The hidden cost of untrusted data at AI scale

My daughter recently came home talking about the Great Fire of London. Like many school stories, it started with the dramatic details: the bakery on Pudding Lane, the match-stick dry wooden buildings packed tightly together, the flames spreading from street to street. But what stayed with me was not just the fire itself. It was how preventable so much of the damage sounded.

London had rules. It had people who understood the risk. It had warning signs everywhere: crowded streets and touching structures, timber-framed buildings, and open flames. But having rules and recognising risk is not the same as having control to prevent it. And having people responsible for parts of the system is not the same as having a system that works when pressure hits.

That is the part that feels familiar today.

Many organisations are now racing to scale AI. They have data policies, governance frameworks, catalogues, dashboards, and committees. On paper, the foundations look solid. Yet, Gartner found only 14% of IT leaders are very confident their data is ready for AI¹. Despite years of investment in data platforms and governance initiatives, many organisations still struggle to translate visibility into trust.

When data is incomplete, inconsistent, poorly governed, or not trusted, AI does not amplify value. It amplifies risk.

The cost of untrusted data at AI scale is often hidden at first. It shows up as failed pilots, inconsistent outputs, regulatory exposure, duplicated effort, and a quiet loss of confidence in decision-making. Then, as AI adoption accelerates, that cost compounds quickly.

This is where the Great Fire analogy becomes useful: not because data risk is the same as physical risk, but because both expose the same weakness, which is visible warning signs without effective enforcement.

When AI scales, so do data problems

Most organisations have already invested heavily in data visibility, governance, and analytics. Yet many still struggle to turn those investments into trusted, operationalised data.

The gap is simple: knowing your data is not the same as controlling it or fixing it. Documenting a policy is not the same as enforcing it. Seeing a risk is not the same as preventing it from spreading.

Common issues include:

• Data quality that varies across systems, teams, and business units
• Governance policies that exist in documentation but are not consistently applied
• Data moving across fragmented, federated environments with different standards and controls
• Teams lacking visibility into how data issues affect real business outcomes

At a small scale, these problems can be worked around. Someone checks a report manually. Someone reconciles a spreadsheet. Someone asks another team whether a field can be trusted.

At AI scale, those workarounds break down. Small data issues can quickly become enterprise-wide risks.

If a machine learning model is trained on incomplete, inaccurate, unrepresentative, or inconsistent data, the outputs will reflect those weaknesses. In use cases like credit decisioning, fraud, risk management, regulatory reporting, or customer operations, the consequences can include financial loss, poor customer outcomes, regulatory scrutiny, and reputational damage.

And once trust breaks, people compensate in predictable ways. They slow down. They validate manually. They duplicate work. They stop relying on the model. Eventually, AI initiatives stall or are cancelled, not because the ambition was wrong, but because the data foundation was not trusted enough to support it.

The real business impact

Untrusted data does not just affect model accuracy. It affects operational efficiency, regulatory compliance, customer outcomes, and ultimately business performance. Organisations see:

When organisations get their data foundations right, the impact can be substantial.

In one example with a UK public-sector bank, better data quality automation through Aperture Data Studio saved around 1,500 FTE hours per year, reduced data processing time by more than 85%, and cut manual data corrections by 80%. Importantly, those gains did not come from collecting more data. They came from improving the quality, consistency, and governance of the data already available.

That is the real difference between data being visible and data being operationally trusted. One creates awareness. The other removes friction, reduces cost, and gives teams confidence to act.

In high-stakes environments such as financial services, risk management, and compliance reporting, this is not theoretical. Every decision must stand up to scrutiny. Every model output needs a defensible foundation. Every data-driven process needs to be explainable, repeatable, and controlled.

When trust is uncertain, every data decision starts to feel like a gamble.

Why traditional approaches fall short

Many organisations still treat data quality and governance as separate, often passive activities.

Catalogues describe data but do not control it. Without active management, they can quickly become stale or incomplete.

Policies define standards but are not always embedded into operational workflows.

Monitoring identifies issues, but often after those issues have already affected downstream systems, reports, models, or decisions.

This is the modern enterprise version of having fire rules posted on the wall while the city is still built from dry timber.

The problem is not that organisations lack intent. Most have invested heavily in data management. The problem is that intent does not scale unless it is operationalised. AI systems need continuously reliable inputs. Distributed data environments need controls that work across systems, teams, and platforms. Business users need confidence that the data they rely on is fit for purpose before it reaches the point of decision.

To scale AI successfully, trust must move from static documentation into active execution.

Operationalising trust in real time

To reduce risk and realise AI value, organisations need to rethink how they manage data.

Trust cannot be something that is checked only at the end of the process. It needs to be embedded into the workflows where data is created, transformed, governed, shared, and used.

That means organisations need to:

• Validate and standardise data at the point of capture
• Continuously monitor quality and business impact
• Trace data lineage to understand downstream effects as data changes
• Apply governance policies automatically across systems
• Give business and technical teams a shared view of data risk and trust 

This shift turns trust into something operational, measurable, and scalable.

It also changes the culture around data. Instead of treating quality and governance as compliance exercises, organisations can make them part of how decisions are made every day.

Building AI-ready data

At its core, AI success depends on the quality and reliability of its inputs.

AI-ready data is:

Reliable, accurate and consistent
Structured and enriched
Governed and traceable
Continuously validated
Fit for purpose
Understandable to the teams that use it

Without these foundations, AI models can still produce outputs. But those outputs may not be reliable enough to support real-world decisions.

That distinction matters. A model can generate an answer without the answer being trusted. A dashboard can display a metric without the metric being defensible. A workflow can automate a decision without the organisation being able to explain it.

AI-readiness is not just about having more data. It is about having data that can be trusted when the decision or process matters.

Designed for complex, federated environments

Modern enterprises do not operate in a single system.

Data is spread across cloud platforms, legacy systems, third-party applications, regional teams, business units, and specialised operational environments. Internal organisations often choose the software and vendors that best meet their own needs.

That reality is not going away.

Trying to centralise everything can introduce cost, complexity, vendor lock-in, and performance trade-offs. Well-intentioned simplification efforts can also lead to frustration, low adoption, and inefficient workarounds.

A more effective approach is to apply trust where data already lives.

That means improving quality, governance, and control across federated environments without forcing every team into the same system or every dataset into the same platform. It allows organisations to maintain flexibility while improving consistency, accountability, and confidence.

In other words, the answer is not always to rebuild the whole enterprise architecture. It is to apply consistent controls across the environments where data already lives.

From data effort to business outcomes

The goal is not better data for its own sake. The goal is better business outcomes. By operationalising trusted data, organisations can: