I've been in this meeting more times than I can count. A room in a bank's headquarters. Nice chairs, bad coffee. On one side, the innovation team. On the other, a vendor with a demo. The conversation is enthusiastic. The pilot gets approved. Everyone shakes hands.
Six months later, nothing has shipped. The pilot is "still running." The vendor is still billing. The innovation team is already talking to the next vendor about the next pilot. And the business line that was supposed to use the product has never heard of it.
This is pilot purgatory. And it's the default state of bank innovation in 2025.
The numbers nobody talks about
I don't have a McKinsey report to cite here. I have something better: my own deal pipeline and the conversations behind it. Over the past three years, I've been involved in AI-related conversations with banks across Italy, Germany, the UK, and the Gulf. Here's what I see.
A typical Tier 1 or Tier 2 bank in Europe has between five and fifteen active innovation pilots at any given time. These span AI, blockchain, open banking, embedded finance, whatever the current theme is. Of those pilots, fewer than 10% ever reach production. Not "fail" — that would imply someone made a decision. They just... persist. They sit in a sandbox, generating quarterly reports that say "promising results" and "further evaluation needed."
In Milan, I worked with a bank that had been running an AI pilot for document processing since 2022. Three years. The pilot "worked." The accuracy was good. The compliance team had even signed off. But the product was still in a test environment because nobody in the operations team had been asked to change their workflow. The innovation team had done their job. Operations had done theirs. The two jobs never intersected.
In Frankfurt, same story, different technology. A RegTech pilot that had been approved, tested, validated, and then left to die because the IT department had a two-year backlog and no mandate to prioritize innovation projects over regulatory deadlines.
In Doha, a bank had signed pilot agreements with four different AI vendors simultaneously. Four. Each one was working in isolation, each one was "going well," and none of them had a path to production because the bank hadn't decided what problem they were actually trying to solve. They were collecting pilots like stamps. Most AI products fail this way — not because the tech is broken, but because nobody validated the premise.
The incentive problem
Here's the thing nobody in the room wants to say: innovation teams are measured on pilots launched, not products shipped.
Think about that for a second. The KPI is activity, not outcome. The Head of Innovation goes to the board and says: "We launched eight pilots this year. We're exploring AI, blockchain, and embedded finance. We're working with leading vendors." The board nods. The budget gets renewed.
Nobody asks: "How many of those pilots are in production? How many are generating revenue or saving costs? How many have been killed because they didn't work?" Those questions would be uncomfortable. They might reveal that the innovation program is an expensive research lab that produces demos, not products.
I've sat in rooms where the innovation team privately admitted that they knew a pilot wasn't going anywhere, but they kept it running because killing it would mean admitting failure. In some banks, the pilot had become a political artifact — a signal that the institution was "doing innovation" — rather than a genuine product development effort.
The incentive structure creates a specific behavior: launch many, ship none, report all. And as long as the board measures innovation by volume of activity rather than volume of impact, nothing changes.
The sandbox illusion
Every bank has a sandbox. Some call it an innovation lab. Some call it a test environment. Some call it a "digital factory." The name doesn't matter. The function is the same: it's a safe space where new technology can be tested without touching the real systems.
In theory, this makes sense. You don't want untested AI models making decisions on live customer data. You need a controlled environment.
In practice, the sandbox becomes a trap. A pilot that "succeeds" in a sandbox has proven exactly one thing: that the technology works under controlled conditions with clean data and no integration constraints. That's the easy part. The hard part — connecting to legacy core banking systems, handling real data quality issues, meeting production SLAs, passing security reviews, training end users — none of that happens in the sandbox.
I've seen pilots that had 95% accuracy in the sandbox drop to 60% when exposed to real production data. Not because the model was bad, but because real data is messy. Real data has edge cases. Real data has formats that were never documented because the system was built in 1997 and the person who designed it retired in 2008.
The gap between sandbox success and production readiness is enormous. And most banks don't have a process for crossing it. The pilot ends at the sandbox. The "go to production" decision requires a completely different set of stakeholders, a completely different budget approval, and a completely different risk assessment. It's essentially starting over.
The organizational wall
This is the structural problem underneath everything else. In most banks, the innovation team is organizationally separated from the business line.
The innovation team reports to the CIO, or the Chief Digital Officer, or sometimes directly to the CEO as a "strategic function." The business line — wealth management, retail banking, corporate banking, operations — reports through a completely different chain. They have their own budget, their own priorities, their own roadmap.
When the innovation team develops a pilot, the business line hasn't been involved. They haven't asked for it. They don't own it. They didn't allocate resources for it. And when someone from innovation walks in and says "we have this great AI tool for you," the business line's first reaction is not excitement. It's suspicion. Who asked for this? Who's going to support it? Who's going to retrain my team? What happens when it breaks?
I saw this play out in London. A bank's innovation team had built a genuinely impressive AI-powered client analytics tool. It worked. It was useful. The wealth management team looked at it and said: "This is nice, but we're in the middle of a CRM migration. We can't adopt anything new for eighteen months." The innovation team had solved a real problem. But they'd solved it for a customer that wasn't ready to buy.
The organizational separation means that innovation operates in a vacuum. It can explore, but it can't deploy. It can build, but it can't ship. Because shipping requires the cooperation of people who were never part of the conversation.
The vendor's dirty secret
I'm going to say something that might sound counterintuitive, given that I work on the vendor side. Most AI vendors love pilots. Pilots are revenue. They're low-risk, high-margin engagements. The vendor gets paid to set up a demo, run a proof of concept, and deliver a report. If it goes to production, great. If it doesn't, the vendor still got paid.
This creates a perverse alignment. The bank wants to look innovative. The vendor wants revenue. The pilot satisfies both needs without requiring either party to do the hard work of actually deploying something.
But here's the dirty secret: vendors hate the cycle too. At least the good ones do. Because a pilot that never converts to production is a customer that never scales. It's revenue today but a dead end tomorrow. The best vendors — the ones building real businesses, not just consulting shops — want production deployments because that's where the long-term value is. That's where the recurring revenue comes from. That's where the case studies come from.
At Streetbeat, we've had to make a deliberate choice about this. We've walked away from pilot engagements where we could see from the first meeting that there was no path to production. The innovation team was enthusiastic, but there was no business line sponsor, no IT commitment, no compliance involvement. We knew the pilot would "succeed" in the sandbox and then die. We'd get paid, but we'd waste six months and have nothing to show for it.
That's a hard conversation to have with a potential customer. "We don't think you're ready for a pilot" is not a great sales pitch. But it's honest. And in the long run, it builds the kind of trust that actually leads to real deployments.
What the 10% do differently
So what separates the banks that actually ship from the ones stuck in pilot purgatory? I've seen enough of both to identify the pattern.
First, there's a CEO or C-level mandate with a deadline. Not "explore AI" — that's a blank check for endless exploration. Instead: "By Q3, I want automated KYC remediation in production for the retail segment." Specific. Measurable. Time-bound. When the mandate comes from the top with a real deadline, the organizational barriers dissolve because nobody wants to be the person who explains to the CEO why they missed it.
Second, the business line owns the project, not the innovation team. The Head of Retail Banking or the Head of Wealth Management is the sponsor. They define the requirements. They allocate their own resources. They're responsible for adoption. The innovation team plays a supporting role — scouting vendors, evaluating technology — but the business line drives.
Third, they use a 90-day kill-or-ship framework. The pilot has ninety days. At the end, there are two options: go to production or kill it. Not "extend for further evaluation." Not "phase two exploration." Kill or ship. This forces every stakeholder — IT, compliance, operations — to engage from day one, because the clock is ticking and nobody wants to be the bottleneck.
Fourth, compliance is at the table from the beginning. Not as a gate at the end that can veto everything. As a design partner from week one. The banks that succeed treat compliance as a feature, not a blocker. "This AI is compliant by design" is a competitive advantage. The EU AI Act makes this even more concrete by providing a clear regulatory checklist. "We'll figure out compliance later" is a death sentence.
I saw this model work at a Gulf bank that went from zero to a fully deployed AI-driven advisory tool in four months. The CEO set the mandate. The wealth management head owned it. Compliance was embedded in the project team. And they had a hard deadline that everyone knew was real. No sandbox theater. No quarterly "progress reports." Just: build it, test it, ship it, or kill it.
What Streetbeat taught me about getting past the pilot
Building and selling an AI product to banks has taught me something I couldn't have learned from any other vantage point: the technology is almost never the problem.
We can build models that work. We can demonstrate value in a sandbox in weeks. That part is straightforward. The hard part is everything around it. Getting IT to allocate integration resources. Getting compliance comfortable with the risk framework. Getting the business line to change their workflow. Getting the budget approved through the right channel.
What I've learned is that the vendor has to do more than deliver technology. You have to help the bank navigate its own internal politics. You have to map the stakeholders before you write a line of code. You have to ask, in the very first meeting: "Who owns this if it succeeds? Who has the authority to deploy it? What's the timeline, and is it real?"
If the answers are vague — "the innovation team is exploring this," "we'll figure out deployment later," "there's no hard deadline" — that's a pilot that's going nowhere. Walk away or fix the governance first.
The best engagements I've been part of started not with a demo, but with a conversation about organizational readiness. Can this bank actually ship something? Do they want to? Is there a person with enough authority and enough urgency to make it happen?
If yes, you can move mountains in ninety days. If no, you can run pilots for three years and ship nothing.
The real question
Every bank I talk to asks me the same thing: "How do we innovate faster?" They expect me to talk about technology. Better models, faster infrastructure, more data.
My answer is always the same. Your technology is fine. Your operating model is broken.
Stop measuring pilots launched. Start measuring products shipped. Stop separating innovation from the business. Start embedding it. Stop giving pilots unlimited timelines. Start killing them when they don't convert.
The banks that figure this out in the next two years will own the decade. The ones that don't will still be running pilots in 2030, wondering why nothing ever ships.
I know which group I'd rather work with.
