Why 95% of Generative AI Pilots Stall (and the Malaysian Version of the Problem)

MIT NANDA’s The GenAI Divide found that about 95% of enterprise generative AI pilots deliver little or no measurable impact on profit and loss (MIT NANDA, The GenAI Divide: State of AI in Business 2025, reported by Fortune). The cause is rarely the model. It is the data layer underneath, and in Malaysia the gap between AI ambition and data readiness is now documented.

What did the MIT study actually find?

The researchers call it the GenAI Divide: a small share of pilots cross into measurable value, while the rest stall in what they term the learning gap. Generic tools, off-the-shelf chatbots and assistants, demo well in isolation, but they do not adapt to the specific workflows, data and approval chains inside a real enterprise, so the gains disappear once the pilot leaves the sandbox.

One detail matters for anyone planning a budget: the study found that specialised solutions bought from a vendor succeeded far more often than tools built in-house. The bottleneck usually is not ambition or model quality. It is the surrounding organisational and data work most internal build teams skip.

Why do pilots stall?

In Malaysia, the pattern MIT documented shows up before a pilot even starts. Cecily Ng, Databricks’ VP and GM for ASEAN and Greater China, has described most organisations as still working with fragmented data estates spread across multiple clouds and legacy systems, siloed by business unit, which makes it “very difficult to operationalise AI beyond isolated use cases” (TechWire Asia, May 2026).

Corinium Global Intelligence, May 2026 makes the same point from a different angle: Malaysia’s push toward becoming an AI Nation by 2030 is outpacing organisational readiness, with many enterprises still working through siloed systems, inconsistent data quality and governance that has not caught up. Speed without readiness, in that framing, becomes a liability rather than an advantage.

Neither source puts a number on the gap, and we are not putting one on it either. What they agree on, qualitatively, is the shape of the problem: the data foundation, not the model, is where most Malaysian pilots run into trouble.

Is prompt engineering training enough?

Most corporate AI training in Malaysia is prompt-deep: it teaches the tool layer, prompting patterns and chat interfaces, and stops there. That is useful, but it is also exactly the layer MIT’s data describes as demoing well and then failing quietly, because the workflow underneath was never connected to real data, real approvals and real systems.

Data-deep training goes further. It builds the layers underneath the prompt: the data engineering, data science and governance that decide whether a workflow survives contact with production, not just a demo in a training room.

Prompt skills matter, but they are the front door, not the whole building. Teams that stop at prompts tend to end up back at square one the first time a pilot meets a real data source, a real approval chain or a real audit.

What does the National AI Action Plan mean for employers?

Malaysia’s National AI Action Plan 2026-2030 sets out the government’s roadmap toward an AI Nation by 2030, and it raises the pressure on every employer to show visible AI progress, not just intent. That pressure is exactly the trap MIT’s research warns about: when leaders fund AI activity to be seen moving, without first checking whether the data underneath can support it, they produce precisely the stalled, no-measurable-impact pilots the GenAI Divide report measured.

National ambition is not the problem. Funding training and pilots ahead of data readiness is.

What should you do before funding AI training?

The order matters more than the budget:

1. Check data readiness first. Before committing to a training programme, find out whether your data can actually support the use cases you have in mind. Our AI readiness and data plumbing diagnostic is built for exactly this: a fixed-scope engagement that scores your data maturity and returns a go or no-go before you spend further.
2. Train role-specific, not generic. Once you know what is ready, train the people who own the workflow on the workflow they actually run, not a one-size course.
3. Pilot on your own data. Prove the result on a real use case, with a measured outcome, through something like the AI Capstone, before you scale.

Reverse that order, and you are funding the pilot MIT already measured failing.

How do the 5% succeed?

The organisations that land in MIT’s 5% treat generative AI as the top of a stack, not the whole thing. Everyone else teaches the front door: the prompts, the chat interface, the demo. The 5% build the building underneath it, the data engineering and data science that decide whether a workflow survives contact with real operations, then connect that foundation to the front door their teams already use.

That is also why a training programme that ends in a capstone, a working pilot on your own data instead of a certificate, tends to produce a very different outcome than one that ends in a slide deck.

Not sure where your organisation should start? Book the Free AI Strategic Briefing and we’ll map it with you.