Two decades of gold, buried under 2000 metres of granite

Why 2026 will be the year businesses finally unearth the real value of AI

If you are a service business with extensive domain experience, then this article might interest you. It shines a light on our journey so far. It contains the beginnings of a blueprint for your own company AI journey. I will diarize each step here so you can tag along for the journey and experience success yourself.

After two years of feverish experimentation, 2026 should see the AI era finally moving forward on the maturing scale . The early hype is giving way to something more disciplined - a recognition that while the tools are extraordinary, the real challenge lies in what we feed them.

For most companies, the gold of AI lies buried under 2,000 metres of granite: decades of unstructured data, scattered files, and uncaptured expertise. The difference now is that we're finally learning how to dig properly.

The billions invested in AI haven't been wasted. We are starting to build the muscle memory to understand what truly drives intelligent systems and also why so many pilots have failed to deliver. According to MIT research, 95% of enterprise AI pilots are still failing to achieve meaningful results, and businesses that rushed to deploy AI functions have actually seen a temporary productivity drop of 1.33 percentage points. But this isn't failure - it's calibration. It's the necessary excavation before the gold.

Think of it as building a sophisticated library system. You can't just throw books on shelves and expect people to find what they need. A proper library doesn't just store books; it catalogues them by subject, assigns each a unique identifier, cross-references related topics, and maintains an index that lets you find exactly the right chapter in exactly the right book within seconds. That's what we're building here, except instead of the Dewey Decimal System, we're using vector mathematics to map the relationships between ideas.

The granite problem ( We are here!)

I saw this up close while starting our own project of a building digital twin and AI agents for our own business. Despite clear use cases, early results were underwhelming. The AI couldn't make useful recommendations or replicate the nuance of nearly three decades of strategic experience. Not because the models weren't capable - but because our knowledge wasn't ready for them.

We were sitting on a geological layer of unorganised information: files buried in servers, overlapping archives, and decades of client work without structure or context. Before AI could think, we had to mine.

We culled 98,000 documents down to two groups. The one data set now has 10,000 documents around handling issues and crisis and the other data set has around 20,000 documents of more broad PR, Content and Communications background. We reduced 1.2 terabytes of historical content to 275 gigabytes of knowledge that actually mattered. Even then, that remaining quarter-terabyte needed categorising, tagging, and contextualising before AI can begin to detect meaningful patterns.

That's the granite. The drilling takes months, not minutes. From here, we're stripping out client names and sensitive information, converting everything into a format the AI can actually learn from, then running quality checks to make sure the system isn't learning our mistakes alongside our successes. We're also building the guardrails - protocols that determine who can access what, how the AI should respond in different situations, and how we audit its outputs before they reach clients.

Each preprocessing step serves a critical purpose. Extraction ensures readability. Cleaning removes noise that would confuse the model. Chunking creates the right granularity for retrieval. Indexing maintains provenance. Only after these foundations are solid can vectorization capture true semantic meaning. To extend the library analogy: extraction is unpacking the boxes, cleaning is removing damaged pages and duplicate copies, chunking is dividing books into chapters and sections, indexing is creating the card catalogue, and vectorization is training a librarian who understands not just where things are, but what they actually mean and how they relate to each other

Most mid-sized companies are in the same position as our team. Underneath the clutter lies extraordinary value - decades of insight, patterns, and hard-won expertise - but it's trapped inside chaotic file systems, inconsistent documentation, and tribal knowledge locked in people's heads.

Wipro's State of Data4AI report found that only 14% of business leaders believe their data maturity supports AI at scale, while 76% admit their data management capabilities can't keep up with business needs. Deloitte adds that 62% of executives cite data access and integration as their top obstacle to adoption. The issue isn't compute power or model quality - it's the missing connective tissue between data and institutional memory.

The excavation mindset

The companies seeing results now are the ones that approached AI methodically - as an excavation process, not a product purchase.

When AI is trained on clean, connected, contextual data, it becomes something more than a chatbot. It becomes a digital apprentice - one that recognises the company's unique tone, decisions, and logic.

This is where 2026 will be different. The conversation is shifting from "What tool should we buy?" to "What knowledge do we actually possess?" and "How do we make it usable?" The early adopters are realising that AI doesn't replace expertise but helps to scale it.

The data dividend

The return on doing this foundational work is exponential. Once the granite is cleared, the gold flows fast.

IDC estimates that companies lose $31.5 billion annually to poor knowledge sharing and disconnected systems. Simply reducing that friction can unlock immediate ROI - not through a new AI model, but through data readiness.

We're already seeing a divide emerge between businesses that treated data preparation as strategic and those that didn't. The former are integrating AI seamlessly into workflows - automating client research, generating insights, and supporting decision-making with confidence. The latter are discovering that without clean, organised knowledge, even the most advanced models produce noise instead of intelligence.

The great uncovering

The next 18 months will be defined by what might be called The Great Uncovering - a global effort by businesses to excavate their institutional knowledge before the AI era fully scales.

The winners will be the ones who:

• Treat knowledge management as a board-level function, not an HR afterthought.
• Make data governance and architecture part of their AI strategy from day one.
• Pair AI investment with data archaeology, extracting and structuring the wisdom already in their organisations.

 These aren't glamorous projects. They're slow, steady, and deeply human - involving conversation, curation, and the digitisation of decades of intellectual capital. But once done, the benefits compound.

When knowledge is structured and shared, AI can finally reflect the best of the business - not just what's online.

The gold ahead ( can we do it?)

The market is ready. The tools are ready. And the leaders who have the patience to drill through the granite will strike the richest knowledge seams first.

The gold was never lost. It was simply waiting for us to learn how to mine it. Our team are learning. Come along for the journey and see if we actually do strike gold. This really should be a reality TV show with all the ups and downs.