Decoding AI ROI Paradox: Why 85% of Projects Die in Production

Here’s a familiar scenario: The chief executive officer talks up a cutting-edge AI pilot and makes it an imperative, gets the boardroom excited about why AI is the future and why the organization needs to get on the bandwagon, and then gets the budget approval. That was how it looked during the last Chief Digital and Data Officer Asia Summit 2024.

In the latest edition — Chief Digital and Data Officer Asia Summit 2025 — the mood had turned somber. That’s because a lot of POCs are now being shelved, deployment is becoming DOA, and AI aspirations are facing the budget-blowing costs. The result: a digital graveyard, where 85% are left to die, abandoned Jupyter notebooks, and deflated data science teams.

Which is why this year’s keynote, “Turning Potential into Real-World Performance – The Hunt for AI ROI,” grabbed much attention. Grace Chin, manager specializing in data transformation at Wavestone and the keynote speaker, started the Summit with a sobering insight: “AI is not just a tech problem, it is a change problem.”

Chin, who has spent over a decade shepherding digital transformations across financial services and healthcare, linked AI trust with ROI. “Even if you have the best model in the world, if nobody trusts it, nobody uses it. Don’t even talk about ROI — ROI will never come.”

The four horsemen of AI failure

Chin identifies four critical blockers that kill AI ROI before it can breathe:

• People and change management: The human element remains the biggest wildcard. Engineers might architect flawless inference pipelines, but if end-users don’t trust the black box making decisions about their workflows, adoption flatlines. “Change is hard,” Chin notes, “and without trust, there’s no adoption.”

• Technical debt: Legacy systems are the quicksand of AI projects. Data scattered across fragmented APIs, brittle pipelines, and unreliable data warehouses creates a foundation of sand. Chin cited a survey of global data leaders that found 31% cite unreliable data as actively blocking their AI projects — that’s a diplomatic way of saying “garbage in, garbage out.”

• Regulatory complexity: From Singapore’s PDPC framework to the E.U. AI Act, compliance landscapes shift faster than model training cycles. Chin shared that only 25% of APAC companies have implemented robust AI compliance protocols — a regulatory time bomb waiting to detonate.

• Talent scarcity: The war for AI talent isn’t just about data scientists anymore. You need MLOps engineers, model governance specialists, and business champions who can translate between C-suite priorities and neural network architectures. The best practitioners are already locked up by Big Tech or well-funded startups.

The European bank that cracked the code

One of Wavestone’s clients — a European bank drowning in legacy systems and regulatory constraints — managed to escape the 85% failure rate. Instead of spinning endless pilots, they asked a different question: “How do we embed AI into our business in a scalable and measurable way?”

Their approach was surgical. They identified high-impact use cases tied directly to risk, cost, and customer experience metrics. But the real innovation was organizational: they launched an “AI factory” — a new operating model with cross-functional teams owning AI products end-to-end.

The bank didn’t just hire more data scientists; they created an ecosystem. They ran an AI summer school for 2,000 employees, teaching basic AI literacy alongside data ethics. They built an “AI watchtower” to track every use case with measurable outcomes. Most importantly, they embedded governance from day one.

The results speak in the language executives understand: 15% of marketing campaigns now AI-generated, 75% reduction in document handling time, and measurably improved analytics performance. “The success wasn’t just the models,” Chin emphasizes. “It was the system around the models — the people, the training, the watchtower, the culture, the process, the governance.”

From tinkering to transformation

The fundamental shift required is mental: stop asking “what can AI do for us?” and start asking “what do we need AI to do for our business?” Chin notes. It’s the difference between curiosity-driven exploration and outcome-driven execution.

Chin’s framework centers on five pillars: Strategy (focus on business-critical use cases), People (address fear and build capability), Foundation (clean data and robust tech), Governance (embed compliance from day one), and Talent (build cross-functional teams, not just hire data scientists).

“Do not build what you cannot measure,” Chin warns. “Pick two to three pain points that can really move the needle.” The graveyard of failed AI projects is filled with technically impressive solutions to business-irrelevant problems.

Singapore’s AI Advantage

Singapore sits in a unique position to break the 85% failure curse. The city-state combines AI infrastructure leadership with a digitally native workforce and collaborative regulators. National models, open-source initiatives, and public-private partnerships create an environment where the transition from pilot to production actually has institutional support.

But even in this favorable ecosystem, the fundamentals remain unchanged. “Design AI with ROI in mind. Get clarity on ownership. Upskill teams. Data is the foundation — invest in it,” Chin advises.

The call to action is simple but not easy: identify the use case where AI can deliver immediate value. Then build and measure it, and then scale. The secret isn't in the code; it’s in the system around the code.

The 85% failure rate is a symptom of treating AI like a technology problem instead of a transformation challenge. The companies that crack this code won't just survive the AI revolution; they’ll architect it.

Image credit: iStockphoto/champja