Does AI-powered digital sales of financial products work?

We started out in early 2018 with the ambition to build the ultimate growth machine for retail banks. Multiple iterations, successes and failures later, we repositioned and re-engineered our solution to be an AI- powered personal finance management platform for banks. This is our story and few lessons learned

In November last year, we were finally going live! fully, completely, with no restrictions!

We’ve had been deploying 3 paid PoCs with banks in Europe and the Middle East region from April to October. In November, we got a green light to go live at full scale with our first customer; after months of sales, PoC, negotiation, deployment and testing.

Our solution has been built to predict what’s the next best digital interaction to propose to banking customer, and engage interactions using a widget added to mobile banking and internet banking platforms. Interactions were first meant to sell products ranging from investment and retirement plans to credit cards and overdraft protection.

When in full production, conversion rates didn’t match those at PoCs. Sales uplift was underwhelming as we built the expectations after deploying very successful PoCs. We needed to get to the bottom of it. So we did some research to understand why.

In parallel, we pursued another experimentation after finding that although conversion rates at scale didn’t match small scale deployments, engagement sky-rocketed and kept growing. Predicting the topics customer wanted to know more about translated into rich and recurring interactions with customers. They didn’t buy the product we pushed, but interacted a lot… substantially more than what we first thought

This is the tipping point of our pivot. We’re using now the same technology we used as a marketing tool to predict what PFM features customers want, and deliver it to customers. From being a tool to sell more products, Attila transmuted into the ultimate tool to engage customers on digital and ultimate tool to collect customer data.

So here is what we learned


We sincerely couldn’t wrap our minds about it at first. The PoCs results were conclusive. We were totally expecting a big hit. The first campaign was close to our PoC results… We breathed. The second’s results were disappointing. Customers exposed to smarter product recommendations seemed less interested in the products we were selling. Compared to a benchmark campaign, the third has just few percent improvement in conversion rates. The fourth campaign has confirmed the trend.

We surveyed customers and organized a couple panel-based research sessions and here’s what we learned


54% of customers surveyed thought that the products they were offered are likely to fit a their needs. But 87% of them declared they didn’t care to research the product nor to consider it for immediate purchase.

There was a clear pattern. One customer said : “I receive plenty of offers and that one was no different from others I have”. Another customer said “OK, I might save a few bucks on taxes subscribing to a retirement plan… so what? if I believe every email I receive, it’s super easy to get rich”

Customers just do not like to be sold stuff, especially when it comes to money management. We learned that this reluctance dilutes any benefit from personnalization


We tried to push product recommendation via physical channels. We’ve found that customers who are approached by sales people in the branches about the products we’ve predicted are 3X more likely to subscribe than on digital.

The uplift of prediction in digital is also 3X less important than in physical channels. Being better at predicting customer needs doesn’t solve it. At least not in the short term.


The findings of our first iteration lead us to trying something new, as engagement kept high despite modest conversion rates

What if we used an tweaked version of our software to address other needs that do not entail selling a product ? at least not right away.

We started out by deploying simple interactions to help customers save for vacation. The use case was obvious, it only took 3 weeks to get it from concept to live deployment.

After few weeks, our algorithm learned to predict with a 82% accuracy who is likely to engage with it. And engagement actually followed. The tool actually reached similar engagement rates than the market average while only targetting a fraction of the banks customers

Since then 5 use cases were implemented using the same solution. 42% of the MAU of the internet banking platform used one of the PFM use cases at least.


Our experimentations are clearly not sufficient to answer this question but one thing is clear. AI- Powered digital interaction do yield substantial engagement, and customer understanding.

Using it to deliver relevant and satisfactory experiences in money management has worked for our clients. The indirect uplift in sales has not yet measured, but hundred of thousands of customers have been interacting with sophisticated tools to simulate retirement, investment, they have interacted about bancassurance products.

One thing is clear: one size doesn’t fit all. Every banks has to find their own recipe. We strive to deliver the best ML based tools to help them use customer insight effectively in crafting their own customer experiences

Board member for SaaS companies, advisor to banks, obsessed by growth, I write about fintech, data-driven growth , product management and Design