When the Data Is Broken - And the Work Still Needs to Get Done

The dashboards disagree, the reports conflict and workarounds are everywhere. Some numbers can be trusted directly, others need a human check before they mean anything. Despite all of this, the work still has to get done.

The assumption underneath most conversations about AI readiness and data strategy is that the data exists, that it is findable and is coherent enough to act on. The research suggests that assumption deserves closer examination.

What makes this moment different is not fragmented data- organisations have navigated imperfect information for decades. What has changed is AI. Because people compensate for broken systems in ways AI cannot.

A necessary boundary

There are environments where imperfect data is simply not acceptable where a single inconsistency can have irreversible consequences and the only acceptable answer is better data. This piece does not speak for those environments.

This is for the rest where imperfect data is the normal condition, where perfection is not coming and the work cannot wait. How to navigate that honestly and still make decisions good enough to keep things moving.

What the research actually shows

Gartner's cross-industry research estimates that poor data quality costs organisations an average of $12.9 million annually as lost revenue, operational inefficiency and compliance costs. This does not include the costs of slower decisions, eroded trust and missed opportunities.

The IBM Institute for Business Value's 2025 report found that 43% of chief operations officers identify data quality as their most significant data priority and over a quarter of organisations estimate losses of more than $5 million annually as direct results. IBM is unambiguous about what this means for AI:

"AI systems inherit and amplify data quality issues. When that data is inconsistent, incomplete, biased or outdated, both models and the agents built on top of them are less accurate and prone to spreading issues at scale."

A 2026 Forrester Consulting study of 310 senior IT decision-makers found that 85% say fragmented data sources and knowledge systems must be unified for AI to succeed and 83% said that the challenge will grow harder as more AI is layered on top of existing infrastructure.

The problem is not static. It compounds.

The fragmentation spiral

When data fragmentation reaches a threshold when dashboards disagree, reports conflict and the same question gets different answers depending on who you ask, something happens to organisational behaviour that compounds the technical problem.

People stop trusting the data. They develop workarounds and rely on the person who knows the real number rather than the system that should hold it. Because the workarounds work well enough, the pressure to fix the underlying data problem reduces. The system deteriorates and the workarounds become more elaborate.

AI does not inherit the workarounds, it inherits the data. When AI is deployed into a fragmented environment, the human compensation that made that environment functional are not part of what AI reads.

For years, organisations survived fragmented systems through human judgment. AI inherits the systems - not the judgment that made them usable.

Why this is a knowledge management and AI governance challenge

Data fragmentation is not just a technical problem for the data team. When data is fragmented, the knowledge built on top of it inherits those qualities. A knowledge system is only as reliable as the data it draws from. Fragmented data produces fragmented knowledge - content that looks and treated as authoritative, while the inconsistencies beneath it remain invisible.

You cannot govern what AI produces without governing what AI consumes. A governance framework that does not extend to the quality, consistency and ownership of the underlying data is not governing the AI. It is governing the surface while the foundation remains unexamined. It is one problem with three faces and the organisations addressing all three from a single governance conversation are the ones the research identifies as managing AI effectively.

What to do- five things the research points to

1. Inventory before you fix

   The instinct is to start fixing -deduplicating, migrating, standardising. The research suggests this compounds the problem. IBM found only 26% of CDOs are confident their organisation can effectively use unstructured data. Know what you have first.

2. Name an owner for every data source informing AI

   Gartner consistently finds that absence of clear ownership is the most common governance failure. Accountability without a name is not accountability. Ownership without the time, direction and resource to act on it is accountability in name only.

3. Distinguish broken data from incomplete data

   Broken data actively misrepresents reality, acting on it produces wrong outcomes regardless of how efficiently you process it. Incomplete data is manageable with honest acknowledgment of what is missing. These are different problems requiring different responses.

4. Position human judgment where uncertainty is highest

   Forrester's 2026 research found 45% of IT leaders cite missing organisational context as the primary cause of AI underperformance. Humans need to be positioned where data uncertainty is highest — not to review AI outputs at the end but to provide the context that makes data usable in the first place.

5. Govern continuously

   As ECI Research / Starburst put it directly:

   "AI isn't being held back by models. It's being held back by data that is fragmented, inconsistently defined, and difficult to govern across systems."

   Data that was governed twelve months ago and has not been reviewed since a restructure is not governed. Governance has no completion date.

The question this article does not fully answer

The research on data fragmentation is not new. The mitigation steps have been recommended consistently for at least a decade. If knowing the problem and the mitigations were sufficient, the problem would be smaller by now but it is not. The harder question is why organisations consistently do not do what's required and what needs to change for that to be different.

The conclusion

What AI has changed is not the existence of fragmentation but the visibility of it.

For years, human judgment quietly held fragmented environments together. Experienced practitioners knew which systems to trust, which inconsistencies to challenge and where institutional knowledge lived outside formal structures.

AI exposes what was previously being managed informally. It does not know which spreadsheet the team actually trusts. It cannot detect the undocumented context that experienced employees apply automatically. It works from what the organisation formally captures.

Fragmented data no longer remains a hidden operational inconvenience. It becomes visible decision risk.

The organisations that navigate this well will not be the ones that achieved perfect data. They will be the ones honest about the condition of their systems, clear about where uncertainty exists and disciplined enough to build governance, accountability and human judgment into how decisions are made.

Human expertise was compensating for broken systems all along.

References

1. IBM Institute for Business Value, January 2026 — The True Cost of Poor Data Quality

2. Gartner — Poor data quality costs organisations an average of $12.9 million annually

   Verified via Dataversity(full Gartner report requires subscription).

3. Forrester Consulting / Simpplr, Q1 2026 — AI Highlights the Limits and Potential of the Digital Workplace

4. McKinsey - State of AI 2025

5. DATAVERSITY- Trends in Data Management 2025

6. ECI Research / Starburst, 2026 — Enterprise AI Data Governance

Part of KnowledgeNova Insights