Industries / Finance and business services / Call centers and BPO
Routing
Routing is the work of getting each call, chat or message to the right queue and the right agent, and for outbound work, of running the dialers that place calls within the rules on how they may do it.
7
Milestones achieved, from 1967 to 2021.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1967
Toll-free calling arrives
AT&T introduced 800 service in 1967. These were inbound long-distance lines ("In-WATS") that the business paid for instead of the caller. Mail-order firms took them up first, and by the mid-1970s a call center industry had grown around them, by historian Richard K. Popp's account. Until 1993 each carrier held its own blocks of 800 numbers, so a business that changed carriers had to give up its number; from 1 May 1993 a shared database let numbers move between carriers, and the FCC later found the 800 market "substantially competitive".
LessonThe number customers dial is the real asset, and once it could move between carriers, carriers had to compete to carry it.
Credit
AT&T; US Federal Communications Commission
1973
A switch that queues calls
In 1973 Continental Airlines bought a Rockwell Galaxy automatic call distributor (ACD) for airline reservations. An ACD is a switch that answers incoming calls, holds them in a queue and hands each one to the next free agent. In 1996 Rockwell called itself the inventor of the first digital ACD and called the Continental sale "the beginning of the call center industry"; Continental said Galaxy serial no. 001 had run for 23 years before five networked ACDs replaced it across five reservation centers. The "first" is disputed: Wikipedia says ACDs appeared in the 1950s and that New York Telephone used a large one for hundreds of 411 directory operators in the early 1970s.
Credit
Rockwell; Continental Airlines
2003
Skills routing outruns theory
In skills-based routing, the ACD knows each logged-in agent's skills, such as Spanish or international bookings, and sends each call to a suitable free agent. Computer telephony integration (CTI) links the phone switch to company databases, so the caller's number can pull up their record to choose the route and show it on the agent's screen (the "screen pop"). A 2003 review found that "the majority of advanced ACDs" offered skills-based routing with scarcely any guidance on how to use it, and that "the technology has raced ahead of managers' and academics' understanding".
LessonRouting rules can be set up much faster than anyone can show which rules work best.
Credit
ACD and CTI makers (the review names no single inventor); review by Noah Gans (Wharton), Ger Koole (Vrije Universiteit Amsterdam) and Avishai Mandelbaum (Technion)
2003
A cap on abandoned calls
A predictive dialer calls more numbers than there are free agents, betting that many people will not answer; when too many do, the person who picks up hears silence ("dead air") or a hang-up. The US Federal Trade Commission's amended Telemarketing Sales Rule, published on 29 January 2003 with the National Do Not Call Registry, counts a call as abandoned if no salesperson comes on within 2 seconds of the person's greeting, and gives sellers a safe harbor if they abandon no more than 3% of calls answered by a person and play a recorded message naming the seller. The FTC chose 3% between the direct marketers' 5% guideline and the 0% that consumer groups wanted; today's text measures the rate over each 30-day period and requires letting an unanswered phone ring at least 15 seconds or four rings.
LessonDialing ahead of free agents trades agents' idle time for callers' silence, and the regulator settled that trade with a number.
Credit
Federal Trade Commission
2010
Silent calls from guessing machines
Britain's communications regulator, Ofcom, treats silent and abandoned calls as "persistent misuse" of the phone network and limits abandoned calls to 3% of live calls per campaign or per call center over 24 hours, each with a message saying who called. In its 2010 revision Ofcom said over 22% of people had received such a call in the past six months, and over 70% of those who got silent calls had received two or more in one day from the same company. It blamed most repeat silent calls on answering-machine detection, software that guesses whether a person or a machine has answered and wrongly takes people for machines, and ruled that a number flagged as a machine may be called again only with an agent on the line; Parliament raised the maximum penalty from £50,000 to £2 million.
LessonA detector that misjudges a person once tends to misjudge them again, so a low overall error rate can still mean one household gets silent calls day after day.
Credit
Ofcom
2017
Routing moves to the cloud
Amazon Web Services launched Amazon Connect on 28 March 2017, a cloud contact center service whose queues, routing and call flows run on Amazon's servers; Amazon said it was "based on the same contact center technology used by Amazon customer service associates". South Carolina's Department of Social Services moved its benefits call centers onto it after dropped calls, holds of up to an hour and server outages, and after Hurricane Helene handled more than 100,000 disaster food-aid calls with more than 500 agents, Route Fifty reported in June 2025.
LessonWhen routing runs as a rented service, a center meets a surge by adding agents, not by buying a switch.
Credit
Amazon Web Services; South Carolina Department of Social Services
2021
Signed caller ID
From 30 June 2021 the FCC required the largest US voice providers to use STIR/SHAKEN, shared standards that pass checked caller-number information from network to network so a faked ("spoofed") number can be flagged; smaller providers got until 30 June 2023. The FCC warned that a verified number "does not mean the call itself is legitimate", and the analytics carriers use to block suspected robocalls also block lawful calls. In February 2025 it required a standard "Network Blocked" signal (SIP code 603+) within 12 months so callers learn when and by whom they were blocked, after the call center software firm TCN told it that each wrongly blocked case took 10 to 20 hours to diagnose.
LessonProving who is calling did not prove the call was wanted, and legitimate outbound centers now have to watch how carriers treat their calls.
Credit
US Federal Communications Commission; ATIS and SIP Forum
Watch for
Pairing callers by predicted outcome
Afiniti's Pairing product predicts the outcome of every possible caller-agent match from past interactions and routes each call to the best one. Afiniti says it "has delivered more than $2.5 billion in measurable value to clients", measured by switching the AI on and off at intervals through the day, and in October 2025 said Five9 customers would get access to it. This research found no independent published evaluation.
Achieved whensomeone outside Afiniti, such as a client in an audited filing, a regulator or researchers given the on/off logs, publishes the measured gain and how it was calculated.
Claimed by
Afiniti, January 2026
AI agents that hand off
NICE says its Cognigy AI agent can answer inbound calls inside Microsoft Teams Phone, handle routine requests and hand callers "to the right person, queue, or department with full context", and that this "can reduce blind transfers, hold times, and repeated explanations". It gave no figures.
Achieved whena named customer or an independent analyst publishes before-and-after transfer rates and hold times for calls routed this way.
Claimed by
NICE, November 2026
Learning routers on replayed calls
Sanne van Kempen, Jaron Sanders, Fiona Sloothaak and Maarten G. Wolf replayed a month of calls from a large US bank's call center (public Technion SEELab data) through a simulator and report that their learning router, UCB-QR, earned at least 98% of the best possible payoff on every day, while fixed rules fell as low as 60% on some days. The payoff was a stand-in (calls that ended without a transfer or conference), because the data has no cost or satisfaction measures.
Achieved whenthe router runs in a live center, or another team reproduces the gain with a real outcome measure.
Claimed by
van Kempen and co-authors, June 2025
Open
01
A bake-off for routing rules
A public benchmark that replays real call center arrivals through a simulator and scores several routing rules on the same calls: first come, first served to the longest-idle agent; fixed skill priorities; a speed-weighted priority rule; UCB-QR; and an open version of outcome-based pairing (predict the result of each caller-agent match, then pick the best). Add synthetic call sets with a known "true" caller-agent fit built in, so each rule's share of the possible gain can be measured against a known answer.
The briefCredit · unclaimed02
A checker for dialer logs
An open tool that reads a dialer's call log, one row per dial (campaign, time, ring time, whether a person or a machine answered, the answering-machine guess, seconds until an agent joined, whether a message played), and reports clause by clause whether each campaign met the FTC safe harbor (3% of person-answered calls per 30-day period, a 2-second connection, at least 15 seconds or four rings, a recorded message) and Ofcom's limit (3% of live calls per campaign or per center in 24 hours, with a message). It also lists any number that got repeat silent calls, and ships with a library of synthetic logs with known answers.
The briefCredit · unclaimed
Open 01 · the brief
A bake-off for routing rules
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public benchmark that replays real call center arrivals through a simulator and scores several routing rules on the same calls: first come, first served to the longest-idle agent; fixed skill priorities; a speed-weighted priority rule; UCB-QR; and an open version of outcome-based pairing (predict the result of each caller-agent match, then pick the best). Add synthetic call sets with a known "true" caller-agent fit built in, so each rule's share of the possible gain can be measured against a known answer.
Why it is still open
The nearest work, van Kempen and co-authors (2025), tests one learning router against a few fixed rules on the Technion US bank data with a stand-in payoff, and does not test outcome-based pairing. Afiniti's $2.5 billion figure rests on its own on/off tests, and this research found no independent comparison. In 2003 Gans, Koole and Mandelbaum wrote that the technology had "raced ahead" of anyone's understanding of how to use it.
Built with what exists
Technion SEELab's USBank (about 1,000 agents) and AnonymousBank call logs, free for academic use after notifying the lab; the UCBQR simulator on GitHub (it lists no license, so ask the authors before reusing it); Python with SimPy or Ciw for event-by-event queue simulation.
First step
Ask the SEELab for USBank access, and rerun the UCBQR repository to reproduce the paper's results.
Done looks like
A public repository that runs at least five rules on at least one real month of calls and two synthetic sets with a built-in fit, reporting mean wait, abandonment rate, the share of calls ended without a transfer and, on the synthetic sets, the percentage of the built-in gain captured. A second person reruns it from the README and gets the same figures to within 1%.
Open 02 · the brief
A checker for dialer logs
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open tool that reads a dialer's call log, one row per dial (campaign, time, ring time, whether a person or a machine answered, the answering-machine guess, seconds until an agent joined, whether a message played), and reports clause by clause whether each campaign met the FTC safe harbor (3% of person-answered calls per 30-day period, a 2-second connection, at least 15 seconds or four rings, a recorded message) and Ofcom's limit (3% of live calls per campaign or per center in 24 hours, with a message). It also lists any number that got repeat silent calls, and ships with a library of synthetic logs with known answers.
Why it is still open
Dialers check themselves. The open-source VICIdial dialer, for example, tracks its own abandon rate and adjusts its pace to stay under the US 3% limit, but that check runs inside one system, against US rules only, and no shared test set shows that it counts correctly. Regulators work from estimates after the fact: Ofcom's 2014 fines rested on an estimated 30,296 abandoned calls by one firm and 12,703 silent calls by another.
Built with what exists
Python; synthetic CSV logs with made-up numbers (no real calls or people); 16 CFR 310.4 and Ofcom's persistent-misuse policy as the written specification.
First step
Write both rule sets as a one-page table that maps each clause to the log fields it needs, then generate the first ten synthetic logs.
Done looks like
40 synthetic campaign logs, 20 compliant and 20 each carrying one planted breach (a rate over 3% in one window, a slow connection, a short ring, a missing message, repeat silent calls to one number). The tool flags all 20 breaches, flags nothing in the 20 clean logs, and its abandonment rates match a hand calculation on five logs.