What is predictive pairing?

Predictive pairing decides which collector works each account, using the floor's own history to predict who will do best with it. It sits alongside the three dialer modes most floors know. Preview, power and predictive decide when a call is made. Predictive pairing decides who is on it.

Where it comes from

Contact centers have been answering "who takes this one?" for fifty years. Each answer used a little more information.

• 1970s: the next free agent. Automatic call distributors sent each call to the agent who had been idle longest. Simple and fair, but blind to who is good at what.
• 1990s: skills routing. Calls went to a group with the right skill: a language, a product, a balance band. That is still the right first cut, and most floors use it today.
• From the mid-2000s: predictive pairing in the enterprise. Afiniti, based in Washington, D.C., built predictive pairing for the largest contact centers in the world, mostly inbound: telecoms, insurers and banks with thousands of agents. Its results were measured against control groups, which showed the idea works at scale.
• Since 2011: predictive pairing on outbound floors. The team behind BestPair has matched outbound agents to leads and accounts since 2011, on insurance, banking and collection campaigns, always measured against a control group of accounts assigned the usual way.
• Now: US collection floors. Collection floors dial out, work agent-owned lists and are measured in dollars collected. BestPair brings predictive pairing to them as a dialer mode.

If your own platform offers predictive routing, it belongs on that timeline too. Same idea, different machine.

The theory

1. Comparative advantage. An idea from economics going back to David Ricardo in 1817: each person should do the work where their advantage is greatest, not only the work they are best at in absolute terms. Every collector is stronger with some kinds of accounts than others. A collector who looks average overall can be the strongest on older accounts, or on payment plan restarts. One leaderboard hides that. Splitting results by account type shows it.
2. The pair is what matters. When the result depends on the match, and not only on each side, sorting well beats ranking. That is Gary Becker's work on matching, from 1973. The question changes from "who is my best collector?" to "who is best for this account?"
3. Assignment. With a prediction for every collector and every account, picking the best set of pairs within each collector's capacity is a classic problem with a known solution: Harold Kuhn's Hungarian method, 1955. That is why the best collectors don't simply get everything.
4. People persuade. Collections is persuasion. In a field experiment at a Chinese consumer lender, 21% fewer borrowers made a verbal promise to repay when the caller was an AI voice bot rather than a person (Choi, Huang, Yang and Zhang, "How Good is AI at Twisting Arms?", NBER working paper w33669, 2025). Who is on the call matters, and different people persuade different people.
5. Measured, not assumed. A prediction can be wrong, so the result is checked against a control group: accounts split at random, some assigned by prediction and some the usual way, over the same weeks.

When predictive pairing works

What it needs:

• History tied to collectors. Calls and outcomes, such as payments, linked to the collector who worked them. Ideally 24 months. No consumer names or contact details are needed.
• A choice of collector. The floor decides who works an account, or which collectors a call can go to: agent-owned lists or agent queues on a preview or manual dialer, segments in the skills-based router on a predictive dialer, and a mix of the two on a progressive or power dialer. A purely inbound floor that sends each call to the next free agent leaves nothing to choose.
• Enough collectors and volume. Usually 25 or more. A bigger team gives more choice for each account and more history for each collector.
• Collectors who differ. If they do about the same on the same kinds of accounts, there is nothing to match. Often nobody knows until the data is read.
• Account details at assignment time. Balance, age of debt, placement date, state and similar fields, present on every account.
• Assignments followed, results fed back. The assigned collector works the account, and outcomes come back regularly so the predictions can be rebuilt.

Where it tends to pay most:

• Low conversion rates. When most contacts don't pay or promise, the conversation decides the outcome, and who has it matters most. When nearly every account pays whoever calls, any collector will do.
• Persuasion-heavy calls. Collections, telesales and retention, where the collector changes the result.
• Many similar decisions every day. High volumes of comparable accounts give both the choice and the history.
• Outcomes that show quickly. Payments and promises that land soon after the call, so results can be measured and the predictions improve.
• Reasonably stable teams. More history per collector. New hires are treated as floor average until their own record builds.

What it isn't

• It doesn't replace the dialer or skills routing. It works inside the group routing already chose.
• It doesn't decide whether an account is worked, what a consumer is offered, or any collection terms.
• It isn't AI talking to consumers. People make every call.