Plug-and-Play AI Is a Myth: Why Enterprise AI Projects Stall, and What Works Instead

Enterprises are spending more on AI than at any point in history, and most of them can't tell you what it bought them.

The gap isn't money

Cognizant recently surveyed 600 AI decision-makers and interviewed 38 senior executives across the US, Germany, Singapore and Australia. Eighty-four percent had a formal AI budget. More than half were spending over $10 million a year. And 63% reported a gap between their AI ambitions and what they could actually do.

Sit with that for a moment. These aren't companies that haven't heard of AI. They have committees, budgets, and eight-figure line items. The thing holding them back isn't access to the technology. The top blockers they named were regulatory concern, an inability to show return, and — the one that should worry every board — no clear AI strategy.

I see a version of this in nearly every organisation I work with. Someone signs off the budget, a vendor runs the rollout, and a year later there's a contract, a dashboard, and no honest answer to "what changed?"

Why generic AI doesn't fit

The Cognizant finding that matters most is this: enterprises now cite generic, off-the-shelf AI as a leading reason to reject a provider, alongside lack of industry expertise and inability to integrate with their existing systems.

That tracks with what actually happens inside a business. Every firm has its own language, its own controls, its own reconciliation process that someone built in 2014 and nobody wants to touch. A regulated institution has more of these, not fewer — model governance, audit trails, data residency, lines of accountability. No off-the-shelf product accounts for any of it. It can't. It was built for the average of ten thousand companies, which means it fits none of them precisely.

My colleague Sacha Windisch makes this argument well in One-of-One. At Scale.: custom software used to be a Fortune 500 luxury because building it was expensive. AI has collapsed that cost. Bespoke is now viable for almost anyone — and once it's viable, the generic option stops being good enough. Enterprises aren't rejecting SaaS because they've turned against software. They're rejecting it because they no longer have to bend their workflow around someone else's assumptions.

Build, buy, or partner — what the evidence says

There's a widely-cited MIT study from 2025, The GenAI Divide, with a headline number that did the rounds: roughly 95% of enterprise generative-AI pilots delivered no measurable impact. Treat the exact figure with some caution — it's been challenged as overstated, and the methodology is contested. But the more durable finding inside the same study is the one worth acting on: buying capability from specialised partners succeeded about 67% of the time, while purely internal builds succeeded around a third as often.

Read alongside Cognizant — where decision-makers ranked builder firms above management consultancies for AI adoption, a 23-point trust gap — a clear pattern emerges. The firms that get value aren't the ones that bought the most generic tooling, and they aren't the ones that tried to build everything alone in-house. They're the ones that partnered with people who build for their specific context.

That's the real "build vs buy" answer, and it's neither extreme. Buying generic gets you a tool nobody configured for your work. Building everything yourself means learning lessons the expensive way, on your own time. The third path — a partner who builds with you, on your data and inside your controls — is the one the data keeps pointing at.

What this looks like in financial services

In a regulated firm the stakes are sharper, because the failure modes are not just wasted spend. They're explainability and accountability.

The cautionary tale everyone in this sector should know is Klarna. In early 2024 the company said its AI assistant was doing the work of 700 customer-service agents, handling 2.3 million conversations a month, with around $40 million in projected savings. By 2025 it was quietly rehiring humans, after service quality dropped on the complex, emotionally charged cases the AI couldn't handle. The CEO admitted publicly that they'd gone too far. The lesson isn't "AI doesn't work." It's that the gap between the vendor pitch and operational reality is exactly where the cost lives — and in financial services, that gap has a regulator standing in it.

Because the pressure is now real and dated. The EU AI Act's high-risk obligations become enforceable on 2 August 2026, with credit scoring and insurance pricing named explicitly, and fines reaching €35 million or 7% of global turnover. In the US, the SEC has put AI governance and "AI washing" near the top of its 2026 examination priorities, and has already sanctioned advisers for claiming AI capabilities they didn't actually have. A generic tool you can't explain to an examiner is no longer just an underperforming asset. It's a liability with your firm's name on it.

Contrast that with what purpose-built actually looks like. We did a piece of work for a firm that builds tradable investment indices. They needed to manage the data flowing from the independent financial advisers who use their products, and surface it two ways at once: back to their own team through a back-office interface they could administer, and out to the advisers and the advisers' own clients, so each could see how their indices were actually performing. No off-the-shelf product models those relationships, because they don't exist anywhere else — they're specific to how that business works. We built it on their data and around their own workflow, which is exactly why it fits, and exactly why a generic dashboard never would.

The way through

There's no shortcut, and the way out isn't another licence. It's two steps most firms skip.

First, diagnosis before deployment. Before anyone buys or builds, work out where the real bottlenecks are, which workflows actually justify automation, and what "good" looks like in your world. Most of the $10-million-a-year crowd never did this — they started with the tool and worked backward to a problem.

Second, build for specificity, not for the average. Configure to your data, your terminology, your controls. In a regulated environment, that's not a nice-to-have; it's the difference between a system you can stand behind and one you're hoping nobody asks about.

That's the work we do at Exponential Partners — diagnosis first, then purpose-built. We're a team that has worked inside financial institutions and ships custom products, not a firm that hands you a deck and leaves. If your AI spend is climbing and you still can't point to what it changed, that's the conversation worth having.

References

1. Cognizant, Plug-and-Play AI is a Myth, 10 March 2026. news.cognizant.com
2. MIT, The GenAI Divide: State of AI in Business 2025 (as reported, Fortune, Aug 2025). fortune.com — figure contested; see marketingaiinstitute.com
3. Klarna AI customer-service reversal (Entrepreneur, 2025). entrepreneur.com
4. EU AI Act penalties, Article 99. artificialintelligenceact.eu
5. SEC 2026 examination priorities. wealthmanagement.com