Last month I sat in a meeting room on the 14th floor of a bank in Milan. The Chief Technology Officer had assembled his team to evaluate our product. Twelve people in the room. Three external consultants. A 47-slide deck on the table that the innovation team had spent six weeks preparing.

The CTO told me they were looking for an AI solution for their wealth management division. He said they'd already evaluated nine vendors. They had a shortlist of three. We were one of the three.

Then he asked me a question I've been asked a hundred times: "How is this different from the last AI solution we bought?"

I asked him what happened with the last one. He paused. "We're still in the pilot phase." The pilot had started 22 months earlier.

This is not an unusual conversation. This is the standard conversation. I have it in some variation every week, in cities across Europe and the Middle East. The details change. The pattern does not.

Banks keep buying AI they will never deploy. And the reasons are more structural than anyone in the vendor ecosystem wants to admit.

Innovation theater

There is a performance that happens at large financial institutions, and it goes like this. The board reads a McKinsey report that says AI will generate $200 billion in value for the banking sector. The CEO mentions it at the next leadership meeting. Someone is assigned to "own AI." A budget is created. Vendors are invited to pitch.

What follows is a procurement process that is thorough, documented, and almost entirely disconnected from the question of whether anyone will actually use the thing being purchased.

The purchase itself is the deliverable. The CTO can report to the board: we have an AI vendor. We have a pilot. We are innovating. The box is checked. The annual report gets a paragraph about digital transformation. The investor deck gets a slide with a neural network graphic on it.

I call this innovation theater. It looks like progress. It sounds like progress. But the distance between the signed contract and a working deployment is enormous, and almost nobody is accountable for closing that gap.

I've seen this in Milan with mid-size private banks. I've seen it in London with global institutions. I've seen it in Doha with sovereign wealth-adjacent funds. The scale changes, the budgets change, but the choreography is identical. Buy the tool. Announce the pilot. Move on to the next quarter's priorities.

18 months to buy, 6 months of shelf life

Here is the math that nobody talks about openly.

A typical enterprise AI procurement in European banking takes 12 to 18 months from first meeting to signed contract. By the time the legal team has finished redlining the data processing agreement, the security review is complete, the compliance team has weighed in (if they weigh in at all, more on that later), and the budget has been approved across three levels of hierarchy — a year has passed. Sometimes more.

The AI landscape moves in quarters, not years. The product that was cutting-edge when the evaluation started is mid-tier by the time the contract is signed. The competitive advantage the bank was hoping to gain has already been captured by someone who moved faster.

The shelf life of an AI solution in banking is about six months. Not because it stops working, but because the expectations around it shift. The market moves. The internal champion gets promoted or leaves. The team that was supposed to integrate it gets reassigned to a regulatory project. The vendor releases a new version that the old contract doesn't cover.

I've watched deals close where by the time we began implementation, the person who initiated the evaluation was no longer in the role. Their replacement had different priorities. The project didn't die formally. It just stopped having anyone who cared enough to push it through the last mile.

That last mile — from signed contract to live deployment — is where most AI purchases go to die. And it's nobody's explicit job to prevent that.

Buying for the board, not for the users

There is a fundamental misalignment in how AI gets purchased at banks, and it starts at the top.

CTOs at large financial institutions operate under a specific set of pressures. They need to show the board that the bank is technologically competitive. They need to manage a legacy infrastructure that is decades old and held together by middleware that nobody fully understands. They need to keep the lights on while also appearing to innovate.

When a CTO evaluates an AI vendor, they are often optimizing for a very specific audience: the board. They want something that looks impressive in a quarterly review. Something that demonstrates awareness of the latest technology trends. Something that justifies the innovation budget they fought to get approved.

What they are not optimizing for, in most cases, is the end user. The relationship manager who needs to prep for client meetings. The compliance analyst who reviews transaction alerts. The back-office team processing trade settlements.

These are the people who would actually use the AI tool every day. And in the vast majority of evaluations I've been part of, they are not in the room. They weren't consulted during the requirements gathering. They didn't test the prototype. They find out about the new tool when IT sends them a login link and a PDF user guide.

By then it's too late. The tool doesn't fit their workflow. It adds steps instead of removing them. It requires data entry they don't have time for. They try it once, find it slower than their existing process, and go back to the spreadsheet.

I sat in a meeting in Riyadh where a bank's Head of Digital proudly showed me the dashboard of their new AI analytics platform. Usage stats for the first three months. Forty-seven unique logins. The bank had 600 wealth advisors. Forty-seven logins in three months. The platform had cost them north of two million dollars.

Nobody had asked the wealth advisors what they needed. Most AI products fail for exactly this reason.

The compliance paradox

Here is the part that makes me lose sleep.

Every bank I talk to cites compliance as the primary reason AI deployment is slow. Every single one. "We'd love to move faster, but compliance..." It's the universal explanation. It's also, in most cases, a lie.

Not a deliberate lie. A structural one. Because here's what actually happens: compliance teams are almost never consulted during the AI purchasing process. They're brought in after the contract is signed, after the pilot is designed, after the integration plan is written. They're handed a finished thing and asked to approve it.

Of course they push back. They're being asked to rubber-stamp a system that processes client data in ways they haven't reviewed, with model outputs they can't explain, under a regulatory framework that doesn't have clear guidance on AI. Their job is to protect the institution. Saying "slow down" is the rational response when you've been given no context and no involvement in the design.

Compliance isn't the blocker. The failure to include compliance from the beginning is the blocker.

The banks that actually deploy AI — the ones where it reaches production and stays there — are the ones where the compliance team is in the room from day one. Not as a gatekeeper at the end. As a design partner at the start. They shape the data handling. They define the explainability requirements. They set the boundaries that the product is built within.

This approach is slower at the beginning. It's dramatically faster at the end. Because there's no six-month compliance review holding up a finished product. The compliance review happened in parallel with the build.

I brought this up to a CTO in London last spring. He nodded and said, "You're right, we should do that." Then he paused. "But our compliance team reports to a different division. I can't invite them to my vendor evaluations without going through legal and the COO." He wasn't making excuses. He was describing reality. The org chart is the obstacle.

The pilot graveyard

I keep an informal count. Across every conversation I've had with European and Middle Eastern financial institutions over the past two years, I've identified at least 40 AI pilots that are technically "active" but functionally dead. They exist in a reporting structure somewhere. They have a Jira board. Someone occasionally updates a status document. But nobody is using the product. Nobody is pushing for production. The pilot has become permanent — an undead project that consumes a small amount of budget and attention without producing any value.

Banks that collect pilots treat them like insurance policies. Having five active AI pilots feels safer than having zero. It feels like progress. It provides material for the board presentation and the regulator conversation. "Yes, we are exploring AI responsibly. We have five initiatives underway."

The banks that actually ship AI treat pilots differently. They set kill criteria at the start. If the pilot doesn't hit specific adoption metrics within 90 days, it's shut down. Not paused. Shut down. The budget is reallocated. The team moves to the next problem.

I've met exactly three banks in Europe that operate this way. Three. They are all outperforming their peers in AI deployment by a wide margin. The correlation is not subtle.

One of them — a mid-size bank in Northern Italy — killed four pilots in a single quarter. Their innovation lead told me that each cancellation taught them more than the pilots that succeeded. They learned what their people wouldn't adopt, which narrowed the field for what they would. Within a year, they had two AI tools in full production with over 70% daily active usage among target users.

The banks that never kill anything never ship anything.

What separates the banks that ship

After two years of these conversations across Milan, London, Doha, Riyadh, San Francisco, and a dozen other cities, I've identified one pattern that separates the banks that deploy AI from the ones that just buy it.

The CTO and the business line talk to each other.

That's it. That's the entire differentiator. Not the size of the AI budget. Not the sophistication of the data infrastructure. Not whether they hired a Chief AI Officer. The single best predictor of whether a bank will actually deploy AI is whether the technology leadership and the revenue-generating business units have a functioning relationship.

When the CTO understands what a relationship manager's Tuesday morning actually looks like — which systems they use, where they waste time, what information they need but can't access quickly — the AI use cases that get prioritized are the ones that will actually get adopted. When the business line understands what the technology team can realistically deliver and integrate, the expectations are grounded.

When these two groups operate in silos — which is the default at most banks — the CTO buys what looks good to the board, and the business line ignores it because it doesn't solve a problem they have.

I was in San Francisco last year, talking to a former banker who now runs product at a fintech. He said something I think about often: "At my old bank, the CTO had never sat with a client-facing team for a full day. Not once in four years. He was making technology decisions for people whose jobs he'd never watched."

That's the gap. Not a technology gap. A proximity gap.

The vendor's role in this

I should be honest about something. Vendors — including the company I work for — are part of this problem.

We sell to CTOs because CTOs have budgets. We build demos that impress executives because executives sign contracts. We optimize our pitch for the boardroom because that's where procurement decisions happen.

We are less good at insisting on access to end users before the deal is signed. We are less good at walking away from a deal when we can see that the internal conditions for successful deployment don't exist. We are less good at telling a CTO that buying our product right now would be a waste of their money because their compliance team hasn't been briefed and their business line hasn't been consulted.

The honest vendors — the ones who want long-term relationships instead of one-time contracts — are starting to change this. At Streetbeat, we've begun requiring discovery sessions with end users before we'll commit to a pilot timeline. Some of the most important conversations I've had were with the people who would actually use the product, not the ones signing the contract. Not every bank likes this. Some see it as an obstacle. A few have walked away because they just wanted to buy the product and check the box.

Those are the deals I'm glad we lost.

Where this goes

The current state of AI procurement in banking is not sustainable. Institutions are spending significant budgets on tools that generate PowerPoint slides, not production value. The vendors collecting these contracts are building revenue on a foundation of shelfware. And the banks that are actually deploying AI — the small minority — are building a compounding advantage that grows every quarter.

The gap between the buyers and the shippers will get wider before it gets narrower. The buyers will keep buying. The board presentations will keep getting made. The pilot graveyards will keep growing.

But eventually the market corrects. Eventually a bank's competitors demonstrate measurable results from AI that is actually in production. Not a pilot. Not a proof of concept. A tool that bankers use every morning that makes them measurably faster, more accurate, more productive.

When that happens — and in some segments it's already happening — the institutions that spent three years buying and never deploying will find themselves facing a deficit that no procurement process can close quickly.

The banks that win will not be the ones that bought the most AI. They'll be the ones that shipped it.

And shipping requires something much harder than a signed contract. It requires a CTO who knows what the business actually needs, a compliance team that's been in the room from the start, and an organization willing to kill a pilot that isn't working.

That's a short list of requirements. Almost nobody meets all three.