UpShaqo
Intelligence desk
Enterprise Agents Source-backed analysis

Why Only 16% of Companies See Real Value From Enterprise AI

A new poll finds nearly every company uses AI, but almost none escape pilot purgatory. The sponsor's diagnosis—missing enterprise context—deserves scrutiny as much as attention.

UpShaqo Editorial IntelligenceOctober 28, 20266 min read
Intelligence standard

Independent UpShaqo analysis built from fresh, attributed sources. We explain the impact instead of repeating the announcement.

Read for leverage: focus on the workflow change, the customer problem, and the next action—not only the product announcement.

Every enterprise leader approving an AI budget this quarter faces the same uncomfortable question: do you fund more AI tools, or do you fund the infrastructure that makes the tools you already have actually work? New survey data suggests most companies have been choosing the former and getting disappointing results.

An MIT Technology Review Insights poll found that every organization surveyed is using AI in some form, yet fewer than one in five—16%—describe their deployments as delivering transformative value. The largest group, 40%, reports only moderate gains confined to specific tasks or teams. Another 21% see meaningful gains. The remaining 23% are getting minimal returns or none. Most companies, in other words, are still waiting for AI to pay off at scale.

What the Numbers Actually Say

The adoption-maturity breakdown is the most useful part of the data. According to the poll, 38% of organizations are still piloting AI use cases, and 35% are using it mainly to assist employees through copilots or chat assistants. Only 12% have AI agents running in production, and just 3% say AI is doing most of the work in key processes. That's a steep drop-off curve: nearly three-quarters of companies are stuck at the experimentation or assistance stage, and the leap to autonomous, production-grade agents is happening almost nowhere.

It's worth being precise about what this evidence does and doesn't establish. The report doesn't disclose sample size, respondent industries, or survey methodology in the material reviewed here, and it was commissioned and published as sponsored content—produced by MIT Technology Review's Insights custom-content arm "in association with Atlan," not by its editorial staff. That doesn't make the numbers false, but it does mean they should be read as directional signal from a vendor-aligned study rather than independently audited research.

The Vendor's Diagnosis—and Its Blind Spot

Atlan co-founder Prukalpa Sankar frames the gap in stark terms: "There's a big gap between what the models are capable of and how the models are actually being useful inside organizations," she says, arguing that "for every real-life intent and purpose, artificial general intelligence is already here"—it's simply not useful inside most companies yet, according to the report. Her proposed fix is what Atlan calls a context layer: infrastructure sitting between a company's systems of record and its AI platforms, translating business knowledge into a governed, machine-readable form that agents and models can actually use.

Her analogy is memorable: deploying AI without this layer is like hiring the smartest intern in the world, blindfolding them, and throwing them into the organization. It's a vivid way to describe a real problem—language models have broad reasoning ability but no inherent knowledge of a specific company's product exceptions, approval thresholds, or customer history.

The analysis worth adding here: Atlan sells data cataloging and governance software, which is functionally adjacent to—if not identical to—the "context layer" concept it's describing. That doesn't invalidate the diagnosis, but it means the prescription should be treated as a vendor's pitch for its own category, not a neutral research conclusion. Readers should separate the empirical finding (adoption is stalled, most companies aren't seeing transformative returns) from the interpretive claim (the solution is a specific architectural layer sold by companies like Atlan).

A Concrete Illustration of the Problem

Consider two deployments of the same underlying model inside a retail company. In the first, an employee uses a chat assistant to draft a customer-service reply; a human reviews it before sending, so the model's lack of business-specific knowledge is a minor inconvenience corrected by oversight. In the second, an autonomous agent is empowered to approve refunds directly. Without access to governed context—return-policy exceptions, fraud flags, customer tier, prior disputes—that agent either refuses too conservatively or approves incorrectly, and there's no human in the loop to catch it.

This is a reasonable inference from the pattern in the data, not a specific case study in the research: the 12%-in-production/3%-doing-most-of-the-work split likely reflects exactly this risk calculus. Companies will let AI draft or suggest far more readily than they'll let it act, because acting without context is where errors become operationally and reputationally expensive.

Separating Evidence From Speculation

To keep this brief disciplined, here's what's actually supported versus what's inference:

  • Evidence: Nearly all surveyed organizations use AI; only 16% report transformative value; adoption clusters heavily at the piloting and assistance stages; agent-in-production rates are low (12%) and autonomous-majority deployments are rare (3%).
  • Evidence: The report is sponsored content co-produced with a vendor in the data-context space, a disclosed fact worth weighing when assessing the recommended fix.
  • Inference: That the low-autonomy numbers reflect risk aversion around ungoverned context, rather than purely technical limitations or budget constraints, is a plausible reading but not something the poll data isolates on its own.
  • Speculation to avoid: Any claim that a context layer alone would move organizations into the 16% transformative-value bucket. The report offers no before/after data or case evidence for that causal claim.

Practical Controls for Leaders Evaluating This Pitch

For operators weighing whether to invest in context infrastructure before scaling agent deployments, a few checks are more useful than the marketing framing:

  • Audit where your current AI failures actually originate—model reasoning errors versus missing or stale business data—before buying a governance platform to solve a problem you haven't confirmed exists.
  • Pilot agent autonomy in reversible, low-stakes processes first, and track whether access to structured internal data (not just better prompts) measurably reduces error rates.
  • Treat any vendor's definition of "context layer," "company brain," or similar branded architecture as a sales category, and request specifics on what data sources, governance controls, and access permissions it actually touches.
  • Separate copilot-style assistance, which tolerates imperfect context because humans review output, from autonomous agent deployment, which does not.

What Remains Unresolved

The poll doesn't say how long companies have been piloting, what industries are overrepresented, or whether the 16% achieving transformative value share any common infrastructure choices—context layers or otherwise. It also doesn't address cost: building governed, machine-usable context across legacy systems of record is itself a significant and ongoing investment, and the report offers no estimate of what that costs relative to the returns it claims to unlock. Until independent, vendor-neutral research tracks organizations before and after adopting this kind of infrastructure, the central claim—that context, not capability, is the bottleneck—remains a compelling hypothesis rather than a proven fix.

Sources

#enterprise AI#AI agents#data governance#context layer#AI ROI#Atlan

Two doors. Pick one.

Hire the team.
Or become it.