The right payment reminder depends on who reads it

Picture two customers with the same overdue phone bill. Maria joined four months ago and has just missed her first payment. James has been a customer for nine years and missed one payment after a busy month. They can both get the same reminder. In practice, the message that works for Maria is often not the one that works for James.

Two customers, two reminders (illustrative examples)

To Maria: “Hi Maria, your $84 bill is 5 days past due. Pay in two taps, or split it into two smaller payments.”

To James: “Hi James, your payment didn’t come through this month. Thanks for being with us since 2017. You can pay in two taps here.”

Both messages are approved and compliant. Which one gets paid depends on who reads it. KredosAi’s latest model update learns exactly that.

How the Intelligence engine got smarter, step by step

KredosAi’s Intelligence engine is built to keep improving. As advances in algorithms and computing power make new things possible, we build them in, so clients get more precise results from the same message library.

From the start, the engine has learned which style of message works best for which kind of customer. For example, interactive carousel messages tend to work especially well for newer customers. It has also learned how individual messages perform, since two urgent reminders can look alike and still get different results.

The released update adds a new level of detail. The engine learns how each specific message performs for each kind of customer. That includes patterns that are hard to spot by hand, where a certain wording simply lands better with certain people. 

What this looks like in numbers

On one telecom portfolio, the model now keeps track of about five times as many “this message works for these customers” patterns as before. The more of these patterns it can act on, the more precisely each message goes to the people it works for. The first rollout is underway with a major US telecom provider.

Nothing new for legal to review

The update uses only messages the client’s compliance team has already approved. It doesn’t write anything new. It gets more value out of the library that already exists.

FAQ

Do we need to write new messages?

No. The model makes better use of the messages you already have.

How is this different from A/B testing?

An A/B test usually picks one winning message for everyone. The model keeps learning which message works for which customers and adjusts as results come in.

What are the exact numbers?

On one telecom portfolio, the patterns the model tracks grew from 4,053 to 21,705. For how the model works, see our multi-armed bandit explainer.