Beneath the Surface: The Technical Foundations Every AI Strategy Needs

Article Summary:

Most AI conversations only address the visible part of the iceberg — automation, faster decisions, better products — while ignoring the technical foundations that determine whether any of it actually works.
According to MIT, 95% of enterprise GenAI pilots deliver no measurable financial impact. Not because the technology fails, but because it’s built on an unprepared foundation.
Three foundations separate AI that scales from AI that stalls: Cloud-native architecture built to flex, data that's genuinely trustworthy (not just accessible), and security/governance robust enough to withstand regulatory scrutiny.
Legacy systems and technical debt are already the biggest blocker with 68% of IT decision-makers saying legacy systems are preventing AI adoption, and fragmented systems make implementation delays 30% more likely.

Visible Ambition Vs. Hidden Reality

The iceberg metaphor has been around for decades, and yet it almost feels as though it was tailor made for Generative AI, specifically for its application in modern businesses. Most conversations, strategies and ambitions for AI refer to the visible section of the iceberg. Everything visible represents the benefits of AI that businesses want to harness: automation, faster outcomes, and the creation of all-around better products. But just like how 90% of an iceberg remains obscured by waves, providing buoyancy and balance to stop it from sinking, the unseen, unconsidered and – often unappreciated – elements of AI are ultimately what make the visible sections possible and successful.

It’s for this exact reason that so many AI initiatives underdeliver. It’s not the model that everyone can see and everything it’s capable of, it’s what’s holding it up beneath the waves. According to MIT, 95% of businesses’ GenAI pilots deliver no measurable P&L impact. So how do the 5% manage it? Simply put, they don’t neglect the obscured bulk of the AI iceberg.

What Businesses Want from AI

Ask most executive teams what they want from AI and the answers tend to cluster around four themes:

Smarter Automation: Offloading repetitive, rules-based work — such as bordereaux reconciliation — so people can focus on higher-value judgment calls.
Faster Decisions: Compressing time between a question being asked and an answer being actioned, whether that's underwriting a policy, approving a claim, or repricing a product.
Intelligent Products: Embedding predictive or generative capability directly into what the business sells — dynamic risk pricing, or fraud detection built into the product itself — rather than treating AI as a back-office tool.
Better Customer Experience: Personalisation and responsiveness that make a policyholder's claim or a borrower's application feel immediate, not queued for manual review.

These are legitimate and well-reasoned goals, avoiding hype and mapping to a genuine commercial outcome that any board would sign off on. But these points also represent where most conversations about AI stop. These are the discussions about AI that fit neatly onto a strategy slide but fizzle out before serious conversations about logistics can begin.

The Waterline: Why AI Doesn't Operate in Isolation

So what are the poor foundations that can cause the AI iceberg to sink? It can ultimately be summed up by the intention of the foundation the AI is built on. If the existing environment wasn’t built with AI in mind, then this mismatch can cause major problems.

Data Silos

Data is one of the most important foundations for AI. Most organisations store data in separate systems that don’t communicate well with each other. A CRM here, a finance system there, claims platform somewhere else, with spreadsheets filling gaps. Each department usually has oversight and trust in their own data, but there is rarely a single, reliable view across all of it. This is the most common reason AI pilots that work in a demo fall apart in production: the demo used a clean, hand-picked dataset while the real thing requires pulling from a messy, siloed environment.

API Sprawl

These data issues are compounded by how the silos are connected. Many organisations have grown their technology estate through years of one-off, point-to-point integrations: System A wired directly to System B, then to C, with no shared standard underneath. Every new capability, including AI, then requires its own bespoke integration project before it can access the data or systems it needs. What should be a quick connection becomes a lengthy one.

Monolithic Architecture

Underneath both points sits architecture. AI workloads place erratic demands on computing power and older, monolithic systems built to run steadily rather than flex, generally can't scale to the needs of the AI. All of this is further compounded by technical debt: the accumulated shortcuts, ageing platforms and undocumented systems every long-running business carries, each add cost and delay to anything new that needs to connect to it.

Pega reported found that 68% of IT decision-makers say that legacy systems and technical debt are actively preventing their organisations from adoption modern technology like AI. McKinsey also noted that companies with fragmented or legacy systems are 30% more likely to experience AI implementation delays, specifically due to failures integrating with modern data platforms. If there’s nothing holding up your AI iceberg, it’s bound to sink fast.

But what steps are needed to ensure that your business has a solid foundation to build its AI ambitions on top of?

The First Foundation: Architecture and Infrastructure

Getting this foundation right starts with a genuine shift to cloud-native architecture. Many businesses will cut corners by simply rehosting their existing systems in the cloud without any redesign. With a concentrated shift to the cloud, it enables building around microservices and API-first design, meaning that new AI capabilities can plug into existing systems through standardised connections rather than triggering a full rebuild each time. It also enables event-driven architecture, allowing systems to respond to activity in real time; containerisation for consistent, scalable deployment; and orchestration layers that manage how everything communicates and recovers automatically. If thoughtfully assembled, this shift to cloud allows AI to scale sharply when it needs to — be that during training or interference spikes — and idle efficiently the rest of the time.

As AWS experts, Engineering Insights can assess this foundation against the AWS Well-Architected Framework where relevant, evaluating architecture for reliability, security, and cost-efficiency before any AI conversation begins. This kind of structured assessment — regardless of platform — is one of the key differentiators between AI that scales cleanly into production and AI that stalls in the pilot stage, which is why it's one of the first things we check with our clients.

The Second Foundation: Trusted, Governed Data

The second foundation is closing the gap between data being both accessible and trustworthy. This means ensuring the quality, lineage and governance of the data: where it came from, whether it has remained accurate along the way, and whether access classifications and ownership is clear. It also means genuine AI-readiness: structured and unstructured data pipelines, feature stores for reusable model inputs, and vector databases for retrieval-augmented generation use cases.

This is a principle we've built into our own delivery work. For our reinsurance clients, we built the Bordereaux Intelligence Engine to ensure that all forms of disparate data arriving can be trusted the moment it lands. There’s no point starting analytics or an AI layer if your core data can’t be trusted.

The Third Foundation: Security, Governance, Resilience and Explainability

This final foundation is about ensuring that your ambitions for innovation don’t run away with you and risk exposure and breach of compliance. Explainability of AI models are no longer optional, especially with the EU AI Act now formalising this for high-risk systems, requiring documentation and audit trails to ensure full compliance. The moment an AI model touches sensitive data, security expands exponentially as every integration point becomes a potential vulnerability. Businesses must also be more aware than ever of their resilience to containing an AI failure: is it contained, or does it silently work its way into business-critical decision making? As well as this, businesses must consider compliance, be it from the EU AI Act or sector-specific model risk guidance, to ensure that they can reconstruct and justify AI-assisted decisions after they’re made.

How Engineering Insights Builds These Foundations

We work from our tried and tested 5D framework — Discover, Define, Design, Develop, Deploy — to put the above into practice and stop our clients’ icebergs from capsizing. Discovery sessions asses your existing architecture, data, and governance landscape against the three foundations. Define turns findings into a comprehensive roadmap that sequences investment rather than chasing the most visible and trendy AI use cases. Design builds cloud-native architecture and the data foundations needed to support it, all while keeping governance frameworks and security at the front of mind. Develop implements iteratively, with AI accelerating delivery without compromising the governance from the design stage. Deploy launches with the monitoring and resilience needed to catch issues before they risk exposure. Each stage works to close the gap between AI ambition and AI that works in production, building the enormous foundation to keep your AI iceberg floating proudly above the waves.

Is Your Technology Ready for AI?

The organisations getting real value from AI aren't the ones with the flashiest use case, they're the ones who built what's underneath it first. Before your next AI initiative, it's worth asking honestly what's really supporting it. Talk to our team about assessing your foundations.