TL;DR: Company AI projects are stalling because enterprise data is fragmented, poorly described and out of reach for agents, not because models lack capability.

New surveys published in early October 2026 point to the same bottleneck. A Harris Poll survey conducted for Collibra and reported on Oct. 9, 2026 found 72% of technology decision-makers traced unsuccessful artificial intelligence initiatives to weak data foundations, while an Oct. 11, 2026 report from Teradata based on 1,000 senior technology and data leaders found only 37% report strong positive returns even as 90% expect to increase agentic AI investments over the next 12 months.

72% blame weak foundations for AI shortfalls

The Collibra survey included 300 employees working in data management and privacy as well as AI decision-makers, according to the Oct. 9, 2026 account titled Poor data linked to most unsuccessful AI initiatives. Among those respondents, 72% identified inadequate data foundations as the primary cause of AI initiatives falling short.

That result reframes pilot post-mortems. The problem described in the survey is not a failed demonstration of language capability. It is mismatched keys, stale records, undocumented lineage and a customer table that staff do not trust to support production use.

Companies are responding with organizational changes. About 53% of decision-makers said their organizations are moving oversight of AI functions closer to their data operations. Another 58% of enterprises are working to establish clear internal accountability for AI-generated results.

Regulatory preparation is adding to the shift. Nine in 10 business leaders surveyed by Harris Poll said their organizations were preparing for AI regulations in the United States and other countries. About 51% said their companies were investing in data lineage and documentation, which track the origins, movement and use of data.

Agents see less than half of enterprise data

A separate August report from Google Cloud and MIT Technology Review Insights found AI systems can access an average of 45% of an organization's data. Only half of the organizations surveyed said they trusted their AI agents to produce relevant and accurate results.

The access gap helps explain daily supervision work. About 87% of decision-makers in the Harris Poll survey for Collibra said their teams regularly verify that information available to AI agents remains accurate and current. More than half said employees spend hours manually reviewing and correcting agent outputs.

That manual review burden appears in other October summaries of the same survey. An Oct. 8, 2026 analysis of supervision costs noted that more than half of employees spend hours reviewing outputs, while 87% re-check context for accuracy and currency. The same analysis cited Gartner's 2026 CIO and Technology Executive Survey finding only 17% of organizations have deployed AI agents while more than 60% expect to within two years.

The Google Cloud and MIT Technology Review Insights figures add detail on legacy constraints. In that 300-executive summer survey, 55% said legacy data systems are preventing them from scaling agentic AI, and only 51% trust the accuracy and relevance of agent decisions. Data leaders expose more than 70% of enterprise data to AI systems, while laggards expose 30% or less. Among laggards, trust falls to 22%.

64% traced confident errors to their own data

An Oct. 8, 2026 summary of VentureBeat Intelligence's August 2026 Pulse survey reported 64% of enterprises have traced a confidently wrong AI agent answer to problems in their own company data in the past six months. The Pulse survey covered 184 enterprise respondents with 100 or more employees. Of affected organizations, nearly half experienced such an error more than once.

The types of failures were consistent across recent VentureBeat surveys. The summary listed wrong metric definitions, stale data that no longer reflects current conditions, and missing documents that the agent needed to answer correctly.

Wrong metric definitions occur when finance, sales and product teams define revenue, customer count or other concepts differently. Stale data causes answers that were correct at creation to become incorrect in operation. Missing documents leave the agent to answer without required context.

The trajectory in that series moved upward across quarters. The June 2026 survey found 57% had traced agent errors to bad context. The July 2026 survey found 68%. The August survey found 64% had traced a confidently wrong answer to company data problems, with recurrence at nearly half of affected organizations.

Fragmentation blocks the move from pilots to returns

The Oct. 11, 2026 Teradata account titled Disconnected Data Is Stifling ROI From Agentic AI describes data fragmentation as a primary roadblock for organizations trying to move from agentic AI pilots to enterprise-wide autonomous knowledge that delivers measurable results. The report surveyed 1,000 senior technology and data leaders.

The investment intent is broad. Although 90% of respondents expect to increase agentic AI investments over the next 12 months, only 37% report strong positive returns to date. The report places the bottleneck shift from pilot quality to enterprise-wide data readiness.

Three context figures illustrate the gap. About 77% of executives report that 20% or less of their enterprise data is sufficiently described and contextualized for agents to use. About 78% say it is challenging to unify data and knowledge across business functions so agents can reason across the full enterprise. Among specific barriers, leaders rank data lacking necessary metadata, context and relationships at 43%, nearly equal to data fragmented across systems at 42%.

The underlying design issue is that enterprise data was built for humans. The Teradata report states that even organized data typically assumes a person will interpret context and meaning before acting. Agents now need that context supplied directly.

A related Oct. 8, 2026 report on C-suite perceptions of agentic AI summarized the same Teradata survey and found 40% of respondents say more than 40% of AI pilot projects pause before production because infrastructure is not ready for autonomy. In that summary, 78% of leaders report difficulty creating a connected data foundation for agentic AI to act on.

Only 7% reach operational scale

The Teradata report uses an Agentic AI Maturity Index with four stages. Companies in the Experimenting stage, at 28%, are still exploring localized pilot projects that typically deliver personal productivity gains only.

In the Developing stage, where 40% of enterprises currently sit, organizations have found success with some models and automations but are limited by data connectivity across teams. These organizations are working to shift from personal to organizational gains and encountering interoperability challenges across data sets.

Those in the Building stage, at 25%, have basic structured governance and workflows but have yet to solve the knowledge fragmentation problem. They have more connected data but still lack critical context layers to enable autonomous decision-making at the enterprise level.

Only 7% of enterprises in the Operationalizing stage have dynamic governance and safety rules alongside the data foundations necessary to scale autonomous decision-making across the entire organization. These are the companies starting to see measurable business impact from agentic AI that can execute multistep workflows.

A companion summary found 68% of organizations are stuck in the Experimenting and Developing phases combined. Only 25% have reached Building and 7% have reached Operationalizing. In that summary, 62% of leaders prioritize enterprise-wide ROI over individual productivity gains when evaluating AI, but only 30% actually use margin improvement as a success metric. About 63% admit seeing no more than an emerging, minimal positive return to date.

C-suite sees more progress than vice presidents do

Perception differs by rank. The Oct. 8, 2026 summary of the Teradata-commissioned survey found 69% of C-suite executives say their organization is already operating with agentic AI, while only 57% of vice presidents say the same.

The distinction between personal and enterprise gains helps explain that split. Personal AI gains come from tools like custom chatbots that help individuals or small teams work more efficiently. Enterprise-wide gains require connecting agents to core business systems with shared knowledge, appropriate access levels and governance.

Billing, finance and marketing illustrate the shared-meaning problem described in the summary. Billing counts accounts, finance counts entities and marketing counts emails. Each count can be locally correct while an agent moving across systems has no way to know they disagree without supplied definitions and ordering rules. The report summarizes the requirement as knowing which data, assembled in what order, on behalf of which process, needs to reach an agent at the moment of a decision.

Pilots use clean data, production does not

An Oct. 13, 2026 poll summary from TXP reported an even split on barriers to AI value. Among LinkedIn respondents, 30% cited lack of clear use cases, 30% cited poor data quality, 30% cited skills and adoption, and 10% cited governance concerns.

The TXP account states pilots can often be built around carefully selected information, while production must operate within existing systems, data and workflows. Having data is not the same as having connected, accessible and usable data where AI needs it. TXP's wider research found leaders reported strong confidence in data foundations while technology integration remained a challenge to orchestrating AI across end-to-end workflows.

The same account states a compelling use case can still struggle without the right data, strong data does not guarantee value if people do not adopt the solution, and enthusiastic adoption cannot compensate for a use case that does not solve a meaningful business problem. Critical knowledge around AI can remain with specific individuals rather than being documented and accessible throughout the organization.

Strict ROI definitions leave only 7% fully successful

An Oct. 13, 2026 summary from GBK Collective of the fourth annual Wharton Human-AI Research study, titled The Adaptive Advantage, reported 80% of enterprise leaders report positive returns from AI. The same study found 7% of leaders' organizations clear all three markers of success defined in the report.

The survey covered 1,416 respondents, including 1,007 enterprise leaders and 409 entry- and mid-level employees. Proving value required strongly positive ROI from a department that formally measures it. Workflow transformation required AI fully embedded in core processes, use of agents in some form and significantly changed workflows. While 67% of departments use agents and 73% say workflows have changed, just 20% meet all three transformation conditions.

Organizational alignment required four of five factors: adequate AI training, a clear long-term vision from leadership, meaningful choice over how and when to use AI, not being held back by people and processes rather than technology, and no real resistance among colleagues. Only 7% meet all three markers combined.

The study also recorded a split on training. About 78% of leaders believe employees receive adequate AI training, compared with 56% of employees. About 73% of employees said AI has had a positive effect on their own role, and 84% of those who use it said it saves them time.

Daily use rose 16 points to 62% while the share of managers who describe themselves as excited about AI fell 8 points. In 48% of departments, AI performs work and a person approves it before anything happens. Transformation was weakest where every action is approved at 16% and where a sample is reviewed afterward at 18%, and highest where agents operate under real-time monitoring and guardrails at 31% and where AI only advises at 25%.

What agent-ready data requires

The Teradata report defines progress as a shift to autonomous knowledge, described as data that carries enough context, governance and meaning for AI agents to act on reliably and repeatedly. Teradata CIO Josh Fecteau said in the report that contextualizing an entire data estate is likely the wrong goalpost because data estates keep changing with new types, sources and multimodal inputs.

Fecteau suggested identifying the highest-value portion of data, structured or unstructured, and focusing on getting that portion fully described, governed and agent-ready. He said if most data is unusable, the answer is not to fix all of it at once but to be selective about which 20% to 50% to start with.

The report summarizes steps to reach the Operationalizing stage. First, audit and contextualize the data estate starting with highest-value data sets, applying consistent definitions, lineage and meaning across the organization regardless of team or function. Second, embed dynamic governance in the foundational data architecture itself, not the software or application layer, so guardrails transfer and decisions remain traceable through multistep processes. Third, build for portability by creating reusable, governed assets that any model can interpret, which preserves flexibility in the software layer.

Decision paralysis complicates those choices. The report finds 60% of leaders experience some decision paralysis on technology decisions due to desire for long-term flexibility. Specifically, 30% cite vendor lock-in as an active worry. The report states embedding context and connecting disparate data layers within the foundational architecture removes specific software dependencies and enables reliable autonomous decision-making.

Why data work now decides AI budgets

The pattern across the October surveys is that AI budgets function as data cleanup bills. An Oct. 7, 2026 Forbes Councils account described boards approving AI programs that stall in production and produce post-mortems citing mismatched keys, stale records, undocumented lineage and an untrusted customer table.

Other failure-rate comparisons show where projects stop. An October 2026 comparison of eight studies found MIT NANDA counted 95% of companies getting little or no measurable profit-and-loss impact from generative AI pilots, while IDC with Lenovo found only four of every 33 proofs of concept reached production. S&P Global found 42% of companies abandoning most AI initiatives in 2025, up from 17% a year earlier.

RAND interviews with 50 industry data scientists and engineers found persistent data quality issues raised by 30 interviewees, the most cited cause. BCG's survey of 1,000 executives found about 70% of AI program challenges came from people and process issues, 20% from technology and 10% from algorithms.

For United States technology teams, the near-term implication in these reports is selective, governed data work before broader agent deployment. The shared prescription is to choose a high-value workflow, supply definitions and lineage for that workflow, place human approval and monitoring where errors carry cost, and measure returns against a pre-set baseline.

Sources referenced