Plurio Agent Brought Neobank Finom 38% More Customers While Ad Spend Fell 8%

Finom, a European neobank for small businesses, has reported the results of five months of working with Plurio, an AI agent that manages its advertising. Over that period, the number of new customers rose 38% and ad spend fell 8%. On one social network, whose name is not customarily mentioned in Russia, customer flow from performance advertising doubled, and customer acquisition cost (CAC) dropped 57%. According to the author of the write-up, the trend is similar in Google and other channels.

A caveat up front: this is a case study published on the website of the vendor, Plurio, and retold by the Telegram channel "Дизраптор" (Disruptor). The figures have no independent verification, and the source gives no baseline CAC values, budget volumes or attribution methodology.

Why Finom's advertising could not be evaluated in the moment

Finom is one of the three largest neobanks in Europe with a focus on small business. The team numbers about 600 people, and in its latest round the company raised €115 million. The bank was founded by the same people who created Modulbank. Finom's advertising was already large-scale: more than 500 creatives ran at the same time across several countries and in different languages.

The main difficulty was the structure of the funnel. A person clicks a banner, registers, then spends weeks going through KYC, opens an account and activates it. As a result, the real "click to active customer" conversion arrived with a delay of one to two months. The budget was spent today, but it was possible to tell whether a lead had paid off only weeks later.

Over that period, creatives manage to burn out. As a result, lead cost crept up across hundreds of ads at once, and catching that manually was unrealistic. Previously, this quirk was considered an unavoidable cost of bank marketing.

How the agent works around conversion lag

Plurio received data from the data warehouse, the CRM and the ad accounts, as well as business context and purchasing rules for each country. Its decision logic is split into two modes.

If a creative already has opened accounts, the agent calculates the actual CAC and compares it with the target value for the country. It then decides whether to scale the ad, cut its budget or stop it. If there are not yet enough accounts to judge, the agent looks at the cost of registration and at a forecast of what share of registrations will reach a real account. This guards against the mistake of switching off a decent creative because of "Schrödinger's metrics": its customers are still in the pipeline but do not yet show up in reports.

The agent identifies burnout from two signals at once: impression frequency is rising and customer cost is rising. If both indicators go up, the creative is stopped. One signal is not enough, which lowers the risk of false positives.

The human stays in the loop

The Finom team did not disappear from the process. The agent produces a list of recommendations in the format "scale this, cut that"; staff approve it, and only then does the agent go into the ad accounts and make the changes. For a fintech with strict compliance requirements and different rules in each country, this setup looks sensible: a human decides on the budget, while the agent handles the routine across hundreds of ads.

The source says nothing about the technical side: it is unknown which model underlies Plurio, how many tokens or how much compute each analysis cycle takes, how often the agent is restarted, or what the system itself costs. There is also no data on what percentage of recommendations humans reject. Without this, the economics of the solution cannot be calculated, and part of the effect could have come not from the agent but from seasonality or changes in the product and funnel.

What can be taken from the case

The practical takeaway is broader than advertising in fintech. The most valuable point is that the feedback-delay problem was solved not with a new model but with careful problem framing: actual data first, a forecast from an intermediate metric when data is lacking, and stop rules tied to two independent signals. This framework fits any process where the outcome becomes visible weeks later: subscriptions, B2B sales, insurance, lending.

Teams that want to repeat the approach need three things: a unified data warehouse where clicks are linked to CRM events; clearly defined target CAC for each market; and an approval procedure in which a human signs off on the agent's actions. Without the first, the agent will have nothing to calculate, and without the third, it risks quietly switching off working ads en masse.