A chatbot that answered questions, a model that spotted patterns, or an agent that summarised reports used to be met with enthusiasm and applause in boardrooms and often secured follow-on funding.
Today, enterprises expect more. The pilot is no longer the destination. It is the starting line. Yet most organisations remain stuck there, unable to turn promising experiments into reliable production systems.
According to MIT’s Nanda project, 95 percent of organisations saw no measurable return on their AI investments, highlighting the gap between successful experimentation and real-world deployment.
Edward Lim of Entelechy Asia interviewed industry leaders on what it takes to move AI from successful pilot to reliable, full-scale production.
“The biggest challenge is not the AI model itself, it is getting the underlying data infrastructure ready for production,” said Michael Cronin (below left), APAC Managing Director of Couchbase. “Moving AI from pilot to production is less about building a smarter model and more about building a unified data foundation.”
Philip Madgwick, Regional Vice President of Asia at Alteryx, puts it equally plainly, “Get your foundations right before you scale. The best models and the most sophisticated infrastructure will not save a deployment where the people, processes, and governance are not ready.”
3 hurdles
What separates a successful demo from a reliable production system? Industry leaders across the Asia Pacific region point to three consistent hurdles that trip up even the most promising initiatives.
- Fragmented data
Pilots can work happily with isolated, curated datasets, but production demands real-time access to operational records, vector stores, and context spread across multiple systems.
“Pilots run on sanitised, curated datasets; production does not. Enterprise data is often trapped in silos, backup systems, multiple clouds, or formats that AI cannot easily consume,” said Andy Sim (above right), Vice President and Managing Director of Singapore at Dell Technologies.
Madgwick added that the issue runs deeper than infrastructure. “The first is getting data into a genuinely trustworthy state. This means auditing, consolidating and governing data across fragmented sources before it can reliably feed any AI workflow. Many organisations skip this and go straight to model selection, only to find AI producing outputs that cannot be verified or acted on.”
Alteryx’s 2026 research shows that in Singapore alone, 46 percent of failed AI and analytics projects trace back to data-related issues rather than model or tooling failures. - Persistent context and logic
Production AI must retain user history and conversation state over time, linking interactions to live operational records.
“Without persistent memory connected to live operational data, AI agents risk making decisions based on outdated information,” noted Cronin.
Madgwick expands this into what he calls the business logic layer: “This refers to the rules, calculations and operational context that define how a business actually operates… Without this, AI systems run on logic that is incomplete, outdated or misaligned with how decisions are actually made.” - Governance and security
Pilots often overlook enterprise controls, but live deployments require robust access policies, audit trails, and compliance frameworks from day one. Retrofitting these capabilities into disconnected systems is both complex and costly.
For Mohan Veloo (below right), Chief Technology Officer of Asia Pacific, China and Japan at F5, the shift is even more fundamental: “A pilot proves that AI can work. Production proves that the enterprise can depend on it.”
“True production requires a platform that treats security as an inherent property of the storage and compute layer, rather than a bolt-on feature,” said Daryush Ashjari (below left), Chief Technology Officer and Vice President of Solution Engineering, Asia Pacific and Japan at Nutanix.
Why promising projects stall
A successful pilot can create false confidence.
“The most common pattern we see is organisations treating the pilot as the destination rather than the starting point,” observed Sim. “A successful pilot proves a concept, but it does not prove the organisation is ready to scale it.”
Data problems that teams can work around during testing often become production blockers.
“Fragmented, ungoverned, or poorly protected data will stall even the most promising project the moment it hits real-world conditions,” he added.
Costs also change dramatically at scale. Ashjari described “token shock” as the moment organisations realise that public-cloud inference costs have multiplied far beyond their pilot estimates.
Gartner projects that more than 40 percent of AI agent projects will be cancelled by the end of 2027, with escalating costs and unclear business value cited as primary drivers.
Governance often arrives too late too.
“When governance arrives at the end, it becomes a brake. When it is built in from the start, it becomes an accelerator,” said Veloo.
The operating model matters just as much. “AI does not scale because the model gets bigger. It scales because the organisation becomes capable of operating it,” he said. “The biggest scaling risk is not shadow AI itself. It is fragmented accountability.”
Regional realities
The Asia Pacific region adds layers of complexity that make the transition harder than in many Western markets.
Data sovereignty and localisation rules vary widely across borders. “Privacy regulations often prevent organisations from using centralised global data stores, requiring them to build localised sovereign AI architectures that keep data within national borders,” noted Cronin .
Ashjari likens this to a “tug of war” between innovation and compliance. “Organisations need a platform that can enforce these rules at the software layer, regardless of where the physical hardware sits.”
Linguistic diversity is another distinct challenge. “Many enterprise AI applications rely on models trained predominantly on English-language data, which limits their reliability across Southeast Asian contexts,” said Sim.
But Singapore faces a hurdle all its own: an exceptionally high bar for trust.
“There is a distinctly high bar for trust when it comes to acting on AI outputs in this region,” Madgwick (above) observed. “Our research shows that 61 percent of Singapore respondents favour a human-in-the-loop approach, the highest of any region we surveyed.” This means organisations in the country must invest more deliberately in explainability and transparency before seeing meaningful adoption.
Legacy infrastructure and physical constraints compound these issues. In Singapore, limited data centre space and power regulations make traditional scaling difficult. “Asia is not one market. It is a collection of jurisdictions with different rules, infrastructure maturity, languages, data requirements and attitudes towards technology providers,” said Veloo.
Talent gaps add further pressure. About 54 percent of APJ organisations cite shortages of skilled professionals who understand both AI and business operations as their biggest barrier.
“The platform must automate the complex ‘Day 2’ operations, like patching, scaling and troubleshooting, allowing IT generalists to manage what used to require a specialist,” said Ashjari.
Bridging pilot and production
Technology providers are responding with approaches designed specifically to close this gap between experiment and enterprise readiness.
Couchbase’s AI Data Plane unifies operational data, vector search and persistent agent memory on a single platform to replace fragile custom integrations with built-in capabilities such as Agent Memory and native Model Context Protocol support.
“By bringing together operational data, vector search and persistent agent memory on a single platform, Agora eliminated sync delays caused by fragmented, multi-database architectures,” shared Cronin of one customer who achieved a 0.5-second conversational latency target and moved four workloads — outbound sales, interactive marketing, automated customer service, and physical AI — into live production.
Dell’s AI factory and AI data platform combine hardware, software and services to activate enterprise data at scale, while Dell Deskside Agentic AI lets teams run production-grade workflows locally, cutting token spend by up to 87 percent over two years. Working with Singapore’s Certis Group, Dell helped scale the Mozart orchestration platform across aviation, healthcare and smart city operations to bring local innovation to global markets.
F5 helps enterprises separate the application from the model, maintain consistent control across hybrid and sovereign environments, and build visibility into every API call and decision point. “Models will change faster than enterprise architecture. The architecture must be designed for that reality,” said Veloo.
Nutanix takes a “Sovereign AI first” approach that Ashjari said decouples workloads from location so organisations can develop in the cloud and repatriate to local data centres without disruption. For a major regional financial institution, this shifted sensitive inference to private infrastructure, meeting residency rules while improving latency and cost.
Alteryx addresses the gap from a different angle — the logic and trust layer that sits between raw data and AI outcomes. Its VURA framework serves as a production readiness checklist. “Visible — every employee should be able to answer where this answer came from and how we reached it. Understandable — explainable to non-technical users. Repeatable — the same question produces the same answer. Auditable — there is a traceable trail of ownership and oversight,” explained Madgwick, who cites FWD Hong Kong as a case in point. By starting small and building governed workflows, the insurer cut reporting times by over 95 percent and data preparation error rates by 80 percent, turning pilot success into enterprise-wide adoption.
The road ahead
What should enterprises do to turn promising tests into reliable business tools?
“Start with the outcome, not the technology,” advised Sim. “Identify what genuinely differentiates you, define clear success metrics, and assign accountable owners before committing to a solution.”
Cronin puts it simply. “Enterprises must shift their focus from model experimentation to infrastructure reliability by consolidating their fragmented data layer onto a single unified platform.”
Madgwick urges leaders to look beyond infrastructure entirely. “Organisations pour significant investment into selecting the right models and building the right infrastructure, but overlook the rules, risk appetite, governance requirements, and institutional knowledge that define what a good decision actually looks like.” His advice is to build a governed logic layer, keep business teams in the lead, and test readiness against the VURA principles before going live.
Ashjari urged leaders to “build a platform, not a silo. Focus on building a robust, hybrid infrastructure that gives you the choice of models and the freedom of location.”
Veloo’s advice cuts to the heart of the matter: “Do not scale the pilot. Scale the capability behind it.”
As organisations move from the hype of demos to the rigour of production, one truth stands out above all others. “AI models are rented commodities, while an organisation’s data architecture is a long-term operational asset,” said Cronin.
“The winners in enterprise AI will not be those that experiment the most. They will be those that industrialise learning the fastest,” Veloo concluded.
By Edward Lim
Top image: AI generated
Entelechy Asia 