Most Banks Will Automate the Wrong Thing
Leonid Goriev
Founder Alty
July 1, 2026
A reaction to Deloitte's Tech Trends 2026, from the team that builds banking products underneath these decisions.
There is one number in Deloitte's Tech Trends 2026 worth sitting with. After two years of agentic AI being the thing every board wanted, only 11% of organisations have agents running in production. Thirty-eight percent are piloting. Forty-two percent are still writing the strategy. Another 35% have no strategy at all.
The usual read on a gap like that is "the technology isn't ready yet." The sharper read is the one Deloitte and Gartner arrive at separately. Gartner expects more than 40% of agentic projects to be cancelled by 2027, and its own recommendation is blunt: in many cases, rethinking the workflow from the ground up is the path that works, because dropping agents into legacy processes mostly disrupts them. Deloitte's CTO puts the same idea in a line I haven't been able to shake: apply advanced AI to an existing workflow and you weaponise its inefficiency.
We've spent fifteen years building banking products, including the first version of monobank and platforms now used by tens of millions of people. So when a bank tells me it's adding an AI agent to a process, my first question is never about the agent. It's whether anyone has looked at the process underneath it, because that's the thing the agent is about to scale.
An agent inherits the process it's dropped into
Here's the part the "add an agent" framing skips.
An agent is a multiplier. Point it at a clean, well-understood process and it makes that process faster. Point it at a tangled one, full of steps that exist only because nobody ever questioned them, and it makes the tangle faster too. The inefficiency doesn't disappear under automation. It scales, with a budget line attached and a launch post announcing it.
This is why so many of the banking AI launches over the past year haven't moved the numbers that mattered. We've written before about how AI tends to expose a weak product rather than fix it; the same logic applies one layer down, at the process. The agent worked. The process it was bolted onto was the actual constraint, and that went untouched. A bank that automates its existing onboarding flow doesn't get faster onboarding. It gets a faster version of a flow that was already losing people, now with an AI line in the budget.
The teams Deloitte found getting value did one unglamorous thing first. They looked at how the work actually happens, decided what the work should be, and only then asked where an agent fits. The order is the whole point. Redesign, then automate. Reverse it and the automation locks the old shape in place.
What redesigning first actually looks like
By 2021, a leading Nigerian bank we worked with had spent years as one of Africa's most respected institutions and still had the ambition. What it had lost was the ability to act on it. Years of incremental fixes had left its mobile platform risky to change, and the teams had started avoiding the core rather than touching it. Every new idea ran into the same wall: the process of shipping a change had become more dangerous than the change itself.
Dropping an automation layer onto that would have automated the fear. The faster path to "innovation" would have been a visible AI feature on top, with the same brittle foundation underneath and the same teams still afraid to touch it.
What the work actually required was slower and underneath the surface. The team mapped what was constraining the platform, simplified the product architecture, and rebuilt the technical foundation so that a change stopped being a gamble. Then came the part that mattered more than any feature: structured handover, so the bank's own people could ship changes again without bracing first. The app rating moved from 3.4 to 4.7 stars and the platform grew to roughly three million active users. None of that came from adding intelligence to the old process. It came from rebuilding the process so intelligence would have somewhere stable to land.
That sequence is what "redesign, don't automate" means in a bank. Not a slogan about ambition. A decision to fix the shape of the work before pointing anything fast at it.
The question worth asking before the agent
The report frames the choice as organisational rather than technological, and in banking that's exactly right. The constraint is rarely the model. It's the process the model is about to inherit, and whether anyone has the appetite to redesign it before automating it.
So before the next agent goes into a roadmap, there's a narrower and more uncomfortable question than "where can we add AI." Take the process you're about to automate and ask what it would look like if you designed it today, from scratch, knowing what you now know. Where the honest answer is "nothing like it looks now," the next step is redesign, and the agent comes after. It will be worth more when it does.
Most banks will skip that question, ship the agent, and show the board momentum. A smaller number will fix the process first and let the agent compound something worth compounding. Eighteen months from now, the gap between those two groups is the only number that will matter.
Leonid Goriev is the founder of Alty, a product partner for fintech and banking. Alty works with banks and fintechs on the decisions that shape what gets built before the build starts.
Sources: Deloitte, Tech Trends 2026 (December 2025); Gartner press release, 25 June 2025.
