How to Grow Deposits From the Customers You Already Have

Deposit growth comes from solving the problem each customer has, with a message written for that problem. A random holdout then proves the message added deposits the bank would not otherwise have had.

Deposit growth is slow. Competition for deposits is hard. The usual response is to send more messages to more customers. The number does not move.

It does not move because the message is the same for everyone. Most of the people who receive it were never stuck for the reason it addresses. A reminder to fund an account does nothing for a customer who cannot find a document. It does nothing for a customer who does not trust the transfer screen. It reaches thousands of people and speaks to almost none of them.

The approach below starts with the customer's own reason for stopping. It ends with a holdout that shows what the fix added.

Why deposit growth is getting harder

Funding has become cheaper. Deloitte's 2026 banking and capital markets outlook reports that the cost of interest-bearing deposits had already declined to 2.5 percent in the first six months of 2025. For a bank, that means each dollar of deposits earns more than it did two years ago.

Competition has not eased. The same Deloitte outlook says deposit betas may remain relatively low, particularly for regional banks, because competition for deposits remains high. A deposit beta measures how much of a market rate change reaches the depositor. Low betas mean banks are not winning deposits on rate alone.

Margins have widened. Mercer Capital's June 2026 bank watch covers banks with assets between $100 million and $5 billion. It reports that median net interest margin expanded from 3.23 percent in the first quarter of 2024 to 3.78 percent in the fourth quarter of 2025. It eased to 3.74 percent in the first quarter of 2026. The cost of deposits stood at 2.28 percent, down from 2.70 percent in the third quarter of 2024. A funded account is worth more to the bank now than it was.

Those figures point the same way. The money is worth having, and every bank is chasing the same customers. Growth has to come from customers the bank already has, including the ones who opened an account and never funded it.

The hard part

A customer who opened an account and never funded it did not stop for a general reason. She stopped for her own reason. One customer is waiting on a document she cannot find. One does not trust the transfer screen. One was told the wrong thing at a branch. A single message written for all of them speaks to none of them.

The bank's systems can see where each of these customers stopped. The core platform shows an account opened and never funded. The digital channel shows a transfer screen visited and abandoned. Behavior shows where a customer stopped and never says why.

The only source of a reason is the customer, asked. A reason cannot be read from a click. A guess about the reason stays a guess until a customer confirms it.

Getting a reason for every customer

Few customers answer surveys. Ask the ones who will.

Send a short survey close to the moment the customer stopped. Ask what got in the way, in the customer's own words. Then sort the answers into a small number of recurring reasons, kept in the customers' language. On a funding journey the list is usually short:

• a document the customer cannot find or cannot produce
• a transfer screen she did not trust with her money
• wrong or conflicting information from a branch or the call center
• a better rate somewhere else
• an intention to fund later, and then life

Most customers stay silent. For them, the bank estimates which reason is most likely, using how they behaved. A customer who opened the document upload page four times and never finished is probably stuck on a document. A customer who reached the transfer screen and left is probably stuck there. Each estimate comes with a confidence level.

When the estimate is not confident enough, say nothing. A message about the wrong problem is worse than no message. It tells the customer the bank does not know her.

It is worth being exact about what is estimated. The estimate covers the problem a customer most likely has. Whether a given message will work on her is a different question, and no model answers it. That question belongs to the test described below.

The message

The message has to speak to that one reason. A customer stuck on a document gets a link to upload it, with a note about which one is missing. A customer who did not trust the transfer screen gets a plain explanation of what happens to her money in transit. A customer who was told the wrong thing at a branch gets the right answer and an apology.

The wording has to come from language the bank already approved. Nothing goes out that compliance has not seen. The system chooses which approved message fits the reason. It does not write a new one.

That is where the cut-through comes from. A message about the customer's own problem gets read. A reminder about a problem she does not have gets deleted, however many of them the bank sends.

Proving the message added anything

Banks have a poor record of showing a return on what they buy. Only 4 out of 50 banks analyzed by Evident in 2025 reported realized ROI from AI use cases, according to Deloitte's 2026 outlook. A program built on customer reasons should be held to a higher standard than that. Outside banking, the proof goes by the name incrementality testing. Inside a bank, the tool is a holdout.

Take the customers eligible for a message and set some aside at random. Those customers receive nothing. The rest receive the message. At the end of a set window, compare the funding rate in the treated group with the funding rate in the holdout. The difference is the lift. The lift is what the message added.

Every number in the example below is illustrative and comes from no client. Suppose 40,000 customers are eligible. The bank holds back 4,000 and treats the remaining 36,000. At day 90, 12 percent of the treated group has funded and 10 percent of the holdout has funded. The lift is 2 percentage points. So 36,000 times 2 percentage points is 720 accounts the message caused to fund. At $200 a year per funded account, and assuming the balance stays for the full year, that is about $144,000.

A before-and-after slide would count every funded customer in the treated group. On the same figures, that is 12 percent of 36,000, or 4,320 accounts, and $864,000. That is six times the defensible figure. The difference is the 10 percent who were going to fund anyway.

The lift comes with a confidence interval, because random assignment balances the two groups in expectation and not exactly. The dollar figure carries that interval, so it should be reported as a range.

The holdout has to be large enough to read. A small lift on a small population cannot be detected, so the design should state before launch the smallest lift it can detect. Holding customers back means forgoing the lift on them for the length of the test. That cost is bounded and temporary, and the alternative is to pay for the program with no way to size what it returns.

Proof is what makes a message worth repeating. It also tells the bank which message to send next. A message that earned its lift on the document problem goes into regular use. One that did not is replaced, and the next version is tested the same way.

What a holdout cannot tell you

It measures the group. It does not say which customer moved, and no one can point to one customer and say the message did it.

It measures this message, on this population, in this window. A lift measured in the spring on unfunded accounts says nothing about a different message to a different group in the fall.

It cannot separate two changes made at once. If the message and the offer changed together, the lift belongs to both.

A significant lift can still be too small to be worth having. Significance means the difference is unlikely under chance alone. It does not mean the difference is large. Turn the lift into dollars and set it against the cost before deciding.

Why deposits come first

In the programs we work on, banks start with deposits, and the reason is practical. An account is funded above a threshold on a given day or it is not. So the value per completion can be stated without a long argument about attribution. Opened is not the same as funded, and a funding test measures exactly the step between them, a distinction the retail banking page draws out.

Lending is a harder place to begin. A program that nudges applicants through a loan pipeline touches credit decisions, fair lending review and CRA considerations, and more people have to be satisfied before anything runs. Those reviews mean the first holdout takes longer to launch and longer to read. What we see is banks running deposits first and taking the method into lending once the organization trusts it.

The method for finding a customer's reason is described in more detail in Ask a few. Act for everyone. The counting version of the proof argument, applied to customer experience spending, is in CX ROI by counting, not modeling.

Common questions

Why do bulk messages not move deposit growth? The message is the same for everyone, and most recipients were never stuck for the reason it addresses. A reminder does nothing for a customer who is missing a document. The few customers it does fit would often have acted anyway.

How can you know why a customer stopped if she never told you? You cannot know it. You can estimate it. The customers who did answer supply the reasons. A model then estimates which of those reasons a silent customer most likely has, from how she behaved. Where the estimate is not confident, no message goes out.

What makes one message land better than another? It speaks to the problem the customer has. A customer stuck on a document gets the document. The wording comes from what the bank already approved. Which version of a message works is settled by a randomized test against a control group, and the winner is the one that goes into regular use.

How do you prove the program added anything? With a random holdout. Part of the eligible population receives nothing for the length of the window. The lift is the funding rate in the treated group minus the rate in the holdout. Multiply the lift by the treated population and by the value of a funded account, and carry the confidence interval through to the dollar figure.

Numr runs its PXI interventions this way. Every program has a random holdout. The lift is reported as treated versus held out, in dollars, with its confidence interval, and the arithmetic is available to the client.