Multi-armed bandit is an efficient way to run experiments in collections

Recommendation systems have spent years answering a question for streaming platforms: what should show up for this person, right now, based on everything learned so far? Communications for collections is a similar problem: less choice for entertainment, higher stakes, and a lot more regulation. The central question is still the same. Is the business testing at the speed the market demands, or is it still doing experiments as it did five years ago?

Collections teams love experimentation

Most collection and payment recovery programs already conduct various experiments, such as using a new subject line, sending the message at a different time, employing an alternative call script, offering a settlement in a single payment rather than in three installments, etc. These kinds of collections messaging experiments are starting to become more prevalent

In the conventional 'champion and challenger' method, the entire volume is split equally among a small number of variants, and a statistically significant sample size is obtained before declaring a winner. This method works in situations where time is not an issue, such as agricultural trials, but falls short when it is necessary to act on the results in real time.

Collections communication goes beyond simply conducting a single message test, or a handful of tests, as it involves a wide variety of message types, different sending times, a range of calls to action, multiple offer structures, and a variety of channels; the number of possible combinations is therefore so large that no team could possibly run A/B tests to test every possible combination. For example, a payment recovery team might run five or six of these tests simultaneously across different segments, and if every important combination were evaluated in the traditional champion-and-challenger manner, it would take years to reach a conclusion. By the time an answer materialized, market conditions would have changed, and it would need to be retested. This issue arises from the methodology, which is why so many programs end up choosing one approach and sticking to it for a long time.

Rising delinquency rates require faster and more frequent experimentation

The rate of subprime auto loan delinquencies after 60 days reached a record of 6.90% in January 2026, the highest figure ever recorded since the early 1990s. The trend extends to other sectors as well: in Telecom, rising delinquency is compounded by increased competitive pressures

One-size-fits-all messaging schedules no longer drive results. Sending identical messages to all accounts regardless of customer behavior or response. For telcos especially, late-payment communications represent a critical moment for building customer loyalty, making it essential to personalize these interactions and respond quickly.

A multi-armed bandit test runs experiments continuously

The multi-armed bandit is a classic problem in probability. Imagine a row of slot machines. Each one has a different chance of paying out, but you don't know which is best. Do you keep playing one machine or spread your bets to learn more? KredosAi takes the same approach, spreading the bets to quickly learn which messages deliver the biggest payoff.

Netflix employs the same basic method when deciding which thumbnail to display to a particular viewer, or when a social platform decides which content to present next. Many people have already experienced a bandit-style system even if they never referred to it as such. The system learns from user behavior as it gathers more data and continually refines its recommendations, while maintaining a balance between “earning” and “learning”.

Rather than dividing the volume equally and then waiting for the results, a bandit immediately begins shifting the volume towards the options that appear strongest as soon as results start coming in, while preserving some degree of exploration. This exploration is important because the current leader might still not be the best, and some options may remain uncertain even if the early results are promising. As more data becomes available, the model becomes more confident and efficient than a fixed one

A bandit system has two main components: a model that estimates which message is most likely to perform well, and a method for distributing volume among the winning options and the uncertain ones. Although Thompson sampling is generally regarded as the best method for this allocation, it requires careful engineering to implement it properly.

In reality, it's a matter of the weights assigned to an account as it progresses through its delinquency cycle, the appropriate tone to use at any given moment, which channel the customer actually responds to, and what call to action is suitable given all these factors. These are exactly the same variables that a champion-and-challenger test would look at. The difference is that a bandit runs the experiment across all of them simultaneously and continuously, shifting effort from account to account, rather than applying the single best setting across the whole portfolio once the test has ended.

Multi-armed bandit theory is well established, but running it at scale is recent

Although the theoretical idea of a multi-armed bandit goes back several decades, it is the ability to carry it out at scale that has changed: the amount of computing power and data required to make it feasible became available only recently, in a way that is similar to machine learning, which had existed in theory for a long time before becoming usable on a large scale. Net net, the idea itself has been around for a while, but running it effectively on live data volumes at scale is a more recent development.

Dimensions of experimentation

When it comes to messaging for collections, there are several different dimensions of experimentation.

First, there is message tone, where you can test whether an empathetic tone results in the right outcome versus a direct approach. In addition, there are several layers of behavioral science that can be embedded into the tone. For example, does gain framing work better or does loss framing work better in a particular context

Second, there is the channel component. Some people respond to SMS better, and some respond to Email better. In addition, there are the more innovative channels such as Rich Communication Services (RCS) and WhatsApp, which allow for a layer of interactivity that doesn’t exist in traditional channels

Third, there is the calls to action component – what offers to present to which customers. Presenting the wrong offer has significant economic impact. For example, you don’t want to offer a discount to a customer who would have purchased without a discount

Finally, there is the timing element – what time of day to send the message. Some people respond better to messages sent early in the morning, and some respond better to messages sent later in the day. There is no one-size-fits-all approach that works universally

A multi-armed bandit approach allows for experimentation across all of these variables at the same time. It evaluates tone, channel, calls to action, and timing as a single continuous experiment, dynamically shifting traffic toward the combinations that work best for specific account segments. As delinquent accounts progress, the bandit adapts in real time. When testing ambitions outgrow fixed methodologies, bandit algorithms provide the scale needed to keep moving forward

Deterministic guardrails keep this experimentation compliant

The message sequencing and template logic on a bandit-driven platform must be deterministic to prevent the system from sending a message at the wrong time (for example, sending a near-suspension message to someone early in the delinquency cycle). Each template must be reviewed and approved by the customer's legal and compliance team, so no message reaches a customer unless it has first received that approval. In addition, appropriate guardrails must be automatically enabled in the platform to ensure hard constraints are respected. Any AI-driven experimentation debt collection platform needs to incorporate these regulatory guardrails as table stakes.

Continuous multi-armed bandit testing is what produces a return at enterprise scale

KredosAi's own results, drawn from more than 200 million customer engagements, show a return on investment of 20x or more at Fortune 50 enterprises, driven by write-off reduction and improved customer lifetime value across the portfolio. These results come from running many MAB-driven tests simultaneously and continuously, rather than running one test at a time with long waits between results.

The general concept is closely linked to the idea elsewhere referred to as "pay and stay": recovering payment is the simpler part of the task, whereas retaining the customer is what ensures long-term revenue.

FAQ

What is a multi-armed bandit?

A multi-armed bandit is a decision-making procedure that continuously determines which option to favor among a set of choices by relying on actual results rather than following a fixed testing schedule. When it comes to email marketing, this means determining which message, subject line, or timing option is most effective for a particular account, then allocating more resources to that option as more evidence becomes available.

Is multi-armed bandit testing the same thing as collections experimentation overall?

On the contrary, most collections and payment recovery programs currently run a variety of experiments, examining different versions of messages, channels, timing, and offer structures, typically using a champion-and-challenger approach. The multi-armed bandit is just one particular method of conducting these types of tests, as it differs in that it continuously varies the volume rather than waiting for a predefined test period to end.

Multi-armed bandit vs AB testing: what's the difference?

In traditional A/B testing, or champion-challenger testing, traffic is split equally, and nothing is changed until a conclusive result is reached. With a bandit approach, however, the allocation is adjusted continuously as new data comes in, meaning fewer accounts receive a poorly performing message while the test is still ongoing.

Does champion and challenger testing still have a role once a team uses a multi-armed bandit?

Certainly. Teams usually retain both tests when it is necessary to have a clear before-and-after record, for example, with a brand-new message template that is still awaiting compliance sign-off, while employing a bandit algorithm to optimize allocation among options that have already been approved.

Is multi-armed bandit technology new?

The theory first appeared in the 1900s, but what is new now is the practical ability to apply it to live data, just as machine learning existed in theory long before today's computers made it possible to use it on a large scale.

Does this replace human assessment in collections?

No. The system determines which message or channel is likely to work. People still determine risk tolerance and recognize when an account needs a person rather than an automated message.

How quickly can a business see results from this kind of testing?

Meaningful results generally emerge within weeks rather than the months or years required by a traditional fixed test, since the system starts learning from the first responses rather than waiting for a predetermined sample size.

What happens when account behavior changes partway through testing, such as due to an economic shift or a regulatory update?

A bandit adjusts as new data comes in rather than locking in a result from a fixed test window. When customer behavior shifts, the system starts favoring different messages accordingly, instead of continuing an approach that no longer works.