AI has rapidly moved from experimentation to expectation. But while organisations are racing to deploy models and agents, many are discovering a harder truth. AI doesn’t create agility – it exposes whether it exists in your data layer.

According to research from Nexer Group, more than 80% of AI projects fail. From our perspective, it’s not the model that breaks, but the underlying data estate. Fragmented schemas, inconsistent structures and legacy artefacts make it difficult to produce output that is reliable, explainable, or even relevant.

When organisations merge or transform systems, they often uncover years of accumulated inconsistencies: custom tables, redundant fields, conflicting data definitions and so on. These issues can directly impact latency, reporting accuracy and the ability to generate real-time insights.

Left unresolved, these problems compound and at scale, they undermine accuracy and create operational risk right across the business. Decision-making slows, costs increase and the organisation loses the ability to respond quickly, which is exactly what AI is supposed to enable.

In practice, AI performance is bound by the quality and structure of the data it’s trained on and how well internal teams can collaborate. If data is incomplete, duplicated, or misaligned across systems, your models will inherit those issues and exacerbate them.

So, the question is: how do teams start to tackle underlying data challenges?

What business agility means in the age of AI

Business agility is often framed as an organisational capability. For example, how quickly can the business react and stabilise in the face of changing market conditions, legal requirements (M&As or cutoffs), or regional governance and industry regulations?

From a technical standpoint, business agility comes down to how quickly you can query, trust and act on your data. This means being able to find the right datasets among thousands if not hundreds of thousands of tables. It also requires being able to understand the lineage and context of the data and that’s before you add on the final layer of deploying AI models (LLMs or SLMs), ideally without compromising business critical processes.

If your data lake is too tightly coupled, poorly documented, or fragmented across incompatible systems, your organisation is constrained from the start. Those that suffer from these qualities need to prioritise the establishment of a clear, structured, flexible data architecture.

Preparing your data for AI

In 2025, Gartner analysts predicted that this year organisations will abandon 60% of AI projects due to a lack of AI-ready data.

As systems evolve, requirements shift, or new business objectives are realised, data often starts to accumulate without being organised. This is often called unstructured enterprise data, e.g. emails and logs that are harder to standardise and it only adds to the complexity of data migration projects to prepare for AI.

The first step to understanding your data landscape is to look at the whole picture – not just the primary systems, but everything connected to them. This includes legacy tables, unused schemas and data flows that may no longer reflect how your business operates.

From there, the goal is to design data structures that align with the current operational reality. This means standardising schemas, removing duplications and creating a clear framework. This allows you to ensure and monitor consistent data definitions across all your business systems.

Critically, this process isn’t just about cleaning data, it’s about making it reliable and usable. LLMs and SLMs both require data that is accurate and structured in a way that can be easily interpreted across use cases, e.g. internal agent chat queries.

Deciding what data to keep or archive

A persistent misconception about AI-readiness is that more data improves outcomes. In practice, often the opposite is true. According to research from Proofpoint, 41% of enterprises with more than 10,000 employees are managing more than a petabyte of data – 1,000 terabytes. Much of it is redundant or obsolete, including the ‘dark data’ that many organisations fail to even recognise, let alone use.

This is putting pressure on cloud services, not to mention the associated costs. Aside from the clear security risks involved, obsolete or redundant data will only introduce more noise into AI models. Training AI models on these datasets increases the likelihood of inaccurate output, model drift and higher computational costs.

This makes data curation a very critical next step once you have your structure in place. Businesses should strive to identify datasets that are relevant to current operations, consistent in structure and sufficiently recent to reflect real-world behaviour and use cases. A well-curated dataset will deliver better results than a large unfiltered one. The best thing to do is create a separate dataset to test it so other business-critical services aren’t disrupted.

Once teams know which dataset gives the best response or drives better organisational efficiencies, it is easier to decide which segments are needed most and which can be relegated to the archives.

Knowing where to standardise or customise

Standardisation is essential for scalability. It reduces complexity, improves operations and makes data easier to process across interconnected systems.

However, excessive standardisation can remove differentials that may benefit your business in areas like customer engagement, pricing strategies, or market-specific operations. So, what can businesses do to overcome this challenge?

Ideally, they should standardise at the system level while still preserving differentiation at the business logic level. In practice, this means standardising core processes that don’t change often, such as finance, supply chain and order management, while retaining flexibility in areas where that competitive advantage is most likely to be felt, like marketing or sales.

Taking this approach allows organisations to maintain clean, scalable data architectures without losing the nuance that drives business and brand value.

Building an AI-ready foundation

Becoming AI-ready isn’t a one-off transformation. AI highlights data problems and organisations that treat data as a byeproduct will struggle to scale AI effectively.

Business agility in the age of AI is about data integrity and how quickly you can access and act on it. If organisations lack clear ownership, consistent governance and continuous monitoring, all data issues will persist and AI will amplify them. Those that rise to these challenges will stand above their competitors.

By Peter Bosley-Sharpe, President EMEA at SNP Group.

Over the past 18 months, many organisations have experimented with AI through pilots and isolated initiatives. While those experiments have generated excitement and much speculation, they have also created a measurement challenge. The question is no longer whether AI works, but whether it will deliver the measurable return on the significant investments companies have made in data infrastructure, skills and systems.

Today, AI’s impact is often assessed through activity metrics – usage rates, prompts generated, or time saved – rather than whether it improves business performance. Yet, activity does not equal impact. If organisations want to measure AI’s real contribution, they must anchor it to business outcomes that define competitive position: win rate, customer retention, time-to-market, risk exposure, cost per case, or revenue growth. Therefore, measuring AI meaningfully requires a shift in thinking – from treating AI as a standalone tool to embedding it directly into how the business operates.

Redesigning workflows for the AI era

The main reason AI initiatives fail to scale and, subsequently, prove business value is that they are introduced into existing processes. AI tends to be added as a “bolt-on” step in a legacy workflow rather than being part of the redesigned process.

To unlock value, organisations must therefore rethink and redesign end-to-end workflows around AI rather than inserting it after the fact. I believe that the real opportunity lies in improving decision-making across the whole process – from data architecture to insight generation to operation execution.

When AI is embedded in business operations from the start, its impact becomes measurable through outcomes such as faster product launches, improved forecasting accuracy, reduced operational losses, or stronger customer retention. However, redesigning workflows in this way quickly exposes another challenge: ensuring reliable access to trusted enterprise data.

Building the data foundation for scalable AI

As I’ve outlined, AI pilots often rely on fragmented, project-based datasets, where the data often sit outside of the organisation’s data architecture. While those datasets may be useful for experimentation, they rarely provide the reliability or governance required for enterprise-wide deployment, compromising scalability.

To embed AI into core business processes, organisations must therefore adopt a ‘data product’ mindset. This means creating curated, domain-owned datasets with clear ownership, quality standards and built-in governance. The data products then become reusable assets capable of supporting analytics, operational systems and AI models across the organisation.

It is only when AI operates on trusted data products, instead of one-off extracts, that it can be reliably integrated into operational workflows, scaled across the enterprise and measured meaningfully. At that point, AI stops being an experimental capability and becomes a business asset.

But getting to that point requires another shift to happen. Data must be treated as a governed, enterprise-wide strategic asset rather than a by-product of IT systems or a collection of disconnected silos. The shift here isn’t about better AI models; it’s about building a unified data foundation that enables AI to drive durable competitive advantage.

Scaling up security in the AI era

As organisations start embedding AI into decision-making processes, data governance becomes critical. Scaling AI without robust upstream governance structures introduces significant legal, operational and reputational risks. Before AI influences business-critical decisions, it is imperative that organisations can ensure clear data ownership, enforceable access controls, lineage visibility, and compliance with regional and regulatory requirements.

Importantly, governance cannot simply exist as policy documentation. It must be technically enforced across the entire data estate so that organisations can confidently scale AI while maintaining control over how and what data is accessed and used. Without that level of governance, AI risks amplifying existing data fragmentation and compliance challenges rather than delivering enterprise value.

People first, every time

AI adoption also raises the important question of the relationship between human expertise and machine intelligence. Much of an organisation’s competitive advantage exists not in data, but in tribal knowledge held by humans – judgment, context and business understanding built over years of activity.

AI systems are only as strong as the data and the context they are given. While organisations should progressively codify that institutional knowledge into governed data products, metadata and documented business rules, we are still far from a point where AI can fully replicate that depth of human expertise.

Today, AI is enhancing human capabilities rather than replacing them – surfacing insights, accelerating analysis and providing decision support – while experts remain in the loop to validate, refine, apply judgment and ensure decisions reflect the broader business context.

As institutional knowledge becomes more structured and accessible, AI’s role will naturally expand. But in the near term, as I’ve outlined, the most successful organisations will be those that use AI to augment their workforce rather than attempt to automate expertise prematurely.

From experimentation to enterprise value

Organisations are now moving beyond AI experimentation, shifting the focus from novelty to measurable business impact. Achieving true AI business value requires embedding AI into how the business actually operates – supported by trusted data foundations, strong governance and workflows redesigned around better decision-making. It is my view, then, that AI should ultimately be judged not by how often it’s used, but by whether it delivers meaningful business results.

By Matt Fuller, Co-founder and Vice President of AI/ML Products at Starburst.

MIT reported that up to 95% of enterprise AI projects fail. Businesses are racing to adopt the technology and rushing execution in the process, significantly eroding ROI and potentially leaving thousands of businesses at risk of regulatory exposure.

Despite the urgency, there lies an invisible burden which is undermining progress. It is called AI debt and it may be the single greatest obstacle standing between AI ambitions and commercial reality.

AI debt is the accumulated result of incomplete digital transformation. It is every decision which meant legacy infrastructure was never fully retired, or the fragmented data siloes that were never unified. Any new platforms layered on top of such large and complex data sets that haven’t been organised properly, are creating complexity rather than clarity.

On their own, these shortcuts may have seemed pragmatic but collectively, they are now stalling innovation.

The risk of unfinished business

Recent analysis from McKinsey highlights the scale of the missed opportunity. Despite AI tools becoming commonplace today, 63% of organisations reported that they are still experimenting or piloting early-stage AI projects. This shows that most organisations are yet to embed AI deeply enough into their workflows and processes to capture its full value, estimated globally at between $2.6 trillion and $4.4 trillion.

The number one reason for this is years of bolt-on systems that have created tangled IT estates that slow decision-making and make rapid innovation nearly impossible. Running legacy and modern environments side by side inflates maintenance costs and introduces operational confusion which can result in failed implementations. These poorly executed migrations waste capital and introduce security and compliance risk, particularly under regulations such as General Data Protection Regulation (GDPR) and Digital Operational Resilience Act (DORA).

These consequences are measurable. Projects are delayed by months, which has a knock-on effect for budgets. Overall, the whole process inevitably grinds to a standstill because the data required to fuel AI models remains locked away in silos. Estimates suggest that between 50 and 70 per cent of enterprise data remains unconnected and inaccessible for advanced analytics.

The rise of AI debt

The push towards autonomous systems capable of independent decision-making is amplifying this risk of failure. While a majority of organisations plan to deploy AI agents in the near term, only a fraction have centralised their data or ensured their infrastructure can handle the projected surge in workloads.

The statistics are sobering. A recent report from Cisco found that fewer than one in five companies have fully centralised their data for seamless AI access. Over 60 per cent expect workloads to increase by more than 30 per cent within the next few years. Less than a third feel fully prepared to secure agentic AI systems against emerging threats.

Even the most digitally advanced firms are grappling with spiralling compute costs and persistent talent shortages in cybersecurity and AI engineering. In the same way that technical debt slowed software development in the 1990s and 2000s, AI infrastructure debt threatens to stall the current wave of transformation before it delivers meaningful returns.

At its core, AI debt is a data problem. AI systems amplify whatever they are trained on. If the data is incomplete or contextually degraded, the outputs will be flawed, often in ways that appear plausible but lack integrity. This phenomenon, sometimes described as AI slop, is not merely a technical nuisance but a commercial and reputational risk.

This occurs when organisations migrate or modernise without preserving metadata, lineage and governance and so its meaning is lost along with trust. In regulated industries, that erosion has legal implications. In competitive markets, it has revenue implications.

The path forward requires paying down the debt.

Paying off AI debt

At Cirata, we always advise that the best way to eliminate AI debt is to address fragmentation at its source. This means employing a system that can create a unified, interoperable data foundation that supports AI at scale. By decoupling data orchestration from underlying infrastructure, organisations can move, replicate and integrate data seamlessly across on-premises, hybrid and multi-cloud environments without disrupting production systems.

This approach delivers several strategic advantages. Automated data flows across clouds and platforms ensure models are trained and updated with the latest data. Businesses should always look to leverage open standards, such as Apache Iceberg, to prevent vendor lock-in and preserve long-term flexibility.

IT leaders should explore solutions to help them make sense of their data. By centralising governance and eliminating brittle integrations, organisations can feel confident they’ll be on the right side of AI project success. Most importantly, they can break the cycle of making short-term compromises that accumulate into long-term risk.

No algorithm will compensate for structural weakness

The promise of AI remains immense. Autonomous systems and generative models will continue to reshape industries. But no algorithm can compensate for a weak foundation. Just as a building requires structural integrity before additional floors are added, AI requires a unified and trusted data infrastructure before it can deliver sustained value. The organisations that thrive in 2026 and beyond will not be those that launched the most pilots. They will be those that had the discipline to eliminate their AI debt first.

By Stephen Kelly, CEO of Cirata.