Fragmented enterprise knowledge is now the stated reason companies will not let their own AI agents act. In a Tata Consultancy Services survey of 1,333 senior leaders in 25 countries, released on 4 November 2026, 84 percent ranked productivity as the leading source of value from AI, and 44 percent said important business knowledge still sits inside employees or in scattered documents. Only 10 percent allow AI agents to execute decisions without human approval across multiple functions. The distance between those numbers is the position of enterprise AI in late 2026. The assistance is welcome. The authority is withheld. The reason leaders give sits in the record an agent would have to act on.

The gains are individual. The authority is not.

McKinsey's State of AI 2026 survey (1,719 respondents in 97 nations, fielded from 4 May to 8 June 2026) found 80 percent of respondents reporting that AI improved their individual productivity and 50 percent reporting that it helps them make better decisions. At company level the line does not move: 37 percent attribute any earnings impact to AI, unchanged from a year ago, and the share McKinsey classifies as high performers held flat at about 6 percent.

Accenture's first CIO Pulse survey, published on 3 November 2026 from 300 CIOs and IT leaders surveyed in August and September 2026, shows the same split inside the technology function. Seventy percent use AI mainly for individual productivity and copilots inside a task, against 37 percent using agents for end to end orchestration. Sixty four percent say they are scaling deployment, and 7 percent say they are optimizing at scale.

TCS puts the authority figure directly. Just 9 percent of its respondents report AI agents capable of securely orchestrating workflows across enterprise applications. Deployment is rising on every survey. Permission to act is not. Individual gains that stay on one desk are the pattern this site tracked in shared AI context: an insight one person holds does not become a company capability until it enters a record other people and agents can use.

Fragmented enterprise knowledge is the blocker leaders name

The TCS figure of 44 percent is not an outlier. In a survey of 1,000 senior technology and data leaders conducted by Wakefield Research for Teradata (fieldwork from 23 March to 5 April 2026), 77 percent said only 20 percent or less of their enterprise data and knowledge is ready for AI agents to act on reliably, and 78 percent reported difficulty creating a connected data foundation for agents to act on. On the report's maturity index, 68 percent of organizations sit in the Experimenting and Developing phases and 7 percent have reached Operationalizing.

Accenture's CIOs name the same condition from the data side: 53 percent cite poor data readiness and data quality, and 51 percent cite the absence of a trusted, governed data foundation. Accenture's recommended remedy is a knowledge layer that gives agents a structured understanding of what data exists, what it means and how it connects across the organization. That layer describes the data. It does not by itself settle which version of a fact is current, who decided it, or what a correction should reach. Those are record questions, and they are the questions an acting agent raises, as this site set out in operating knowledge.

Governance is the gate, and leaders say so

In the TCS study, security, risk and governance rank as the leading factor in how and where companies build and deploy AI, ahead of cost, speed and ease of deployment. Amit Kapur, TCS Chief AI and Services Transformation Officer, said what is holding organizations back, "and rightly so", is "concerns around security, governance, and resilience". Accenture found 54 percent of CIOs naming governance and responsible AI concerns as a top barrier to scaling, while only 32 percent have reached Advanced or Leading governance maturity, and 49 percent in IT operations still run agents with manual approvals.

The economics are part of the same caution. TCS found 39 percent of enterprises still assessing whether AI returns enough value against token and compute costs, and a wide gap between concern and preparation on dependence: 90 percent call model independence an important factor in investment decisions, and 7 percent report an organization wide plan for switching providers.

Read together, the refusal of autonomy is not resistance to the technology. On a fragmented record, an unsupervised agent does not fix the fragmentation. It acts on it, at machine speed, and the error inherits the authority of the action it took.

The Act-Ready Record Test for fragmented enterprise knowledge

Autonomy can be widened safely when it is granted against a record, function by function, rather than granted to a model in general. Before an agent acts without approval in a function, the record it acts on should pass three checks.

First, provenance. Every fact the action depends on carries its source, its date and the version it superseded, in a form a person can open and read without asking the agent.

Second, correction. When a fact changes, the correction reaches the derived summaries, the standing instructions and the actions already queued, not only the source file it started in.

Third, receipt. The action leaves a record of what was done, on which facts, under whose delegation, in a form a person can inspect and reverse.

Where the record passes, approval can move from every action to sampled review. Where it fails, the manual approval Accenture's CIOs report is the correct control, and removing it transfers the cost rather than cutting it. The performance data points the same way: McKinsey's high performers, the roughly 6 percent reporting significant earnings impact, report fundamentally redesigning workflows at nearly three quarters, against one quarter of other respondents. The difference is work done on how the organization operates and records, not a different model.

One limit applies to all four surveys. Each measures what executives report about their own organizations, on different samples and field dates, so the percentages are not interchangeable and none of them audits a live company record. The perception gap inside the Teradata data, where 69 percent of C-suite respondents against 57 percent of vice presidents said their organization is already operating with agentic AI, is a warning against reading any single figure as a measurement.

A company earns the right to delegate the way an employee does, by keeping work that can be checked. Fragmented enterprise knowledge is not a search failure and it will not be closed by a better model. It closes when the facts an agent acts on sit in one reviewed record, with a source, a date and a superseded version, which is the machinery described in How Jilu Works and in the context engineering pillar. Until that record exists, the 10 percent figure is not a lag. It is a judgment, and on the evidence of these surveys it is the right one.