AI Call Routing: How Intent, Time and Caller Type Should Decide Where a Call Goes

What routing has to fix

The traditional answer to call routing is a menu. Press one for sales, two for support, three to hear these options again. It works in the narrow sense that calls end up somewhere, and it fails in every other sense: callers guess, guess wrong, get transferred, and repeat themselves to the second person. The menu optimises for the switchboard's convenience rather than the caller's.

Intelligent call routing replaces the menu with a question. The caller says why they are calling, in their own words, and the system decides where that belongs — using what they said, who they are, what time it is and who is available. That is a better experience and, more importantly, a more accurate one, because callers describe their problem more reliably than they classify it.

But 'smart routing' is claimed by every vendor and means very little on its own. What makes routing work is a written, priority-ordered rule set that someone can read, test and argue with. This article is about building that.

The four inputs

Voice routing reads four signals. The first two are shared with text channels; the last two matter considerably more on the phone.

1. 1

   Intent: what the caller said they need

   Captured from an open question rather than a menu. The classification needs a confidence threshold, and below it the correct behaviour is one clarifying question — not a guess, and not a menu as a fallback.

2. 2

   Caller identity: who is calling

   The phone number itself is a strong signal that text channels do not have. A number matching a customer record tells you the relationship before the caller says anything, and it is the cheapest routing input available.

3. 3

   Time: when the call arrived

   Time-based routing carries far more weight on the phone than in chat, because a call cannot wait politely in an inbox. Opening hours, lunch cover, holidays, out-of-hours and the difference between a Tuesday morning and a Friday evening all change the correct destination.

4. 4

   Availability: who can actually take it now

   Real-time agent state, queue depth and expected wait. Routing to a destination with a twelve-minute queue when another could answer immediately is a failure the caller experiences directly.

Writing the rule set

Express routing as a priority-ordered list where the first match wins. Order matters more here than in text routing, because the cost of a wrong transfer is higher — the caller has to explain themselves again, live.

• 1. Emergency or safety-critical language → the emergency destination, immediately, bypassing everything below. Define these categories explicitly for your sector.
• 2. Explicit request for a person or for a named individual → the fastest staffed route, no retention attempt.
• 3. Caller number matches an account with a contractual response commitment → the named owner, falling back to that team's queue.
• 4. Caller number matches an open ticket or active deal → whoever owns it, so the caller does not restart.
• 5. Existing customer + service issue → support, with the account context already attached.
• 6. Existing customer + commercial question → account management, not new business.
• 7. Unknown caller + buying intent → sales, live if staffed, booked if not.
• 8. Unknown caller + documented information request → handled automatically, with an offer to reach a person.
• 9. Anything arriving outside staffed hours → the out-of-hours flow with an honest commitment.
• 10. Unmatched → the default queue, reviewed weekly.

Two disciplines keep this maintainable. Every rule has exactly one owner, and the default queue has a named person whose job includes reading what lands there. Both points apply equally to text-channel routing; the difference on the phone is that failures are immediate and audible.

Time-based routing in practice

Time deserves its own treatment because it is where most routing tables are wrong in ways nobody notices until a caller complains.

• Opening hours per team, not per business. Sales, support and accounts rarely keep identical hours, and a single business-hours setting routes calls to empty desks.
• Holiday calendars, maintained by someone whose job includes it. This is the most common source of wrong routing and the easiest to prevent.
• Lunch and shift-change cover, which is when a surprising share of abandonment happens.
• Predictable peaks — Monday mornings, the day after a campaign, month end — where overflow rules should be pre-planned rather than improvised.
• Time zones where you serve multiple markets, so a caller is not told the office is closed when the relevant office is open.
• A distinct out-of-hours behaviour per intent: an emergency at 2am is not the same as a pricing question at 2am, and routing them identically is a mistake.

The hourly distribution from a missed call analysis is the right input for setting these rules. Most businesses find their time rules were written from assumptions rather than from data.

What the caller should experience

Routing decisions are invisible when they work and painful when they do not. A few design choices make the difference.

1. 1

   Ask once, in an open question

   'What can I help you with today?' produces better routing than a menu and takes less time. Callers who do not know how to answer can be offered options as a fallback, not as the default.

2. 2

   Confirm the destination before transferring

   'I will put you through to the service team' sets expectations and lets the caller correct a misroute before it costs them a repetition.

3. 3

   Carry the context through the transfer

   The receiving person should see the stated reason, the identified account and anything already collected. A transfer that arrives cold wastes the routing entirely.

4. 4

   Be honest about waits

   If there is a queue, say so and offer a callback that holds the caller's place. An accurate wait estimate with an alternative converts abandonment into a scheduled conversation.

5. 5

   Never loop

   A caller who has been transferred twice should reach a person who can resolve or own the problem, not a third destination. Build that as an explicit rule.

The callback-instead-of-queue option is consistently the highest-value single addition to a routing design, because it addresses the one part of the experience the caller finds least acceptable.

Data and integration requirements

Routing quality is bounded by what the system can look up while the caller is waiting — which on the phone means within a second or two.

• Caller number lookup against customer records, fast enough not to add audible delay.
• Entitlement or contract data expressing service commitments, so priority routing is driven by a field rather than by memory.
• Open ticket and active deal state, to route callers back to whoever already has context.
• Real-time agent availability and queue depth per destination.
• Per-team hours and a maintained holiday calendar.
• Telephony integration capable of transferring with attached data, not just transferring the audio.
• Logging of every routing decision with its inputs, retained long enough to investigate a complaint.
• A defined behaviour for every lookup failure — usually route on intent alone and flag the call, never stall the caller.

The transfer-with-context requirement is the one most often discovered late. Many telephony setups can transfer a call and cannot carry the data with it, which undoes most of the benefit. Our automation services page covers how this layer is typically built, and an AI call centre handles much of it natively.

Metrics

Voice routing has its own measures, and answered rate is not among the useful ones.

1. 1

   Transfer rate after first routing

   How often a call moves destinations after the initial decision. The clearest single indicator of routing accuracy, and it should be broken down by source rule so you can find the wrong ones.

2. 2

   Time to the right person

   Not time to answer. The gap between the two is the cost of misrouting, and it is what the caller actually experiences.

3. 3

   Repeat-explanation rate

   How often callers had to restate their problem after a transfer. Sampled by listening rather than measured automatically, and the most honest indicator of whether context is really travelling.

4. 4

   Default-queue volume and composition

   What is falling through. A rising number means the business changed and the rules did not.

Review transfer rate by rule weekly at first. Routing tables decay quietly as products, teams and hours change, and this is the first place it shows.

Failure modes

These recur in voice routing specifically.

• Replacing a menu with an open question and then falling back to the menu whenever classification is uncertain, which gives callers the worst of both.
• Routing on intent while ignoring the caller's number, so known customers are treated as strangers.
• One business-hours setting for teams with different hours.
• Holiday calendars nobody maintains.
• Transferring audio without data.
• Queues with no callback option.
• Loops, where a caller is passed between destinations.
• No defined behaviour when the customer lookup is slow, so the caller hears silence.
• Routing rules that require a deployment to change, so hours and owners drift out of date because updating them is a project.

Where routing should stop

Automated routing assigns calls. It should not make judgements about the relationship.

• Anything with legal, regulatory or press language should reach a named person rather than a queue.
• Escalated complaints, where routing by topic sends the caller to the team they are complaining about.
• Callers in distress, who should reach a person immediately regardless of what the intent classifier concluded.
• Strategic accounts, which should bypass the rules entirely.
• Any call where the routing decision is obviously wrong to a person listening — and agents should be able to re-route and record why, since those records are the best source of rule improvements.

That feedback path is what keeps routing accurate over time. Agents know where calls should have gone long before any report does.

Decision framework and next step

Four questions before writing rules.

1. 1

   What destinations genuinely exist, and when are they staffed?

   Start from the rota. Rules cannot promise coverage that does not exist.

2. 2

   Can you identify callers from their number?

   If not, that is the highest-value thing to add before anything else.

3. 3

   Can your telephony transfer context with the call?

   If it cannot, solve that before refining classification — otherwise every correct routing decision still costs the caller a repetition.

4. 4

   Who owns the default queue?

   Name them. An unowned fallback is where callers go to be forgotten.

Start with the smallest table that covers real traffic: emergency, explicit request, known-customer lookup, one intent split, out-of-hours, and a default. Add rules as the default queue shows you need them. For choosing the operating model behind the rules, see AI receptionist vs answering service; our AI solutions overview covers staging.

Frequently asked questions

1. 1

   What is AI call routing and how does it differ from an IVR menu?

   It replaces a numbered menu with an open question and decides the destination from what the caller said, who they are, the time and who is available. Callers describe their problem more reliably than they classify it, so routing accuracy improves and transfers fall.

2. 2

   What data and integrations does it require?

   Caller number lookup against customer records, entitlement data, open ticket and deal state, real-time agent availability and queue depth, per-team hours with a maintained holiday calendar, telephony that transfers data alongside the call, and routing decision logging.

3. 3

   Which metrics show whether routing works?

   Transfer rate after the first routing decision broken down by rule, time to the right person rather than time to answer, repeat-explanation rate sampled by listening, and the volume and composition of the default queue.

4. 4

   What are the most common mistakes?

   Ignoring the caller's number, using one business-hours setting for teams with different hours, unmaintained holiday calendars, transferring audio without context, queues with no callback option, and rules that require a deployment to change.

5. 5

   When should a call bypass the rules?

   Emergency and safety-critical language, explicit requests for a person, legal or regulatory language, escalated complaints, callers in distress, and strategic accounts.

Routing quality is bounded by two things: what the system can look up while the caller waits, and whether the transfer carries context. Both are integration questions rather than model questions.