Build AI That Works: Fix Your Data Foundation First

IDC projects that AI and generative AI spending in APAC will reach US$110 billion by 2028, growing about 24% annually. That’s a lot of budget chasing a lot of ambition. But here’s the uncomfortable truth I keep seeing in Singapore: many “AI initiatives” are really procurement projects wearing an AI badge.

Most companies don’t fail at AI because they picked the “wrong model.” They fail because their data infrastructure can’t reliably feed AI, and their leadership teams aren’t aligned on what “ready” actually means. The result is predictable: pilots that stall, proof-of-concepts that never ship, and dashboards that look impressive but don’t change decisions.

This post is part of our AI Business Tools Singapore series—focused on practical adoption for marketing, operations, and customer engagement. The stance is simple: before you buy more AI tools, fix the foundation that makes those tools useful.

AI hype collapses when your data stays fragmented

If your data is split across on‑prem systems, multiple clouds, and edge devices, your AI will be slow, expensive, and brittle. That’s not a theory; it’s how AI systems behave in production.

In the source article, NetApp’s You Qinghong highlights the execution gap: leaders announce AI readiness while IT teams see infrastructure limits. I agree—and I’d add one more layer: even when IT can support an AI pilot, the operating model (who owns data quality, access policies, costs, and outcomes) often isn’t defined.

Here’s what fragmentation looks like in a typical Singapore business:

• Customer data sits in a CRM, but transaction history sits in an ERP, and web events sit with a marketing vendor.
• Ops data lives in spreadsheets shared over email, while IoT/edge logs are stored separately by a facilities team.
• Different business units buy different SaaS tools, creating multiple “sources of truth.”

AI systems don’t magically unify any of this. They just expose it faster.

Snippet-worthy reality: AI doesn’t remove data silos. It makes the cost of silos visible.

The “intelligent data infrastructure” that AI actually needs

An intelligent data infrastructure is a set of capabilities that makes data accessible, governed, and efficient—wherever it lives. It’s not only storage. It’s how you move (or don’t move) data, secure it, observe it, and serve it to AI workloads.

In practical terms for an AI program, you want three outcomes:

1. Unified access to data across on‑prem and cloud (without endless manual copying).
2. Security and governance that’s built into workflows, not bolted on later.
3. Elastic performance and cost control so training spikes don’t become a permanent bill.

That aligns with the article’s three pillars—and it’s also exactly what Singapore organisations need if they’re adopting AI business tools at scale.

What “unified access” means (without boiling the ocean)

Unified access means your teams can find, trust, and use data across systems with consistent controls. The biggest misconception is that you must centralise everything into one mega data lake. You don’t.

What works better (and ships faster) is a federated approach:

• Keep data where it makes sense (compliance, latency, cost).
• Standardise metadata, identity, and access patterns.
• Create reliable pipelines for the specific AI use cases you’re deploying.

If your marketing team wants an AI assistant that generates campaign insights, it needs consistent definitions of:

• customer segments
• product taxonomy
• conversion events
• revenue attribution windows

Without that, the assistant will confidently produce nonsense.

Why “AI-native” operations aren’t optional

AI workloads aren’t steady. Training is spiky. Inference is predictable—until you launch a new feature or run a seasonal campaign.

Singapore businesses feel this sharply in Q4 retail peaks, Chinese New Year promotions, and large-scale regional events. Your infrastructure needs to scale up for training and experimentation and scale down without waste.

A useful internal metric: time-to-first-production (from approved use case to deployed workflow). If this is measured in quarters, your bottleneck is usually data access + governance, not model selection.

Pillar 1: Modernise data architecture for AI tools that deliver ROI

Modernising data architecture means designing data flows around decisions, not around departments. If the business outcome is “reduce customer churn,” the architecture should prioritise access to churn signals across support tickets, usage logs, billing, and marketing engagement.

A concrete scenario (common in Singapore’s B2C and telco-adjacent firms):

• Support tickets stored in one system
• App usage in an analytics platform
• Billing in ERP
• Campaign history in a marketing tool

If these systems can’t be joined reliably, your churn model will be trained on partial truth. You might get a decent AUC in a notebook, but the moment you deploy, the model can’t access the features it was trained on. That’s how “promising pilots” die.

What I’ve found works: start with 1–2 high-value use cases and build a reusable data layer (identity, pipelines, governance). Don’t build bespoke plumbing for every team.

Practical checklist for this pillar:

• Define a canonical customer ID and mapping rules across systems.
• Build a feature-ready dataset (not just raw dumps): timestamped, deduped, documented.
• Set SLAs for freshness (e.g., “web events within 15 minutes”).
• Instrument data quality: missing values, schema drift, and pipeline failures.

Pillar 2: Security and governance must start on day one

If you add governance after the pilot, you’ll rebuild everything under pressure. In Singapore, where PDPA expectations and industry regulations are real constraints, “we’ll secure it later” is not a plan—it’s an expensive delay.

The article calls out a zero-trust mindset and AI-specific governance. This matters because AI introduces unique failure modes:

• Prompt injection and data exfiltration via chat interfaces
• Training data leakage into logs or third-party tooling
• Model output risk (hallucinations presented as facts)
• Policy drift where a model’s use expands beyond its approved scope

A governance model that doesn’t slow teams to a crawl

You don’t need a 40-page policy to start. You need crisp rules that teams can follow.

A workable baseline:

• Data classification: public / internal / confidential / regulated.
• Access control: role-based with approval workflow for regulated data.
• Lineage: track where training data came from and which version trained which model.
• Human-in-the-loop for high-stakes outputs (pricing, credit, medical, legal).
• Red-team testing for GenAI tools before broad rollout.

For regulated industries (finance, healthcare), the source mentions the ability to train across distinct locations without moving sensitive data. In practice, that can mean architectures and processes that support data residency, controlled computation, and auditable access—so your AI program can grow without triggering compliance emergencies.

Snippet-worthy stance: If you can’t explain where the data came from, you can’t defend the AI outcome.

Pillar 3: Business–IT alignment is the real AI multiplier

The fastest way to waste an AI budget is to treat IT as an order-taker. AI changes processes, incentives, and risk. That requires shared ownership.

The gap described in Singapore—CEOs declaring readiness while IT sees unprepared infrastructure—usually comes from two issues:

1. Different definitions of success (PR headline vs operational KPI)
2. Different risk tolerance (move fast vs don’t break compliance)

Run an AI readiness assessment that leaders can’t ignore

A useful readiness assessment isn’t a technical scorecard. It’s a joint document signed off by business and tech leaders.

Include:

• Target use cases (ranked by value and feasibility)
• Data availability (systems, owners, quality score)
• Infrastructure gaps (storage, compute, network, observability)
• Security requirements (PDPA, sector regulations, vendor risk)
• Operating model (who owns model performance and incidents)
• Cost model (training spikes, inference baseline, vendor spend)

Then set one rule: no “AI tool” purchase without a data path to production. If you can’t describe the data inputs, refresh rate, and access controls, it’s not ready.

Cost control is a strategy, not a finance exercise

AI bills can surprise teams because training is bursty and experimentation is messy. The answer isn’t “stop experimenting.” The answer is to architect for cost visibility:

• Tag compute and storage by use case
• Separate sandboxes from production
• Use quotas and approvals for large training runs
• Track cost per prediction or cost per automated workflow

If your CFO asks, “What did we get from this spend?” you should be able to answer with numbers.

What Singapore companies should do in the next 30 days

The timeline has compressed—what used to take years now has to show progress in weeks. If you’re trying to roll out AI business tools in Singapore (customer service copilots, marketing content generation, forecasting, fraud detection), these steps move you from talk to traction.

1. Pick one use case with a hard KPI (e.g., “reduce average handling time by 15%” or “improve forecast accuracy by 10%”).
2. Map the data end-to-end (source → pipeline → storage → model → app). Assign owners.
3. Fix identity and definitions (customer ID, product taxonomy, event names).
4. Implement minimum governance (classification, access control, logging, lineage).
5. Design for production from day one (monitoring, rollback plan, human review where needed).

If you do only one thing: stop measuring AI success by the quality of the demo. Measure it by the reliability of the data feeding it.

“When leadership teams are not aligned on both opportunity and challenge, AI initiatives either stall or deliver disappointing results.” — You Qinghong, NetApp

Where this fits in the AI Business Tools Singapore series

AI tools for marketing, operations, and customer engagement can produce real gains—but only when the data foundation is simple, secure, and sustainable. Singapore’s push toward AI-ready infrastructure (including major public investment) is a signal: capability is becoming a baseline expectation, not a differentiator.

The differentiator is execution discipline. Companies that modernise data access, embed governance early, and align leadership will ship more AI use cases per year—and they’ll spend less cleaning up after rushed pilots.

If you’re planning your 2026 roadmap, the question to ask your team isn’t “Which AI model should we use?” It’s this: What would it take to trust our data enough to automate decisions with it?