Neither is a great model – here’s what actually moves the needle.
Most organizations believe they have solved their data problem. Records are cleaner than they were three years ago. Duplicates are down. Core domains are more structured. By traditional standards, this should be enough to support advanced analytics and AI. Then AI is introduced, and confidence drops almost immediately.
Outputs feel inconsistent. Teams hesitate to act on recommendations. Leaders start questioning the results. Which forces the uncomfortable question: if the data is ‘good,’ why doesn’t it work?
Because AI doesn’t expose bad data. It exposes misaligned systems. And cleaning data is not the same as aligning systems.
“The model becomes the visible failure point. But the model is not the root cause. The system is.”
The Investment Is Real and So Is the Gap.
Enterprise AI investment reached an estimated $252 billion in 2024. Against that, MIT’s Project NANDA finds that only 5% of AI initiatives produce measurable returns. Eight in ten companies cite data limitations as the primary roadblock to scaling agentic AI, per McKinsey. And only 7% of enterprises say their data is completely ready for AI – a figure that should reframe every conversation about AI readiness.
These are not early-stage problems. Organizations investing heavily in data quality, in governance programs, in MDM initiatives are still hitting this wall. The issue is not effort. It is what the effort is directed at.
Data quality improves individual datasets. It does not fix how those datasets behave across the enterprise. That is a different problem – and it is the one that AI exposes.
Where AI Actually Breaks Down
Traditional analytics tolerates inconsistency because people compensate for it. A data analyst reconciles a discrepancy. A manager applies judgment. A team agrees on which number to trust. AI does not do any of those things. AI pulls from multiple sources simultaneously, combines structured and unstructured inputs, and produces outputs that reflect whatever inconsistencies exist beneath the surface. It does not reconcile differences – it operationalizes them.
What we see in practice is consistent. An organization invests significantly in improving customer data. Confidence rises. Then AI is introduced to drive recommendations. Trust erodes almost immediately. Customer records don’t fully align across systems. Context is incomplete at the moment it’s needed. Outputs can’t be traced to a clear source. The model takes the blame. The system is the cause.
The failure points are almost never in the data itself or in the model. They are at the seams between systems – where data moves across integrations, where definitions drift, where access creates inconsistent views, where lineage breaks down. On paper, everything works. In production, this is where trust erodes.
“Poor data quality costs the US economy an estimated $3.1 trillion annually. In an AI context, that drag does not stay constant – it compounds. An agent operating on unmastered data produces inaccurate outputs at machine speed, across every transaction it touches.”
AI Requires a Data System – Not just Datasets
The shift is from ‘do we have good data?’ to ‘do we have a system that makes data usable across the business?’ That distinction changes the investment thesis entirely.
A data system, not just clean datasets, requires four things to function under AI:
- Aligned definitions. Core entities need to mean the same thing everywhere they are accessed. Customer, product, location, supplier – if definitions drift across systems, AI will surface the inconsistency in every output.
- Consistent integration. Connection is not consistency. API infrastructure built for human-paced workflows was not designed for agents running continuously at machine speed. That mismatch is where underperformance compounds.
- Governance at the point of use. Policies that exist only in documents will not hold when agents are executing decisions in real time. Governance must operate where data is used, not where it is stored.
- Ownership with teeth. Who owns data definitions across systems? Who resolves conflicts when systems disagree? And who maintains quality over time? These questions need people assigned to them with actual authority to act.
The Unlock Most Organizations Are Missing
When organizations ask why their AI is underdelivering, the answer is almost always one of two things: the data that the agent needs is not trustworthy, or the system that delivers it to the agent is not reliable. Often both.
Master data management is the answer to the first problem. A properly implemented MDM platform – Informatica C360, for example – creates a single governed source of truth for core entities, synchronized in real time across systems. When an AI agent needs to know everything relevant about a customer, a supplier, or a product, it draws from one mastered record instead of reconciling conflicting information across dozens of systems. The intelligence of the agent is only as good as the data it operates on.
Model Context Protocol (MCP) addresses the second problem. Launched by Anthropic in November 2024 and universally adopted within months by OpenAI, Google, and Microsoft, MCP is the integration standard that allows agents to connect to enterprise systems reliably without rebuilding integrations for every new use case. Instead of custom connectors between every agent and every system, integrations are written once – to the MCP server. Every agent draws through that layer. BCG characterizes MCP’s impact precisely: without it, integration complexity rises quadratically as agents multiply; with it, complexity scales linearly.
MDM ensures the data is trustworthy. MCP ensures it is delivered reliably. Together, they are what converts a model with potential into an agent that performs in production.
“The organizations pulling ahead did not invest in better models. They invested in the system that makes any model work – mastered data, governed integration, and the infrastructure to deliver both reliably at scale.”
The Question Has Changed
The organizations that are extracting real value from AI made a different investment decision. They did not spend more time selecting models. They spent that time on the infrastructure that determines whether any model can perform in their environment. Data aligned across systems. Integration designed for machine-driven consumption. Governance that operates at runtime. Ownership that resolves conflicts quickly.
The question for most organizations is no longer ‘is our data good enough?’ It is ‘is our data system good enough?’ Those are not the same question. And until the difference is clear, the same pattern will repeat: better inputs, unreliable outcomes.
The Bottom Line
Good data is necessary. It is no longer the differentiator. AI doesn’t underperform because the data is bad – it underperforms because data isn’t coordinated, controlled, and usable across the enterprise. Until that changes, the results will keep falling short of the investment.
By Michael Martin, Managing Director, Data & AI