From Pilot to Enterprise-Wide: Scaling AI Across the Organization

In an era where every organization is accelerating its AI Transformation, a shocking recent statistic reveals that 46% of AI pilots are scrapped before they ever reach production.

Why do so many global organizations get stuck in "Pilot Purgatory"—an endless vacuum of experimentation that ultimately leads to abandonment? More importantly, how can these enterprises evolve from running small experiments to building an "AI Factory" capable of driving impact across the entire organization?

In this article, we summarize Key Insights from the session "From Pilot to Enterprise-Wide: Scaling AI Across the Organization" at SCB 10X AI-Volution. We dive deep into the perspectives of Kees Lemmens, CTO (ASEAN) of Microsoft, moderated by Kasima (Aom) Tharnpipitchai, Head of AI Strategy of SCB 10X.

The 6 Pitfalls Keeping Organizations in “Pilot Purgatory”

When discussing the 46% failure rate, Kees Lemmens offered a striking observation: "I think the number is actually on the low end." The shock isn't just in the volume of failures, but in how "remarkably consistent" the drivers are across different enterprises. Regardless of the organization, those stuck in Pilot Purgatory almost always fall victim to these six causes:

• No Business Hypothesis: Projects start as "tech experiments" without a proper business hypothesis or real impact analysis. Because they don't necessarily solve a real business problem, they fail to secure funding or sponsorship.
  
  
• Bad Data: Many pilots leverage sandbox environments or batch extracts, which "just doesn't match today's reality." When moving to production requires secure, low-latency connections to core systems, the pilots fail because the data foundation isn't there.
  
  
• Operation Ops: Most pilots are just "scripts put together" lacking observability, rollbacks, or retries. This leads to a "brittle handoff" to engineering and operations teams who cannot support or own the solution, so it never sees production.
  
  
• Compliance Crisis: Especially in banking, pilots often lack audit trails, PII protection, or model explainability. Because the "depth with the business" regarding human-in-the-loop processes hasn't been worked out, compliance teams red-flag the project, halting progress.
  
  
• Siloed Efforts: Pilots are often "bespoke, isolated projects" with no "AI Factory" behind them to drive reusability. As you try to scale isolated projects, they become more complex and costly, driving abandonment.
  
  
• Trust & Adoption: Users aren't incentivized or trained to leverage the system. Crucially, "If you don't trust the output, you will not trust the system, and when you don't trust the system, you will not use it."

Case Study: When "People" Are the Barrier to Scaling

Kees emphasized a key metric: 70% of implementation challenges are people and process-related, not technical. He shared a real-world example of a European bank attempting to automate heavy, document-intensive work like regulatory summarization and compliance checks. The result was a valuable lesson:

• The Start: The initial pilots were impressive, delivering a summarization of hundreds of pages in seconds.
  
  
• The Fail: When rolled out to staff, adoption stalled. Employees didn't trust the output and resorted to manual
  double-checking. Furthermore, they feared the efficiency project meant they were "about to lose their job."
  
  
• The Pivot: The bank shifted from a POC mindset to an "AI Factory" mindset.
• They adopted Azure AI Foundry as a central orchestration and governance layer.
• They shifted the narrative from "replacing humans" to designing for "humans-in-the-loop."
• They built systems with traceable citations and allowed employees to drive corrections via a feedback loop.

The Result: Once employees saw the system was transparent and traceable, they shifted from skeptics to "AI Champions," driving peer-to-peer training and widespread adoption.

"Most of these AI initiatives don't fail because of model quality; they fail because the organization isn't ready to use and embrace it." -Kees Lemmens, CTO (ASEAN) of Microsoft

Integration Shock: How to Survive Connecting AI to Legacy Systems

One of the hurdles Kees identified as "one of the most underestimated challenges" is taking a pristine cloud-based pilot and connecting it to live legacy systems like Core Banking or ERP. He calls this "Integration Shock."

These legacy systems were not designed for the "hot path data flows" required by real-time AI, leading to system overloads and quality issues. To solve this, Kees recommends a three-step approach:

1. Stabilize Data Foundation: You must identify your "gold data"—where it resides and what the AI can trust. Tools like Microsoft Fabric help create consistent data products with clear lineage.
   
   
2. Standardize Connection: Stop writing point-to-point scripts. Instead, create management across reusable connections, turning data into a "governed service layer."
   
   
3. Scale with Orchestration: Bring everything onto AI Foundry to get unified modeling and service telemetry. This allows you to "orchestrate around" legacy systems, giving you the necessary observability.

The Future Model: From Centralized to "AI Factory Mesh"

Historically, organizations started with a central AI team doing everything. However, Kees notes, "I can't imagine a large bank pushing out hundreds of use cases through that model."

He proposes the organizational model of the future: the "AI Factory Mesh."

• The Hub: Defines the guardrails—governance, security, evaluation, and responsible AI policies.
  
  
• The Spokes: Domains like Risk, HR, or Retail Banking have their own empowered AI teams building domain-specific Copilots or agents under the central policies.

This federated model allows for innovation speed while maintaining the safety and control required for sustainable AI Transformation.

Microsoft’s Next Steps: Tools to Watch

To help enterprises bridge the gap from pilot to production, Kees revealed how Microsoft is shifting focus from just "models" to "system trust," specifically through the Microsoft Agent Framework:

1. Microsoft Agent Framework: The Bridge to Enterprise Grade This unifies two key frameworks—Semantic Kernel (for developers) and AutoGen (for experimental multi-agent orchestration)—under Azure AI Foundry to solve classic enterprise problems:

• Make the Invisible Visible: Uses OpenTelemetry to trace end-to-end performance, including latency, lineage, and content safety.
• Resiliency: Adds state management, retries, and rollbacks to prevent pilots from crashing when an API breaks or a schema changes.
• Governance by Default: Embeds "policy-as-code," RBAC, and content moderation directly into the framework, ensuring every agent built is safe and compliant.

2. Looking Ahead: Ignite and Beyond Kees teased upcoming features designed to accelerate enterprise adoption:

• Agentic Patterns & Models-as-a-Service: More complex agent behaviors and flexible model access.
• Fabric & Foundry Integration: Deeper connections between the data platform and the AI platform.
• AI Factory White Paper: Kees and his team are developing a comprehensive playbook on how to set up an AI Factory specifically for the ASEAN region.

Conclusion: Escaping Pilot Purgatory isn't about building the most models; it is about creating "Clarity" of purpose, "Confidence" through transparent governance, and a "Community" of people ready to change. As Kees concluded: "Trust is more than just providing transparency... it's really helping the user... ultimately feel comfortable with the outcomes."

Watch the full session at https://www.youtube.com/watch?v=sDiYniIKLV4