AI Can Reach 45% of Your Data. The Firm Measuring That Also Sells the Remedy

A note on the file before the argument. The angle assigned for this piece concerned state attorneys general, a media merger and Meta's engagement design, and none of the five sources supplied mentions an attorney general, a merger, a bond or Meta ^[0]. What the file did contain was a vendor-sponsored survey about AI agents and enterprise data, the product page of the vendor that sponsored it, and a history of the people who predicted government by machine before the punch cards had cooled. That is a real subject, and it is the one below.

The development: MIT Technology Review Insights, in partnership with Google Cloud, has published a study reporting that AI today can access an average of 45% of enterprise data, while more than two out of every three organisations plan to deploy AI agents widely within two years ^[1][2][3]. It matters because the binding constraint on agentic AI is now being named, quantified and ranked by a party that sells the fix, and the fix has a published price list ^[4][13][15].

The arithmetic of the gap

Two numbers do most of the work. The report says the average organisation's AI reaches 45% of its data, and it sorts respondents into "data leaders" who give agents more than 70% access and "data laggards" who give less than 30%, with leaders said to achieve stronger results and to scale more easily ^[3][5]. The average organisation therefore sits 25 percentage points below the threshold the report uses to define competence ^[10]. That is the shape of an effective sponsored study: the diagnosis is universal and the cure is a named category.

The trust figure deserves slower reading. For data leaders, according to the report, 100% report "consistently" or "mostly accurate" agent decisions, and the report says the trust gap effectively vanishes ^[6]. But that 100% is the sum of two adjacent response bands, and one of them explicitly permits inaccuracy ^[11]. "Mostly accurate" is the interesting half of the total, and it is not broken out in the material supplied.

Then the headline grievance: 55% of executives confirm that their current data systems actively prevent them from scaling agentic AI across the enterprise ^[4]. That is a self-assessment by executives of infrastructure they own, gathered in a study whose full findings sit behind a report download ^[9]. The supplied summary does not state a sample size, a sampling frame, or how "data access" was measured ^[12]. The gap between the share of data stores an agent is connected to and the share of questions an agent can actually answer is the entire substance of the claim, and nothing supplied resolves which of the two the 45% describes ^[3][12].

The remedy is the catalogue

The report names the initiatives leaders prioritise: improving all data access, and enhancing governance with business context ^[7]. Set that beside the sponsor's own front page, which advertises one platform for agent development, orchestration and governance, and an agentic Data Cloud described as powering the system of action ^[13]. The two recommended remedies map onto two advertised product lines ^[20]. Adjacent copy offers infrastructure "scaling for the agentic era" and a threat-defence product, and the build path is named as the Gemini Enterprise Agent Platform ^[14][18].

Then the ladder. New users get a $300 welcome credit plus free usage of more than 20 products ^[15]. A programme called Gemini Enterprise Agent Ready hands out 35 monthly credits to learn to build enterprise-grade agents ^[16]. Early-stage funded startups can receive up to $350,000 in cloud credits ^[17]. Top to bottom that is a spread of roughly 1,167 to one ^[19]. Credits are denominated in the seller's own product, so the discount is paid out of margin rather than cash, and what it buys is position: default proximity between the agent and the place the data already sits ^[15][17].

None of this makes the 45% wrong. It makes it a number with an owner.

What survives is not what happened

Two other files in today's material are about how evidence gets preserved, and they are more useful here than they look. At Altamira in northern Spain, painting began about 34,000 years ago and continued for millennia until a rockfall sealed the cave mouth, and the reds and blacks are still solid and vivid because the landslide imposed near-quarantine conditions ^[30]. The current thinking, according to Diego Garate Maidagan of the University of Cantabria, is that our ancestors painted their way across western Europe and that "cave art" is only what survived on the deepest, darkest surfaces; elsewhere the pigment was eaten by bacteria, effaced by calcite or scoured away, leaving vestigial chisel marks ^[32][33].

That is the structure of a "data leaders" cohort. A group defined by its own surviving outcomes reports on preservation conditions at least as much as on method, and presenting that cohort as "a clear blueprint for others to follow" asserts a direction of causation that a survey of self-reported access percentages cannot establish ^[9][5].

The access detail is sharper still. Altamira opened to the public in 1917, was partly closed in the 1970s and shut for good in 2002, once a century of admiration revealed the paint-stripping effect of moisture and carbon dioxide from visitors' breath; a replica cave with replica artwork was built on an adjacent site, and only Garate and a few other scholars now enter the original ^[31]. Enterprises are buying a version of that arrangement: the agent reasons over the accessible copy, and what sits in the unreachable remainder is a governance question rather than a model question ^[3][4].

One more from the same file, on scaling. Garate's department had three people and, by his reckoning, would each need three lifetimes to explore all the caves, so they trained and deputised cavers from the Union of Basque Speleologists to angle their head torches and adjust their gaze ^[33]. The constraint was trained attention, and the answer was more trained humans rather than a brighter light.

The pitch has a 1958 draft

The oldest file in the folder explains the newest. In 1958 a team of American researchers and advertising men built a machine they claimed could both predict and influence voting behaviour ^[21]. Edward L Greenfield, who ran a Madison Avenue firm, proposed compiling hundreds of thousands of punch cards - election returns, opinion surveys, census data - into an "information bank" that sorted voters by type and opinions by issue, a "macroscope" able to simulate the whole US electorate ^[23][25]. Greenfield and the MIT political scientist Ithiel de Sola Pool wrote that year that there was "nothing mysterious about this on the surface, the input to the machine being the information about real individuals obtained in surveys", but that "once this information is inside the high speed storage facilities of the machine, it is a different world" ^[24]. The following year they incorporated with IBM's Alex Bernstein as the Simulmatics Corporation, meaning the automated simulation of human intelligence ^[26].

That is the same two-step on sale now: the inputs are ordinary records, and the claimed transformation happens after ingestion. Greenfield's version had to go out and buy the punch cards; the current version says the records are already inside the building and only 45% of them are reachable ^[3][23]. Between the 1950s and the 1980s, electronic data went from a molehill to a mountain as sizeable chunks of public life became automated ^[29]. In the same year as Greenfield's proposal, speakers at the National Physical Laboratory in Teddington described a future in which computers would compose music, resolve legal disputes, control air traffic, conduct surgery and, in the administration of government, replace not only clerks but executives ^[22]. In 1981 the Japanese sociologist Yoneji Masuda set out two destinations: a "computopia" that erased distinctions of class and nation, or a "compudystopia" he called an automated state, government by machine ^[27]. The Guardian's long read names the through-line the artificial state: democratic deliberation replaced by prediction by calculation, the public sphere supplanted by data-driven commerce ^[28].

Read the enterprise version against that history and the stake is legible. Once the bottleneck is defined as access, every decision downstream is a decision about who grants access, on what terms, and into whose orchestration layer ^[7][13].

A closing note on substrate. The dinosaur database in today's file is built on PaleoDB, a scientific database assembled by hundreds of paleontologists over two decades, and the compensation the site names for the illustrations it curates is credit to the original authors ^[35]. The interface is cheap; the substrate took twenty years; attribution is the only currency named. Enterprise data carries the same asymmetry, and the party that connects to it captures the option value.

What to watch

Whether the 45% figure is restated next cycle, by whom, and with a published definition behind it ^[3][12]. Whether the named practitioners - engineering and technology leaders at Shopify, Deutsche Telekom and HCA Healthcare - report access percentages rather than modernisation narratives ^[8]. Whether the "mostly accurate" band is ever separated out of the 100% ^[6][11]. And whether governance ends up bought as a product from the same vendor that supplies the agents, or retained as a control by the buyer ^[13][7].