Pilot AI Intelligence isybank
Say the sentence, confirm the summary, the payment goes out.
Project
isybank is the digital bank of Intesa Sanpaolo, and everything its customers do there runs through menus. Comparing what you spent on two things means filtering a transaction list by hand. Finding one receipt means walking an archive. Picking a card means reading four plans side by side. Paying a tax form means knowing codes almost nobody knows by heart, and one wrong digit sends the money to the wrong treasury.
The brief was one assistant that answers whatever a customer asks about their own money, by voice or by typing, and then acts on the answer. Ask it to cross two things and it draws the comparison: how much went on healthy food against junk food last month, plotted rather than described. Ask narrower and it answers narrower, the last few transactions in one category, or how the investments are doing. Ask it to find a document and it searches the archive and returns the statement the request points to.
From there it acts. Transfers and F24 forms are carried out through the chat in steps. It can also read a profile, either described by the customer or inferred from how they actually spend, recommend the card that suits it, and take the purchase of that card plan to the end without leaving the conversation. Revolut shipped this shape of product as AIR about two years later, one chat that handles cards, subscriptions and portfolio questions. We were building it inside a retail bank in 2023, which is a harder room to do it in.
My team and I built a conversational layer over the existing banking services. Six of us: two on the front end, three on the back end, and me. It reads the intent, pulls the fields out of the sentence, and closes the gaps on its own. Transactions get categorised, so a spending question has something to compare and a chart to draw. The document archive gets indexed, so "the November transfer receipt" resolves to one file. Card plans get scored against a profile. Names resolve to identifiers, an IBAN out of the address book, the word IMU into tax codes 3912 through 3915, Biella into municipality code A859. When a request carries less than an operation needs, the assistant asks for one field at a time rather than handing back a blank form.
The pilot covered the whole set. Spending comparisons came back as charts, category questions as short lists, investment questions as performance, document requests as the file itself, and card recommendations as a plan the customer could buy on the spot. The recording walks four of them end to end. A transfer of fifty euro to a named payee, reason "dinner", dated to the next day, sent after one confirmation. A search that returns the unread November transfer document. October outgoing transactions, filtered. And a tax form, precompiled across twelve fields, with the bank's own system posting the two hundred euro payment and pushing the usual notification. The customer typed no field values at any point.
The language model is the least interesting part of this. What makes it usable is the pause before every action that moves money or signs a customer up to something: the summary on screen, then the wait. Answering a question wrongly costs a follow-up. Executing wrongly sends money to the wrong treasury or puts someone on a card plan they did not ask for, so the assistant would rather ask again than guess. On the tax form that meant seven questions before it wrote anything, which nobody enjoys and everybody prefers to the alternative.