Call Center Routing Software Is Judged on Its Fallbacks

Updated 23 min read How we research

TL;DR: Call center routing software is sold on a list of strategy names and judged on what it does when no rule matches. Those names are not a standard. In the Kamailio dispatcher module, selection algorithm 4 is round-robin; in the OpenSIPS dispatcher module the same integer is weighted round-robin, algorithm 8 means “sorted by priority attribute” in one and “the first entry in set” in the other, and OpenSIPS documents that an algorithm it does not implement quietly falls back to the first entry instead of erroring. Agent selection has the same shape: the FreeSWITCH mod_callcenter module recognises ring-all, ring-progressively, longest-idle-agent, agent-with-least-talk-time, agent-with-fewest-calls, random, sequentially-by-agent-order, top-down and round-robin, orders every one of them by tier level first, and falls through to level-then-position when the strategy string is not recognised. Its shipped sample queue and its compiled defaults disagree on two of the four tier parameters, agents whose status is On Break stay in the candidate set alongside Available, and the caller-side controls that decide whether a long hold becomes an abandoned call (max-wait-time, max-wait-time-with-no-agent, discard-abandoned-after, abandoned-resume-allowed) ship at values that do nothing. Buy on the fallback path, not the feature list: ask what happens on no answer, on an empty queue, after hours, and on an unknown value, and get the answer as a configuration file rather than a slide. Kixie includes rule-based routing, ring groups, call queues and an IVR auto-attendant on all three plans, with a seven-day free trial and no credit card.

Here is the thing nobody demos. A buyer calls your main line at 4:52 on a Friday afternoon, two reps are already on calls, a third is listed as on break, and the rule you wrote months ago says the call goes to whoever has been idle longest. There is no idle rep. So what happens to that call?

That answer is the product. Everything above it is marketing.

Most buying processes for call center routing software run the other way around. The vendor opens a rule builder, drags a condition onto a canvas, and routes a test call to a named rep. It works. Of course it works. The demo was built from a call that matches a rule. The hard part of routing is every call that does not.

What Call Center Routing Software Decides on Every Call

Strip the category label off and routing software makes four decisions, in order. Where does the call land first? Which group of people is eligible to take it? Which one of those people gets offered it, and in what order? And what happens to the caller while all of that is being worked out? Four questions. Most products answer two of them well.

Those four decisions live in different layers, and vendors blur them on purpose because one word covers all of it. Number-to-destination mapping is routing. Picking a server or a gateway is routing. Picking a human is routing. They fail in completely different ways, they are usually owned by different people inside the vendor, and the words on a feature list do not tell you which of the three you are being shown on the screen in front of you.

You can see the layers cleanly in open-source telephony, and that is the useful thing about it. Nobody is selling you anything. The behaviour is written down, the code is readable, and the defaults are public. A SIP proxy such as Kamailio or OpenSIPS handles the first layer, deciding which destination in a configured set receives the call. A queue module such as FreeSWITCH mod_callcenter handles the agent layer, holding the caller and offering the call to people in a defined order. Commercial products wrap both and rename them. The vocabulary underneath is the same vocabulary, which is exactly why the next section matters.

So what is the useful question in a demo? Not “can it route by skill,” which every product answers yes to, but “which of those four decisions is this screen actually making, and where do I configure the other three.” Ask it early. The answer tells you whether you are looking at one product or a bundle of three with a shared login.

The Strategy Names on the Feature List Are Not a Standard

Two SIP routing modules with nearly identical documentation, nearly identical function signatures and nearly identical parameter names disagree about what the numbers mean. Is that a corner case? No. It is the normal state of this vocabulary, and it is worth seeing in detail before you trust a comparison table.

Two identical upright frosted glass dials standing side by side on a pale violet ground, each ringed with the same evenly spaced tick marks, one slender glass pointer aimed up and to the left and the other aimed down and to the right.

Kamailio’s dispatcher module documents a selection function that takes a destination set and an algorithm number, and the algorithm number is the whole behaviour. Algorithm 0 is a hash over the Call-ID, 1 through 3 are hashes over the From URI, the To URI and the request URI user, 4 is round-robin, 5 is a hash over the authorization username, 6 is random, 8 selects by a priority attribute, 9 is weight based load distribution, 10 is call load distribution, and 11 is relative weight based distribution. Twelve behaviours behind one integer.

OpenSIPS documents the same function, with the same name, taking the same first eight integers. Then it diverges. Algorithm 4 is weighted round-robin rather than plain round-robin, with each destination’s weight deciding how many times it is chosen before the selector moves on. Algorithm 8 is not a priority sort at all; it is the first entry in the set. Algorithms 9 and 10 are a load pattern variable and a named routing block, which are different mechanisms again. Same numbers. Different product.

Read that again. Same parameter name, same integer, different behaviour, in two systems that share an ancestor and most of their documentation. Now ask yourself what the words “round robin” are doing in the comparison spreadsheet your team built from six vendor websites. They are doing nothing. They are a column that looks like data.

It gets better. OpenSIPS publishes an explicit catch-all entry in its own algorithm table: if the algorithm is not implemented, the first entry in the set is chosen. Not an error. Not a refusal to start. The call simply goes to whoever happens to be listed first, on every call, for as long as it takes somebody to notice that one destination’s volume looks strange next to everyone else’s.

Kamailio has its own version of the quiet substitution. Its hash over the authorization username falls back to round robin whenever there is no username available to hash. That is a sensible engineering decision and a genuinely nasty surprise in production, because hashing is sticky and round robin is not. What does sticky buy you? It sends the same caller to the same destination every single time, which is the entire reason anyone picks a hash. Round robin does not. So if you chose the hash because you wanted repeat callers landing on the rep who already knows the account, one missing field takes that away and tells nobody.

The lesson transfers straight into a buying process. Do not accept a strategy name as a specification, because it is not one, it is a marketing word that two engineering teams have already resolved differently. Ask what this vendor’s round robin does when two reps become available in the same second, when a rep is added at eleven in the morning, and when the system restarts overnight. Those three answers tell you whether you are looking at a counter, a timestamp sort, or a hash, and those three behave differently on exactly the days you care about.

How Call Center Routing Software Picks the Agent

Agent selection is where a sales team actually feels the configuration, and FreeSWITCH mod_callcenter is the clearest public example of how that decision gets made. Its source recognises nine strategy values: ring-all, ring-progressively, longest-idle-agent, agent-with-least-talk-time, agent-with-fewest-calls, random, sequentially-by-agent-order, top-down and round-robin. The sample queue that ships with it uses longest-idle-agent. Nine names. Now watch what they have in common.

Here is the part that is on nobody’s feature list. Every one of those strategies builds its candidate list ordered by tier level first, and only then applies the thing the strategy is actually named after. Longest-idle-agent orders by level, then the time the agent’s last call ended, then position. Agent-with-least-talk-time orders by level, then accumulated talk time, then position. Agent-with-fewest-calls orders by level, then calls answered. Random orders by level, then a random value. Sequentially-by-agent-order uses level, position, and the time each agent was last offered a call. Notice the first column.

Level wins. Always. So random is not random across your team, it is random inside a tier, and if every rep ends up at tier level one by accident, which is what happens when a level is not supplied and the system defaults it, you get one flat pool and none of the tiering you drew on the whiteboard. The configuration will look correct in the interface. The behaviour will be a free-for-all.

This is also where the three fairness strategies separate, and the distinction is worth an argument with your sales manager rather than a shrug. Least-talk-time balances minutes. Fewest-calls balances conversations. Longest-idle balances waiting. Those three produce three different teams, because a rep who takes short qualifying calls and disqualifies fast looks heavily occupied to one of them and permanently idle to another, and the rep handling one long negotiation looks like the opposite. Pick the measure that matches the behaviour you actually want rewarded, then check it against a week of dispositions before you defend it in a pipeline review.

Ring-all is the outlier and it deserves its own line. It offers the call to the whole eligible set at once instead of stepping through them one at a time, and the module tracks a separate serving state for it precisely because a call ringing on six devices is a different object to manage than a call ringing on one. Teams reach for ring-all when answer speed is the only number anyone is being measured on. It works, and it is genuinely the fastest thing in the list. It also teaches a floor of six people that somebody else will probably grab it, which is a behaviour you then have to coach back out.

Top-down and round-robin get their own queries in that module rather than a simple sort, and the reason is that both of them have to remember something between calls. Top-down remembers where the last attempt stopped for this particular caller, resumes after that position, and then falls through to the next level. Round-robin looks up whichever agent was most recently offered a call anywhere in the queue and starts after that agent’s position. Memory is the whole feature. So ask the question directly. Does your round robin forget on restart, on redeploy, or at midnight? If it does, it is not round robin. It is top-down wearing a better name.

The Defaults Are Off and the Mistakes Are Silent

Tiers are the standard answer when somebody asks to route by skill. You put the two reps who actually know the enterprise product at level one on that queue, you put the generalists at level two as a safety net, and the queue escalates to the net when level one cannot take the call. Reasonable. Now look at what actually ships.

In mod_callcenter, tier behaviour is governed by four queue parameters: whether tier rules apply at all, how many seconds a caller waits before the next level opens up, whether that wait multiplies by the level number, and whether an empty level escalates at once instead of waiting out the clock. The module’s compiled defaults are tier rules off, a wait of zero seconds, no multiply, and immediate escalation when a level has nobody available. Hold that list.

The sample queue that ships in the same project sets tier rules off, a wait of three hundred seconds, multiply on, and immediate escalation off. Two of those four values are the opposite of the compiled default. Same project. Same version. Two answers.

Neither file is wrong, and that is exactly the point. An implementer who copies the sample and an implementer who relies on the compiled default end up operating two meaningfully different queues while both of them sincerely believe they are running the documented behaviour. And with tier rules switched off, which is the state in both files, your level two reps are not a fallback at all. They sit in the same candidate pool as level one, simply ordered after them, and every strategy above still sorts by level before it sorts by anything else.

Now add the multiply flag, because this is where a sensible-looking pair of settings turns into a dead queue. With multiply on, the wait before each level opens is the level number multiplied by the configured interval, so level two opens at twice the interval and level three at three times it. Set the interval to three hundred seconds, turn multiply on, and your third-tier reps become unreachable for the first fifteen minutes of every call. Nobody intends that. It is two checkboxes and a number that looked reasonable in isolation.

And the strategy string itself fails quietly, in exactly the way the SIP selector does. Give mod_callcenter a strategy name it does not recognise and it does not refuse to start or warn you off; it falls through to ordering candidates by level, then position, then the time each agent was last offered a call. So you asked for least-talk-time, you got a list in configuration order, and nothing anywhere told you. The queue works. The reports look plausible. The distribution is wrong, and it will stay wrong until somebody counts offers per rep by hand.

That is the single most useful question to carry into a vendor call. What does your system do with a routing value it does not understand? If the answer is that it would never happen, the real answer is that it fails silently and nobody has looked.

On Break Does Not Mean Out of the Rotation

Agent presence looks like one setting and is almost always two. In mod_callcenter there is a status, which the agent or an administrator sets by hand, and a state, which the system sets on its own. The status values are Logged Out, Available, Available (On Demand) and On Break. The state values include Waiting, Receiving, In a queue call and Idle. Two fields. One button on the rep’s screen.

So which statuses actually get offered a call? The candidate query that selects agents for a waiting caller admits three: Available, On Break, and Available (On Demand). On Break is in the set.

Read that the way a rep would read it. Going on break does not take you out of the running for the next call; it changes how the offer is handled, not whether you are in the candidate set in the first place, and only Logged Out actually removes you. So what happens when your team believes the break button stops the phone and the system believes break means something much narrower? You get the complaint second-hand, as a performance issue, in a one-to-one. Somebody says reps are ignoring the queue. Somebody else says reps are not following the process. The process was never what anyone in the room thought it was.

The same module exposes four per-agent timers that decide what happens after an offer goes badly, and these are the ones to ask about by name. A maximum no-answer count before the agent is pulled out of rotation. A wrap-up time after a call ends and before the next offer arrives. A delay after the agent actively rejects a call. A separate delay after the agent’s device comes back busy. The shipped sample uses three missed calls, ten seconds of wrap-up, ten seconds after a rejection, and sixty seconds after a busy. Four numbers. Write them down.

Those four numbers are your entire staffing model wearing a disguise, and wrap-up time is the one that does the most damage. Wrap-up is capacity. If a rep needs forty seconds to pick a disposition, type two lines of notes and set the next step, and the queue offers them the next call after ten, you did not buy throughput, you bought a growing pile of calls that were answered and never logged. Time it on a stopwatch during a real call block, including every click the handoff makes the rep do by hand, and set wrap-up from that number rather than from a round one.

What Happens to the Caller Who Waits

Everything so far has been about picking a person. The other half of routing is the caller, and that is the half that quietly produces your abandon rate. It is also the half nobody configures.

A tall frosted glass beaker on a pale violet ground with a slim collar ring resting all the way down at its base, filled to the brim and spilling a smooth glassy sheet over one side into a wide pool on the ground.

The shipped queue config exposes three separate waiting limits, and they are separate for a good reason. A maximum total wait. A maximum wait when there is no agent in the queue at all. And a grace period for that second case, so a caller who arrives during a momentary gap is not dumped the instant they connect. In the sample, the first two are set to zero and the grace period is five seconds. What does zero mean here? Not five seconds, and not a sensible house default.

Zero means no limit. Out of the box, a caller can sit in that queue indefinitely listening to hold music while not one person is logged in to hear it ring. That is a perfectly good default for a toolkit somebody is going to configure and a terrible one for a sales line somebody forgot to. Somebody has to choose the number. If nobody does during implementation, and implementation is usually the week everyone is busiest, the number stays at zero and the feedback arrives months later as a vague complaint that your hold times are terrible.

Two more parameters decide what an abandoned call even means to the system. One sets how long an abandoned record is kept before it is thrown away. The other decides whether a caller who hangs up and rings back resumes their old position in the queue instead of starting from the back. The sample keeps abandoned records for sixty seconds and does not allow resuming. Sixty seconds is short. Most people take longer than that to redial.

That second setting is a real policy decision wearing a toggle. Allowing a resume rewards the caller who gave up and tried again, which is very often the caller you most want to recover and the one with the clearest intent. Disallowing it is simpler to reason about, easier to report on, and fairer to the people who stayed on hold. There is no correct answer here, only the answer you picked, and most teams never found out they were picking one.

So when you evaluate inbound call queues, ask for the three timeouts and the two abandon settings in writing. If a vendor cannot tell you the maximum hold time their system will allow before it does something else, their system does not have one.

Routing on CRM Data Is a Different Problem

Everything above assumes the routing engine already knows something useful about the caller. Usually it knows almost nothing. It has a phone number and whatever the caller pressed on the keypad. That is the whole input.

So “route this caller to their account owner” is not really a routing feature at all. It is a lookup with a routing step stapled to the end, and the lookup is the part that breaks. The number arrives in one format and the record stores another. The contact exists twice because somebody imported a list. The owner field is empty because the lead was never assigned to anyone. The owner left in March and nobody reassigned the book. None of those are routing bugs, every one of them produces a misrouted call, and all of them land in the lap of whoever owns the phone system.

Walk the sequence out loud before you buy anything. The call arrives. The system normalises the number. It queries the CRM. It gets back zero records, one record, or several. It reads a field. It maps that field’s value to a queue or to a named agent. Then it falls back, because on a real list at least one of those steps returns nothing. Six steps. The demo shows you step five.

Then ask the two questions that matter. How long does the system wait on the CRM before it gives up and routes on something else, and what exactly is the something else? A lookup with no timeout is an outage waiting for one slow afternoon on somebody else’s API, and a lookup with no defined fallback quietly dumps your best inbound call into the general queue the first time an integration token expires on a Saturday.

Kixie ships rule-based call routing on this layer and includes it, along with ring groups, call queues and an IVR auto-attendant, on all three plans rather than reserving routing for a higher tier. The call routing page describes automated routing based on CRM data and on caller location, and the trial is seven days with no credit card. That is the useful thing to compare against a vendor who sells routing as an add-on: check whether the routing you are being shown is in the plan you intend to buy.

How to Test Call Center Routing Software in a Demo

Stop asking vendors what their routing can do. Ask what it does when the rule misses. Here is the list, in the order that exposes the most, and it is short enough to read off a phone during the call.

  1. Show me the configuration file or the API payload for this queue, not the builder. Every value, including the ones that are empty.
  2. What is the maximum time a caller can hold before the system does something else, and what is that something else?
  3. What happens when the queue has zero agents logged in? Is that different from zero agents available?
  4. Define your round robin. Does it survive a restart, a new agent added at noon, and midnight?
  5. Does an agent marked on break still get offered calls? Show me the state table.
  6. How long is wrap-up, and does it start when the call ends or when the disposition is saved?
  7. How many unanswered offers before an agent is pulled from the rotation, and who puts them back?
  8. What is the CRM lookup timeout, and where does the call go when the lookup returns nothing?
  9. What does the system do with a routing value it does not recognise? Type a wrong one in front of me.
  10. Which of these settings can I change myself, and which ones require your support team?

That last question is the one buyers skip and then regret about six weeks in. Why does it matter so much? Because routing changes are never scheduled. A rep quits, a territory moves, a campaign floods one number, and the fix is needed that afternoon. A routing model you cannot change yourself on a Tuesday is not your routing model. It is a ticket queue with your calls in it.

Run the test calls yourself as well, and run the bad ones rather than the tidy ones. Call outside business hours. Call with every single agent logged out. Call from a number that sits in the CRM twice under two different owners. Hang up at thirty seconds and call straight back from the same number. A vendor who is relaxed about all four of those has built and tested the fallback path, because those are the calls that break it. A vendor who would rather set up a follow-up session with a solutions engineer has told you the answer already.

What to Measure After the Routing Goes Live

Answer rate is a lagging number and it hides the thing you can actually fix. Break it into transitions. Each one has a different owner and a different fix.

  • Offered to answered, per agent. This is the no-answer count the system is already tracking. A rep well above the team rate is either away from the desk or being offered calls during wrap-up.
  • Queue entry to first offer. If this is more than a second or two, the delay is lookup or eligibility, not people.
  • Offers per call. How many reps had to be rung before somebody answered. Rising numbers mean the eligible set is too small or presence is stale.
  • Abandons by wait bucket. Split under thirty seconds, thirty to ninety, and over ninety. The first bucket is usually an IVR problem, the last is a staffing problem, and they get confused constantly.
  • Distribution across the eligible set. Count offers per rep per day. If one rep’s number is conspicuously high, you are looking at a silent fallback to configuration order.
  • Calls that hit the fallback path. Tag them. If nobody is counting fallback calls, nobody knows the rules stopped matching.

That last metric is the one that pays for itself inside a month. Routing rules almost never break loudly. They stop matching, the fallback quietly catches everything they used to catch, and your dashboard stays a comfortable green the whole time, because the calls are still being answered, just by the wrong people in the wrong order with the wrong context open on their screen.

Review the distribution number weekly for the first month. Then make it somebody’s job by name. A metric nobody owns is a metric nobody reads.

Call Center Routing Software FAQs

Is routing software the same thing as an ACD

Not quite, and the difference shows up in what you are allowed to configure. An automatic call distributor is the piece that holds callers and hands them to agents, so it owns the queue, the selection strategy and the timers. Routing in the wider sense also covers everything that happened before the queue: which number was dialled, what the caller pressed, what the CRM returned, and which queue the call was dropped into as a result. Most products sell both under one name and one price. So which screen owns which decision? Make them show you.

Does skills-based routing need a separate system

Usually not. In the implementations you can actually read, skills are tiers: an agent is attached to a queue at a level and a position, and the level decides who gets tried first. So what is the honest question? Not whether a product supports skills-based routing, because they all claim it, but whether you can turn tier rules on, set the escalation interval, and see the resulting candidate order for a real call. A skills model you cannot inspect is a label on a checkbox.

What is the difference between round robin and least occupied routing

Round robin remembers who was offered the last call and starts after them, so it balances turns. Least occupied balances a measurement, and which measurement matters. Fewest calls balances conversation count, least talk time balances minutes, and longest idle balances waiting. A rep taking short disqualifying calls looks completely different under each one. Choose the measure that matches the behaviour you want, then audit it against dispositions.

How should a sales team test routing rules before going live

Test the failures, not the happy path. Place calls with every rep logged out, with one rep on break, outside business hours, from a number that matches two CRM records, and from a number that matches none at all. Then hang up in the middle of the queue and call straight back from the same handset. Why those six? Each one exercises a different fallback, and the fallback is the thing you are actually buying.

Can an IVR replace routing rules

No. It moves the decision to the caller, which is a different thing. An IVR auto-attendant collects a digit and hands your routing engine a much cleaner input, and that is genuinely useful when callers already know what they want and your menu matches how they think about it. It still deposits the call into a queue, and that queue still needs a strategy, a tier model, three timers and a fallback path. An IVR in front of an unconfigured queue just makes people wait after pressing two.

Sources

Sources verified and content reviewed by the Kixie Research Team on November 11, 2026.

Ready to close more deals with Kixie?

See how Kixie's AI-powered tools can transform your sales and support operations.

Start Free Trial