new tab), the menu selection, and any customer relationship data, and compares them against each agent’s skills profile.

That profile typically includes languages, product knowledge, authorization levels for actions such as refunds, and any required license. Many routing engines also store a proficiency level per skill, so a call can go to the strongest available match first and fall back to a competent one when needed.

How the engine matches call tags to skills

The engine’s decision depends on two inputs: a tag describing what the call needs and a skills profile describing what each human agent can do. Here is how the engine builds and uses each one:

  • Dialed number: The engine checks which line the caller reached, such as the claims line or the sales line, and uses it as a first clue to the intent.

  • Caller identification: It identifies the caller from the calling number or an account lookup, which pulls in customer relationship data such as a premium tier or an open claim.

  • IVR selections: The caller’s choices in an Interactive Voice Response (IVR) (opens in a new tab) menu feed into the tag and usually carry the most weight.

  • Skill derivation: From those inputs, the engine derives the skill the call needs and searches for a human agent with that skill at the required proficiency.

  • Overflow rules: When no one with that skill is free, overflow rules send the call to a less specialized human agent after a set wait, keeping the queue moving at the cost of a weaker match.

Because menu selection carries the most weight, look at intake first when a call reaches the wrong skill. If the tag is wrong, the match is wrong, no matter how well the skills profile is drawn. That is where most routing failures begin.

Where the routing decision goes wrong

Callers reach the wrong human agent when intake assigns the wrong tag or points to a poorly defined skill. The matching logic rarely picks the wrong agent for the tag it was given. Menu choices and skill definitions can fail in four ways before or around the matching engine.

Coarse intake tags

A menu can offer only a handful of branches, so “account questions” has to stand in for invoices, address changes, cancellations, and refund requests. One tag covers a dozen intents and routes all of them to the same queue. Use detected intent (opens in a new tab) to separate those requests before routing.

Self-selected intent

Callers pick the option that sounds closest, often before the full list plays. A customer disputing a charge chooses “billing” or “complaints,” and each choice labels the same call differently. Capture the request in the caller’s own words before assigning its intent.

Skill sprawl

Every new skill splits a queue. Ten human agents in one queue absorb a spike; the same ten spread across four skills leave two or three per slice, and the wait climbs the moment one is on a long call.

Stale skill profiles

Skills defined at the last reorganization fall behind product launches and policy changes, and behind the human agents who have since been cross-trained. Review those profiles with workforce management before the next routing change.

Coarse intake tags and self-selected intent start at intake; skill sprawl and stale skill profiles start in skill design, and all four surface as transfers and longer handle times. A mislabeled call reaches a human agent who cannot close it, so it becomes a transfer; the second human agent restarts verification and context, so average handling time (opens in a new tab) (AHT) rises for a call that should have taken one conversation.

Skill sprawl adds cost before the transfer: callers wait in a queue sized for a fraction of demand. Adding skills to fix mislabeling worsens fragmentation, because each new skill needs its own menu branch, giving callers one more option to guess at.

Moving intent detection in front of the routing engine

An AI voice agent (opens in a new tab) that answers the call first gives the routing engine what a menu cannot: the caller’s actual intent, in their own words, before any queue is involved. Instead of forcing the caller to guess at a label, the AI agent builds a complete call record during the conversation and hands it to the engine as structured data. Here is how it does that:

  • Open-question intake: The AI agent greets the caller and asks why they are calling. The caller answers in a full sentence, and the agent classifies that sentence into an intent at a level of detail no menu reaches, so “I was charged twice for March” lands as a duplicate-charge dispute on a specific invoice.

  • Live authentication: The AI agent verifies the caller against the account record using the factors the target skill requires, such as a policy number and date of birth, so verification does not have to run again after the handoff.

  • Context collection: It gathers the details the human agent will need, such as the invoice number or the claim reference, and confirms them back in speech.

  • Language detection: If a caller speaks a different language than the line expects, route them to a language-specific AI agent (opens in a new tab), and set language as a call attribute.

  • Containment: Simple requests, such as a balance check or a document resend, can end inside the AI conversation, so the human queues receive mostly the calls that need judgment.

  • Structured handoff: The system writes intent, identity, and collected details onto the call, and the human agent receives a summary of the conversation so far (opens in a new tab).

With that record in place, the routing engine matches on facts rather than guesses: a verified customer with a duplicate-charge dispute on a known invoice goes to the billing skill, with no menu branch in between.

How cleanly the AI agent passes structured data to the routing engine, and how quickly it responds while the caller is talking, depends on the conversational AI architecture behind it. Uelzener Versicherung’s Clara (opens in a new tab) replaced a menu IVR with an AI intake step and reports nearly 100% accuracy in routing calls to the correct skill team and a low call-abandonment rate.

Designing skills for AI-assisted intake

Once intake is accurate, the skill table can shrink. Fewer, broader skills route better than a large set because a well-tagged call no longer needs a menu branch to find the right queue. Narrow skills belong only where a license, an authorization level, or a language requires them. Use the following principles to redesign the table.

1. Collapse menu-derived splits

Check any split that exists only because the IVR needed one, such as “billing disputes” and “billing general,” against the intent written to the call in the caller’s own words. Collapse it back into one billing skill staffed by the full team, and let the intent attribute decide the order in which calls are offered inside that queue.

2. Keep narrow skills only where qualification requires it

The skills narrow only when a human agent cannot legally or operationally cover them without a specific qualification. A regulated product that only licensed advisors may discuss, or a language with three fluent speakers on the floor: each needs its own skill because the wrong match (opens in a new tab) is a compliance failure or a dead end, and no amount of context fixes that.

3. Apply the cross-training test

For every other skill, ask whether a cross-trained human agent with a good summary could close the call. If yes, the skill is a menu artifact and can go. If no, record which authorization or qualification is missing before you decide to keep it.

4. Set overflow rules per skill

Overflow rules belong to each skill, with no global timer. A licensed-advice queue should hold callers longer or offer a callback, because overflowing to an unlicensed human agent at best results in a transfer. A general billing queue can overflow to the customer service pool after a short wait, because a human agent with the intent and the invoice number in front of them can handle the call.

5. Review on a fixed cadence

Review the skill table with workforce management on a fixed cadence and bring the intents the AI agent captured in that period to the meeting. A skill that received a few calls an hour for three months is a merge candidate before anyone argues about it.

Measuring routing quality

Routing accuracy alone is a weak signal. Accuracy is the share of calls that reach the right skill without a transfer, and it’s easy to inflate by drawing skills so broadly that every call qualifies. To know whether routing is actually working, read accuracy alongside average speed of answer, transfer rate, abandonment, and first call resolution (opens in a new tab) (FCR), and treat the combination as the dashboard.

Here is how to read each one together:

  • Transfer rate shows whether intake is tagging calls correctly. A falling transfer rate confirms that the tag reaching the engine matches what the caller actually needs.

  • Abandonment shows whether the wait in each queue stayed short after skills were collapsed or split. A queue that got narrower will register here first.

  • FCR shows whether the receiving skill can actually close the call. When routing accuracy climbs and FCR stays flat, calls reach the skill as defined, but the skill can’t close them; the boundary needs to move, or the human agents in it need authorization they lack.

The misroutes are the most useful output of the month. Every call that transferred out of its first skill carries the intent the AI agent captured and the skill the engine chose, so the difference between the two is the list of boundaries to redraw, intent by intent, before the next skill review.

When the full set of metrics moves together, the gains show up on both sides of the conversation: Swiss Life (opens in a new tab), for example, reports 96% routing accuracy and is 60% faster at addressing customer concerns.

Rebuild skills-based routing around what callers actually say

Skills-based routing fails most often at intake, not at the matching engine. When “account questions” has to stand in for a duplicate-charge dispute, the specialists sit idle, and the general queue grows.

Fixing that starts with capturing the caller’s intent in their own words, authenticating them in the same conversation, and handing the routing engine structured data, not a menu branch. With accurate intent on every call, the skill table can shrink to the skills that qualification actually requires, and the metrics that matter, transfer rate, abandonment, and FCR, start moving in the same direction as accuracy.

Parloa (opens in a new tab) is an AI Agent Management Platform that handles the intake step in front of your existing contact center and routing engine, so you don’t need to replace either. The AI agent captures intent, verifies identity, collects the details a human agent will need, and resolves simple requests in the conversation. When a human agent is required, Parloa passes the full context to the right queue through integrations with contact center platforms, so the human agent who answers already knows why the caller called.

Book a demo (opens in a new tab) to see how AI agents capture caller intent and pass it to human agent queues.

FAQs about skills-based routing

How can you prepare a skill-table review?

Bring the current skill table and captured caller intents to workforce management. For each proposed merge, record which cross-trained human agents could close those calls and which qualifications still require a separate skill. Use that record to agree on the changes.

What should you check in an AI intake handoff test?

Use a duplicate-charge dispute on a known invoice as a test call. Check that the billing queue receives the verified identity and invoice number, then ask the receiving human agent whether the summary contains enough context to continue without repeating verification.

How long does it take to add AI intake to an existing routing setup?

First use cases can go live in as little as a few weeks because the AI agent sits in front of the existing contact center platform and routing engine, and neither needs to be replaced. Later intents follow as a phased deployment.

Abstract dark brown and orange gradient

Ready to turn conversations into lasting loyalty?

Let's build your next great customer experience.