Allica Bank Pushes AI Beyond Use Cases Into Operating Model

Ravneet Shah described Allica Bank’s AI work as a shift from experimentation toward an operating model. The bank, which has operated in the UK since receiving its banking license in 2019, serves SMEs with lending, business current accounts, and deposit products. Shah framed its differentiation as the combination of technology and relationship banking.

The AI work began in 2023, and Shah was explicit that the first phase involved mistakes and uncertainty. Allica did not initially know which use cases would matter or where the technology would fit. The clearer view, in his account, came from treating AI less as a tool for a few specialist teams and more as a change in how the whole organization works.

That required “unlearning,” including for Shah himself despite his technology background. The first priority was adoption across the company: operations, distribution, technology, product, finance, and other functions. Shah said Allica moved from roughly 25% adoption last year to a median workday adoption rate of 77%.

For Shah, adoption alone is not the strategy. He separated Allica’s AI agenda into three layers. The first is broad organizational adoption: people changing how they work. The second is the product-engineering operating model: how teams build software and ship changes. The third is the product layer itself: using AI and agentic applications to do work that traditional software engineering and earlier machine-learning approaches could not address.

That distinction matters because the same AI practices do not apply uniformly across a bank. Shah said what works for the rest of the organization does not necessarily work for product-engineering squads. The bank therefore changed both the tools people use and the structure of the teams expected to build with those tools.

Clement Andre pointed to Allica’s scale of deployment as evidence that the bank’s operating model is unusual for a regulated financial institution. He said Allica made more than 3,700 deployments last year, a number he called notable given the size of the team. Shah said Allica has about 100 engineers, and fewer than 200 colleagues in the broader product-engineering organization when product, data, and related roles are included.

Historically, Shah said, Allica used a Spotify-style model of cross-functional squads. Those squads brought together product, data, backend and frontend engineers, SDETs, and designers. The model had worked for the bank’s earlier stage of product delivery. With AI, Shah said, Allica concluded that the model needed to change.

The main structural change was to reduce squad size. Allica now uses “squadlets,” and Shah said their composition varies depending on the complexity and nature of the product being built. The bank moved away from a fixed squad blueprint and toward smaller team structures tailored to the work.

The reason was not simply speed in the abstract. Shah argued that some handoffs required in the previous model were no longer needed. Allica had believed in T-shaped roles: deep specialization with adjacent skills. Shah said the newer direction is closer to “comb skills,” where people develop multiple deeper capabilities across adjacent areas.

That change has altered job boundaries. Instead of separating backend engineers, frontend engineers, and SDETs, Allica combined those roles into one. Product roles are changing too. In some squads, Allica combined product owner and product analyst responsibilities. In others, Shah said, the team may have a product representative who could come from either discipline. In a smaller number of squadlets, Allica is experimenting with a “product engineer”: someone who can do product work and engineering work at the same time.

Shah emphasized that this remains a journey, not a finished transformation. The most ambitious version is that by the end of the year, product and design colleagues should be able to ship code to production. Engineers can already do so. Shah said this is already happening in some squads, though not yet to the extent Allica wants.

Andre connected that structure to governance and speed. His reading was that Allica avoids some of the “governance drag” that often appears with AI because the people required to make decisions and push a product are inside the same squad. In that model, a product can move from creation to publishing without leaving the squad unless necessary.

The broader implication, as Andre put it, is that companies may need not only to unlearn old habits but also to “unstructure” teams built for an earlier software era. Shah’s examples were concrete: smaller squads, fewer handoffs, merged roles, and a higher expectation that non-engineering product roles can directly participate in production delivery.

Lending is Allica’s core business, and Shah described its lending process as complex and not fully automated. The bank works with introducers and brokers, and in asset finance it receives many applications by email. Allica has portals, but Shah said customers, brokers, and introducers often do not like to use them. Even when the portal is used, the information can arrive incomplete.

The response, in Shah’s framing, was not to force customers and brokers into the bank’s preferred channel. Allica instead changed its own process. The bank introduced an agent that can inspect information arriving by email, identify what is missing, and ask brokers for additional information before the application is fed into Allica’s portal.

That is a different use of AI from a simple chatbot or a front-end automation layer. It is closer to process repair: taking a workflow that depends on messy, incomplete, customer-provided information and using AI to normalize it enough for the bank’s existing systems.

Shah said Allica uses a combination of deterministic and non-deterministic agents in that journey. He did not define the architecture in detail, but the distinction implies that some parts of the process follow fixed logic while others rely on probabilistic AI behavior. He also noted that “harnesses” are in place, acknowledging controls around the agentic workflow without detailing them.

The result, according to Shah, is that some applications have seen time to decision reduced to less than seven minutes or twelve minutes.

The important claim was not only faster decisions. Shah’s point was that AI lets the bank improve the process without demanding that customers change their behavior first. Andre linked this to a broader principle raised around voice and banking: some customers are willing to change channels or habits, and some are not. In that setting, the technology should meet customers where they are.

Shah said Allica had already been introducing technology over the last few years. What has changed is the class of processes the bank can now reconsider. Where manual processes could not previously be solved through standard software applications, Allica is asking where agents can help.

Relationship banking is part of Allica’s stated differentiation, and Shah said the bank had internal debates about how AI should apply there. The obvious path would have been conversational chatbots. Allica did not choose that as the core move.

Shah said the bank did not want to replace relationship managers with AI bots. Messaging may exist as a needed core function, but that is not the same as substituting the relationship manager. Instead, the goal is to support relationship managers with insights and information about their customers.

Andre described this as augmentation: helping employees do more and focus on work that matters. The intended outcome is not necessarily more time with customers, but better-quality time. Shah sharpened that point by focusing on context. Relationship managers should not have to spend large amounts of time assembling background information before a customer conversation. AI should provide enough context to help them drive that conversation.

This is a narrower and more operational claim than “AI improves customer relationships.” Allica’s version, as Shah described it, is that relationship managers remain central, while AI reduces the time they spend gathering and synthesizing customer information. The human relationship remains the interface; AI improves the preparation layer behind it.

Looking ahead twelve to twenty-four months, Shah said Allica now has a clearer direction than it had six months earlier. At that earlier point, he said, the bank was still experimenting. Now, at least in product engineering, the destination is clearer.

Andre referenced Allica’s 3,700 deployments last year, and Shah said the bank wants to double that. But Shah immediately qualified the metric. Doubling deployments could become a matter of “playing with the numbers” or “gaming the system.” The real target is not deployment count for its own sake; it is increasing product increments for customers.

Those increments include both customer-facing improvements and internal-facing improvements. Shah specifically mentioned risk, compliance, and security as areas where internal product increments matter. That is consistent with the regulated-bank context: velocity is not useful if it bypasses the functions that make the bank safe, compliant, and reliable.

Shah also tied the ambition to service quality. Allica wants to maintain the speed of its service while improving quality. Andre summarized that direction as serving customers better, moving faster, and becoming more efficient, drawing an analogy to OpenAI’s own emphasis on better models, cost efficiency, and building for users.

The exchange ended with Andre thanking Shah not only for using OpenAI’s technology but also for feedback. He said that in two conversations Shah had already given several pieces of feedback on things OpenAI needed to improve or ways it could better help Allica.

The underlying tension remains visible throughout Shah’s account. Allica is trying to increase speed in a regulated environment without reducing the role of governance, relationship managers, or customer preference. Its answer is structural as much as technical: broader adoption, smaller and more capable product teams, agentic workflows where traditional software falls short, and AI support for human roles that the bank does not want to replace.