What is next best action in collections, and when is a collector’s call worth making?

The short answer

A propensity-to-pay score tells you who is likely to pay. It does not tell you who will pay *because* you acted. Next best action (NBA) closes that gap. For every account, every day, it asks which action is worth most: a letter, an email, a text, a payment link, an AI voice call, a collector’s call, or nothing.

Score the action, not the consumer. An action is worth the extra payment it buys over the cheaper option, minus what it costs, including the contact it uses up. Cheap channels deliver information and clear the accounts that only needed a reminder. A collector’s call is the expensive, rationed action, so it goes to the accounts a conversation will actually move. And because collectors differ, who takes that call changes what it is worth.

The actions on the menu

Every collection floor has the same menu. What differs is what each action is for, what it costs, and which rules ration it.

The gap between the cheapest and the dearest action is enormous. Prodigal estimates a seat on a well-run predictive floor delivers about 2 to 3 hours of productive talk time a day, which puts an onshore productive conversation hour above $110 (a vendor estimate). Against $0.10 for a thousand emails, a collector’s conversation is the scarcest thing an agency owns. Collectors are also the largest cost line: US bill and account collectors earned a median $47,030 a year in May 2025, and BLS projects employment to fall 10% over the following decade.

Cheap is not the same as a good return. The unit cost tells you what an action costs, not what it earns. An action only pays if the consumer acts on it, and an action that turns the consumer off can cost more than it saves, because it lowers the value of every contact that follows. Next best action compares actions on what they bring in after cost, never on cost alone.

Propensity ranks who pays, not who an action moves

A propensity-to-pay or collection score ranks accounts by their chance of paying within a set window. It is a good place to start and a poor place to stop, because every one of these scores asks "did the account pay?" and never "did it pay because of what we did?"

Marketing learned this the expensive way. At US Bank, targeting the top 30% of customers by a conventional response model produced almost no incremental sales and a negative return: the model "succeeded only in targeting people who would have bought anyway". A model built to predict the *change* caused by the offer did far better on the same budget (Grundhoefer 2009, reported by Radcliffe and Surry, 2011). Peer-reviewed field work on retention found the same: offers aimed at the customers most likely to leave did less than offers aimed at the customers the offer would most change (Ascarza, Journal of Marketing Research).

Collections has a clean example too. Tala, a lender in emerging markets, randomised calling against not calling its late borrowers, then ranked borrowers by how much a call was predicted to raise their chance of repaying. "Over half of the incremental revenue available by calling all borrowers can be obtained by calling only the top 10%," it reported, and after the average cost of outreach, "only the top 50% of borrowers are profitable to call". In production it kept 5% of borrowers on a random-call group and 5% on a never-call group to keep checking (Tala, 2021).

Vendors’ own figures show why propensity still earns its place at the start. Experian says 80% of recovered dollars are captured in the top 40% of accounts worked (Experian Recovery Scores, a vendor figure). That tells you where the money is. It does not tell you how much of it would have come in without the call.

The four kinds of account

Uplift modelling (also called incremental, true-lift or persuasion modelling) sorts every account into one of four groups by how it responds to an action. In collections language:

A propensity score puts sure things at the top of the list, because they are the likeliest payers. That is exactly where a call is worth least. Radcliffe and Surry warn that when the outcome and the effect of an action pull in different directions, "a conventional model will tend to rank the worst group by uplift as the best targets". The persuadables, the accounts a call actually moves, often sit in the middle of a propensity-sorted list.

Digital delivers information; people persuade

The case for digital first is strong where the problem is attention. In Uganda, text reminders produced "a 7-9% increase in the probability of paying on time" (Cadena and Schoar, NBER, 2011). In the Philippines, generic reminders did not help on average, but messages carrying the loan officer’s name cut loans unpaid 30 days after maturity by 5.5 percentage points (Karlan, Morten and Zinman, J-PAL). Even in a text, the relationship carries weight.

Defaulted debt is different. In three field experiments with 32,000 Latvian borrowers holding non-performing loans, "simply nudging defaulted individuals does not work", and reminders after a promise to pay could trigger a repeat default (Saulitis, 2023). Third-party placements lean towards defaulted debt. So in an agency, digital first is best treated as information and triage (the validation notice, the payment link, self-service), not as a replacement for persuasion.

Consumer preferences pull both ways. A McKinsey survey of recently delinquent North American card customers found about 90% prefer digital contact, while an older, traditional-preference segment did 17% better with phone and letter (McKinsey, 2019). The CFPB’s national survey found letters the most preferred method, at 46%, and 63% of consumers contacted about a debt said they were contacted too often (CFPB, 2017). Digital-first vendors report high self-service shares: TrueAccord says 98% of the delinquent consumers it services resolved with no human interaction (TrueAccord, a vendor figure that counts payers it serviced, not all accounts placed).

When the account needs persuading, the person on the call matters. In the largest published study of AI against human collectors, 21 percentage points fewer borrowers promised to repay on the first call they answered when the caller was an AI voice bot, and 18 percentage points fewer repaid on the same day as the call (27.3% against 44.8%) (Choi, Huang, Yang and Zhang, NBER w33669, 2025; cases from China in 2021 and 2022, before today’s large language model voice agents). The authors found AI less cost-effective than human callers except on smaller debts, where it matched them. The damage also outlasted the call: borrowers were less likely to keep a promise made to the AI, and borrowers first contacted by the AI had still repaid less a year later, even though human collectors took over from day 6. The cheap first touch made the expensive later ones less effective. A separate study found borrowers who spoke with a randomly assigned bank agent were more likely to resolve the delinquency, and that the likeability of the agent’s voice affected payment (Laudenbach and Siegel, Journal of Finance, 2025). Will AI voice bots replace human collectors? covers that study in more depth.

Persuasion here means clear information, a workable plan and a promise the consumer can keep. It never means pressure, false urgency or a payment the consumer cannot afford. What the persuasion models say about collection calls sets out how that looks on the phone.

No channel is free

*General information, not legal advice. Ask your own counsel how these rules apply to your agency.*

US rules turn every account’s contacts into a budget, and the budget is not only for calls.

• Regulation F. A collector is presumed to comply if it calls a person about a particular debt no more than seven times in seven consecutive days, and not within seven days after a telephone conversation about that debt. Unanswered calls and ringless voicemail count. Texts and emails do not count toward the presumption, but they remain under the general harassment standard (CFPB, 12 CFR 1006.14). Every email and text must include a simple way to opt out, and the 8 a.m. to 9 p.m. window applies to messages by send time (12 CFR 1006.6).
• Texts. The safe harbour is the hardest to meet: the creditor’s consent does not carry over, so the consumer must have texted the collector from that number or consented directly, and the number must be checked for reassignment (Ballard Spahr).
• AI voice. The FCC holds that AI-generated voices are artificial voice under the TCPA (Mayer Brown), so AI voice calls to cell phones need prior express consent, and TCPA damages run $500 a call, up to $1,500 if wilful (NCLC).
• States and cities. Massachusetts limits calls to two in seven days per debt (Hudson Cook), and its Attorney General’s rule counts texts and recorded messages (Alston and Bird). Washington allows no more than three calls or writings a week (Washington AG). From 1 January 2027, New York City allows three attempts per account per seven days across phone, text, email and social media together (Troutman; Orrick). In New York City an email spends one of three weekly slots.

Two things follow for next best action. First, a right-party conversation locks the phone on that debt for a week, so the first conversation is the most expensive one an account gets. It should go to the right collector at the right moment. Second, contact counts and opt-outs need one record across every channel and vendor. In 2026 a federal court rejected a collector’s bona fide error defence after it kept texting 4,985 numbers that had replied "STOP", because it had no procedures to make sure its texting vendor honoured opt-outs (Orrick, 2026). How to keep collections up inside contact limits goes further on working under a cap.

Five steps to the next best action

The method below is our framework, built from standard parts: eligibility rules before scoring, value net of cost, and work allocated under capacity, as decision engines such as Pega’s and published work on collections optimisation describe. Each step narrows the menu for one account on one day.

1. Remove every action the account may not receive today. Cease-communication, disputes, bankruptcy, attorney representation, channels the consumer has opted out of, missing consent for texts or AI voice, the time-of-day window at the consumer’s location, work email, and state, city and client rules. Compliance is a gate, not a weight in the score.

2. Check the contact budget. For each action left, count the slots remaining: Regulation F calls left in the rolling seven days on that debt, any seven-day pause after a conversation, and state or city limits. A slot used today is a slot not available tomorrow, so it has a cost even when the call itself is cheap.

3. Value each action that survives.

> Value of an action = (chance of payment if we take it, minus the chance with the next cheaper option or no action) × dollars expected if they pay × the agency’s share, minus the cost of the action, minus the cost of the slot it uses.

The first term is the uplift, not the propensity. A sure thing has a high chance of paying but almost no difference between a call and an email, so its call is worth little. A persuadable may have a modest chance of paying, but a call moves it a lot, so its call is worth a lot. This is where the need for persuasion enters the sums: it is the extra a conversation adds over the cheaper channel. Count the effect on later contacts too. An action that annoys the consumer, or teaches them to ignore you, lowers the value of every action after it, and that belongs in its cost.

4. Set the call threshold against collector capacity. Rank every call-eligible account by call value per expected collector minute, and fill the day’s capacity at an occupancy your collectors can sustain. Two thresholds apply, and the higher one wins:

• Break-even: the call must be worth more than it costs. Below this a call loses money, even if a collector is sitting idle.
• Capacity: the value of the last account that fits in the day. When there are more worthwhile calls than collector hours, the bar rises above break-even. When capacity is slack, it falls towards it.

"Keep collectors busy" means letting the threshold flex down to break-even, not dialling below it. Persistent idle time at break-even is a staffing signal, not a reason to dial accounts that cannot pay back the call.

5. Choose who takes the call. The uplift from a call depends on who is on the line. If each connected call goes to the collector who has done best with accounts like it, every call is worth more, the break-even bar falls, and more accounts become worth a human conversation. More on this below.

A worked example

Four accounts on the same morning. The numbers are made up to show how the method behaves, not drawn from any real portfolio. Each account has a 30-day chance of paying after an email and after a collector’s call. The agency keeps 30% of what it collects, and a call attempt costs $4 in collector time. The slot cost is left out to keep the sums short.

A dialer sorted by chance of paying calls A first and D second, the two accounts where a call is worth least or does harm. Sorted by the value of the call, only B is worth a collector’s time. A gets the email and a payment link, C gets the minimum compliant effort, and D is left alone.

Now step 5. Suppose B’s call goes to the collector who has done best with accounts like B, and the chance after the call is 17% rather than 14%. The uplift becomes 11 points, and the call is worth $33 minus $4, or $29. The same lift applied across the floor pulls accounts that sat just under break-even above it.

Start with a score, then learn on your own floor

No agency has uplift estimates on day one, and it does not need them to start. Next best action matures in three levels, and each one uses the same sums with better inputs.

Why level 1 is only a start. A vendor score is read from the consumer’s credit or public record file on the day you ask for it, and your calls never reach it. After two broken promises the account has gone cold, and the score still says what it said on day one. How to get more from the propensity-to-pay score you pay for walks through that with a calculator, turning the vendor’s number into next contact value, which moves after every contact. That is level 2 in practice.

Scores also go stale as a whole. On a 2,000,000-card dataset, a score’s ability to separate good from bad accounts fell by more than half in two years (Jung, Thomas and So, Edinburgh Credit Research Centre). Experian’s own modellers found that adding the holder’s account data to bureau data raised the Gini from 70 to 76 on unsecured bank lending (Deuchars, Experian, 2009), because generic scores "cannot capture client-specific account performance or the effectiveness of each client’s own collections department".

Why level 3 needs random groups. History is biased by who you chose to work. One UK practitioner paper describes a self-cure strategy that hit about 75% at launch and decayed within 12 months, once an upstream strategy removed the same easy accounts (Matthews and others, 2011). Only a group that did not get the action shows what the action changed. Experian’s classic champion/challenger design runs a challenger on a random share of accounts for four to six months (Experian), and practitioner advice is to let a scorecard run a full collection cycle and reweight after about 180 days (Collection Advisor).

What deployed systems have reported

Results from the field, each as its source describes it:

• New York State tax collections put about 300 legal and business rules in front of an optimiser that allocated staff time across actions, and reported delinquent collections up $83 million (8%) "using the same set of resources" (Interfaces, 2012; government tax collection, compared year on year).
• Česká spořitelna, a Czech bank, let an optimiser set contact timing and channel by segment and reported 25% fewer collector call minutes with stable roll rates and the same cash collected (Business Wire, 2022; a vendor announcement).
• A Dutch collection agency field experiment found algorithmic calling decisions achieved "23.4% higher repayment rates compared with human collection officers", with the same agents dialling (Zhou, SSRN, 2024; a working paper).
• Tala, above: half the incremental revenue from calling the top 10% by predicted uplift, and only the top 50% worth calling.

We have not found a published uplift study using US third-party agency data. The parts are well evidenced: human calls move undecided consumers, collectors differ, and ranking by uplift beats ranking by propensity. The combined effect on a US agency floor is something each agency has to measure on its own accounts.

Who takes the call changes what the call is worth

Most next best action systems treat "a collector’s call" as one action with one value. In practice its value depends on who makes it. Agents in US financial services call centres "exhibit very different performance for the same call type" (Mehrotra and others, M&SOM, 2012), the likeability of the voice on the line affects payment (Laudenbach and Siegel), and on collection floors a collector who does well on one kind of account can trail on another. Why collectors convert at different rates explains the gap.

That has three effects on next best action:

• At the same capacity, every collector hour brings in more.
• The break-even bar falls, so more accounts become worth a human call.
• Because a conversation pauses calls on that debt for seven days, getting the first conversation right is worth more than the raw call count suggests.

Predictive dialing decided when to call. Predictive pairing decides who takes the call. BestPair sends each account to the eligible collector who has done best with accounts like it, from the floor’s own call and payment history, inside the rules and workload limits the agency sets, and measures the result against a control group. It does not decide which accounts are called, which channel is used or what a consumer is offered: the agency’s scores, NBA rules and call order stay as they are. How agent matching differs from a propensity score sets out how the two fit together.

In a live test at a North American collection agency, a telecom book of 170,000 accounts was split 50/50 at random before the first call. The matched half collected $98.76 per account reached against $79.17 on the usual process, 24.7% more (the case study). That is one book in one month, not a forecast or a guarantee for any other floor. It is the size of the step 5 effect on that floor.

How to prove it on your own floor

The same discipline that makes next best action honest makes any vendor claim checkable, ours included: a random group of accounts that keeps the current treatment, in the same weeks, compared per account. How to prove a routing change worked shows how to set one up.

For pairing, BestPair’s Floor Assessment works from up to 24 months of your own history for a fixed $15,000 and changes nothing on your floor. It shows how far your collectors differ on the same kinds of account, and an estimated range for what matching might have added. If the report supports it, live pairing splits accounts at random: 80% allocated by BestPair, 20% left on your current routing as the control group, and the fee is 10% of the extra collected against that control, with a $2,500 monthly minimum. Any gain is measured, not promised.

Sources

Scores, uplift and next best action - Real-World Uplift Modelling with Significance-Based Uplift Trees: Radcliffe and Surry, 2011, including the US Bank example (Grundhoefer 2009). - Retention futility: American Marketing Association on Ascarza, Journal of Marketing Research, 2018. - Helping late borrowers repay with uplift modeling: Tala engineering blog, January 2021; a lender outside the US, self-reported. - Recovery Scores product sheet: Experian; a vendor’s description. - Scoring Models for Collections and Debt Recovery: Deuchars, Experian, Credit Scoring and Credit Control XI, 2009; UK data. - When to Rebuild or When to Recalibrate Scorecards: Jung, Thomas and So, Edinburgh Credit Research Centre. - Effectively Deploying Analytics to Support Collections: Connolly, Matthews and Pinch, Credit Scoring and Credit Control XII, 2011; UK. - Champion/challenger white paper: Experian, 2008. - Building a basic collections scorecard: Collection Advisor, 2014. - Tax Collections Optimization for New York State: Interfaces 42(1), 2012. - Česká spořitelna uses FICO prescriptive analytics: Business Wire, March 2022; a vendor announcement. - Algorithmic Decisions in Debt Collection: Zhou, SSRN working paper, 2024, revised 2025.

Channels, persuasion and the person on the call - How Good is AI at Twisting Arms? Experiments in Debt Collection: Choi, Huang, Yang and Zhang, NBER working paper w33669, 2025. - Personal Communication in an Automated World: Laudenbach and Siegel, SAFE working paper 428; Journal of Finance, 2025. - Remembering to Pay?: Cadena and Schoar, NBER working paper 17020, 2011; Uganda. - Text message loan repayment reminders in the Philippines: Karlan, Morten and Zinman, J-PAL. - Nudging debtors with non-performing loans: Saulitis, Journal of Behavioral and Experimental Finance, 2023; Latvia. - The customer mandate to digitize collections strategies: McKinsey, 2019. - Debt Collection Survey Report: CFPB, January 2017. - Consumers prefer self-serve options: TrueAccord, October 2024; a vendor’s figure. - Routing to Manage Resolution and Waiting Time in Call Centers with Heterogeneous Servers: Mehrotra, Ross, Ryder and Zhou, Manufacturing and Service Operations Management, 2012.

Costs - Bill and Account Collectors: US Bureau of Labor Statistics, Occupational Outlook Handbook. - The true cost of debt collection: Prodigal, 2026; vendor estimates. - Amazon SES pricing, Twilio pricing breakdown (textbee), Lob pricing and AI voice agent pricing (Ringly): list prices, 2026.

Rules - Regulation F, 12 CFR 1006.14 and 12 CFR 1006.6: CFPB. - Using email and text messages under the final rule: Ballard Spahr, November 2020. - FCC ruling on AI-generated calls: Mayer Brown, February 2024. - Limits on prerecorded debt collection calls: National Consumer Law Center. - Massachusetts Division of Banks revised regulation: Hudson Cook, 2025; What counts as an attempt to contact: Alston and Bird, 2019. - Collection agencies: Washington Attorney General. - New York City finalizes expanded debt collection rules: Troutman, March 2026; compliance delayed to January 2027: Orrick, August 2026. - Texts after an opt-out violate the FDCPA: Orrick, April 2026.