As enterprises move from AI assistants to fully autonomous agents — systems that don’t just recommend but also act — the stakes of poor data quality have never been higher.
“When people talk about AI failing in enterprises, they usually assume the models are the problem,” says William Benjamin, principal generative AI/ML expert at Allstate. “In practice, the models are rarely the limiting factor. The real constraint is data. Most AI initiatives are actually data engineering problems in disguise.”
To better understand the barriers, Foundry asked the Expert Contributors Network: What data challenges hold enterprises back from realizing AI’s promise?
Based on their responses, here are the key themes and strategies that emerged.
A data and AI puzzle with pieces missing
For most large organizations, data is gathered organically across ERP and CRM systems, operational tools, and cloud services. Every system holds a piece of the story, but the full picture rarely comes together.
“The single greatest impediment to enterprise AI value is not model capability but data fragmentation,” says Javier Campos,Group CTO and chief AI officer at Cape.io. Decades of siloed environments have produced “a patchwork of inconsistent schemas, duplicated records, and undocumented lineage that starves AI models of the coherent, high-quality inputs they require.”
Tom Allen, founder of The AI Journal, sees the same pattern. “Many enterprises still have customer and operational data scattered across different clouds, CRM systems, and productivity tools.”
For AI agents to take autonomous action — scheduling, ordering, resolving — they must reason over live, unified data. A siloed architecture doesn’t just slow humans down; it makes autonomous action unreliable or unsafe.
These disparate systems should instead be treated as a connected data fabric, Allen adds, so AI can reason over a consistent, high-quality view of customers, content, and processes in real time.
However, long-standing silos can make it difficult to know what data even exists.
“Most enterprises don’t truly know where all their data lives or which systems are touching it,” says Scott Schober, president and CEO of Berkeley Varitronics Systems. “Data is often isolated, poorly labeled, or conflicting, which makes it unreliable for producing accurate AI-driven insights.”
“AI is essentially trying to solve a puzzle with many pieces missing,” saysChris Selland, go-to-market lead at Akumina. “The primary challenge isn’t that enterprises lack data, it’s the usability of that data that generally needs to be improved.”
Bringing datasets together is only part of the challenge. AI systems still depend on how data is defined, labeled, and interpreted.
“A major reason AI and ML fail to produce actionable insights is the lack of semantic understanding,” says Peter Nichol, product leader for data and analytics at a large CPG and health company.
For example, a field in a supply-chain database labeled SALES_QTY may appear to represent customer demand, yet it often reflects shipments to distributors rather than purchases by consumers. “An AI model may interpret a spike to 50,000 units in a week as surging demand,” Nichol says, “while a planner knows it simply reflects inventory being staged ahead of a promotion.”
Tip: Simplify the data and AI puzzle with a zero-copy approach, which allows teams to access and use data where it already lives, reducing the risk that comes with moving or duplicating it across systems.
Data quality can hinder the ability to scale
Benjamin at Allstate says challenges quickly surface when organizations begin training models.
“Data quality failures show up in predictable ways: incorrect values, missing fields, stale records, inconsistent definitions across systems, and biased labeling,” he says. “Instead of producing insight, they generate confident but misleading outputs.”
Even when organizations capture large amounts of data, the knowledge needed to interpret it often sits somewhere else.
“The information exists, but it is not structured, contextualized, or continuously refined so that AI agents can truly leverage it,” says Michael Bertha, partner and central office lead at Metis Strategy.
Without a structured knowledge layer linking documentation to operational data, automated systems struggle to understand how work actually happens inside the organization.
“I’ve seen organizations invest heavily in AI pilots while still struggling to answer basic questions, such as which dataset is authoritative or who owns it,” says Will Kelly, a writer focused on AI and the cloud.
This lack of ownership and sourcing creates a crisis of confidence and trust in AI outputs.
There’s another gap that appears as enterprises capture interaction history: “Digital amnesia in the context of users’ interactions with AI — either internal employees or external customers — is a stumbling block for many enterprises,” says Dr. Martin De Saulles, principal analyst at Information Matters.
“Enterprises need to incorporate persistent memory layers through vector databases that capture the semantic meaning of user interactions rather than just keywords,” De Saulles says.
Understanding why customers make decisions or how they engage with content adds depth that transactional data alone cannot provide.
Tip: Work with partners that have developed deep integrations and deliver high-quality, contextualized data. This allows AI agents to transform from simple chatbots to reliable digital labor. Agentic AI systems also benefit from a zero-copy approach, enabling agents to reason over live data where it lives — without the latency or risk of duplication.
The impact of governance and culture on data flows
Data challenges aren’t just technical in nature, according to some of the experts.
“To unleash AI’s potential, companies must treat their enterprise data as a highly managed product, modernizing it before AI model building,” says Robert Siciliano, CEO of Protect Now LLC.
Privacy rules and security concerns can add further complexity, especially when organizations try to train models on sensitive information.
Organizations must also address cultural habits that keep data siloed, says Daniel Jacobs, founder and CEO of Starkhorn. “Before mid-market businesses chase AI’s promise, they must first look honestly inward, auditing not just their data, but the human habits and organizational structures that keep it fragmented and distrusted. Governance means agreeing across departments on what data means, who owns it, and why sharing it serves everyone.”
Bridge the data gaps
The experts agree that organizations should seek a unified data strategy to address these pain points.
For example, “enterprises need strong data governance frameworks, unified modern data platforms, seamless connectivity, and robust practices for strategic governance,” says Elitsa Krumova, a global thought leader in emerging technologies. This strategy helps “bridge the gaps in data management and unlock AI’s promised business value and outcomes.”
Experience how Google Cloud and Salesforce provide a real, ready foundation for the agentic era, including unified data through zero-copy functionality that reduces the need to move or duplicate data across systems. These co-innovations are generally available and represent a system that works together seamlessly, rather than a toolkit requiring integration.