The Trust Gap: Why Enterprises Still Hesitate on Agentic AI Even as Adoption Soars

Key takeaways:

“Agentic reasoning” — AI systems that plan and execute multi-step tasks with limited human intervention — is widely described as 2026’s defining industry trend, succeeding the multimodal wave of a couple of years earlier.
Palantir reported a blowout second quarter, with revenue up roughly 93% year-over-year and net income around $1 billion, while its CEO simultaneously renewed sharp public criticism of frontier AI labs as untrustworthy partners for enterprise customers.
The core argument centers on data and intellectual property: enterprises handing sensitive operational data to third-party model providers, allegedly with limited assurance about how that data and expertise gets used.
Executives across the industry, including at Microsoft and JPMorgan, have weighed in publicly, reflecting a genuine and unresolved debate rather than a settled consensus.
The dispute highlights a structural tension in enterprise AI adoption: raw model capability is advancing quickly, but governance, control, and trust infrastructure are lagging behind it.

A trend and a controversy, arriving together

If 2024 was widely characterized as the year of multimodal AI — models that could see and hear as well as read and write — 2026 is increasingly described as the year of agentic reasoning: systems designed not just to answer a question, but to plan out and execute a sequence of actions toward a goal, often with minimal step-by-step human supervision. That shift shows up everywhere from coding assistants that can independently work through a multi-file software task to back-office systems that reconcile invoices or route support tickets with limited human intervention.

What makes this particular week notable is that the same period saw one of the sharpest public disputes yet over whether enterprises should actually trust the companies building these agentic systems in the first place. That dispute has a name attached to it: Palantir co-founder and CEO Alex Karp, whose public comments this quarter went well beyond typical corporate messaging.

What Karp is actually arguing

Speaking after Palantir posted a standout quarter — roughly 93% year-over-year revenue growth and about $1 billion in net income — Karp used both a shareholder letter and a series of media appearances to renew and sharpen a critique he’s been building for months. His central claim is that enterprises adopting frontier AI models are, in effect, handing over their most sensitive proprietary data, workflows, and institutional expertise to third-party labs, often without a clear return on that investment and without adequate assurance about how that information ultimately gets used or whether it could eventually be repurposed to compete with the very companies supplying it.

Karp’s language has been notably blunt even by the standards of corporate leadership commentary, at one point suggesting that some frontier labs behave as though they are entitled to what he described as a dominant position over their enterprise customers’ operations, and separately invoking politically loaded terminology to characterize the industry’s dynamics. Whatever one makes of the rhetorical style, the underlying business argument is more measured: real enterprise AI value, in Karp’s framing, requires three distinct layers working together — the underlying model, an application and governance layer that sits on top of it, and the compute infrastructure beneath it — and he argues Palantir’s business is built specifically around supplying that middle, sovereignty-preserving layer rather than competing head-on with the model providers themselves.

The financial case behind the rhetoric

It would be easy to dismiss this as a competitor talking its own book, and there’s certainly an element of that — Palantir’s entire commercial pitch depends on enterprises believing they need a governance and control layer distinct from the model itself. But the financial results attached to the argument are hard to wave away. Palantir’s U.S. commercial revenue has reportedly been growing at well over 100% year-over-year in recent quarters, the company has raised its full-year revenue guidance multiple times this year, and outside analysts have pointed to unusually high net revenue retention as evidence that customers aren’t just signing up, they’re expanding their usage over time. Whatever the merits of Karp’s critique of frontier labs, Palantir’s own numbers suggest that a meaningful number of enterprise customers are, at minimum, receptive to the idea that they need a dedicated sovereignty and application layer distinct from a raw model API.

How the industry has responded

The response from other executives has been genuinely mixed, which is itself informative — this is clearly an unresolved argument rather than a settled industry consensus. Some prominent technology and finance leaders have publicly sided with elements of Karp’s framing around data control and return on investment. Others have pushed back or offered more measured takes, with at least one major bank executive noting that companies will simply have to evaluate their own return on AI spending directly rather than treating the debate as settled in either direction. Frontier labs themselves, including those Karp has named directly, have generally continued to argue that broad model access, including through open-weight releases, ultimately serves customers by fostering competition rather than concentrating control.

What “trust” actually means in this context

Stripped of the more provocative language, the dispute points to a real and fairly technical set of enterprise concerns that are worth naming plainly:

Data handling and IP. What happens to proprietary data, prompts, and workflows sent to a third-party model provider, and under what contractual terms can that data influence future model training or product decisions?
Reliability in long agentic chains. An agentic system executing a dozen sequential steps compounds the risk of a single error propagating through the entire chain, which raises the stakes on model reliability well beyond what a single question-and-answer interaction requires.
Auditability and control. Enterprises in regulated industries increasingly want a clear audit trail of what an AI agent did and why, along with the ability to intervene, pause, or override its actions — capabilities that a governance layer, rather than a raw model API, is generally better positioned to provide.
Vendor concentration risk. Relying heavily on a small number of frontier model providers creates a structural dependency that some enterprise leaders are increasingly uncomfortable with, independent of any specific complaint about a given provider’s behavior.

None of these concerns are new, exactly, but the scale of agentic AI adoption this year has made them considerably more urgent than they were when AI use was mostly confined to single-shot question answering.

A debate without an obvious resolution

It’s worth resisting the temptation to declare a clear winner here, because the honest answer is that both sides are describing something real. Frontier labs genuinely are producing systems capable of increasingly sophisticated agentic behavior, and that capability is driving genuine enterprise value in many deployments. At the same time, the governance, contractual, and trust infrastructure surrounding that capability is demonstrably still catching up, and enterprises with the most sensitive data and the highest regulatory exposure have legitimate reasons to want stronger guarantees than “trust us” before handing over their most valuable operational information.

The practical result, at least for now, is a market where both raw frontier model access and dedicated governance-and-application layers are growing simultaneously, rather than one displacing the other. Whether that dual-track structure persists, or whether frontier labs eventually build the trust and governance features enterprises are asking for directly into their own offerings, is likely to be one of the more consequential open questions in enterprise AI over the next year.

What “the application layer” actually looks like in practice

It’s worth unpacking what companies like Palantir mean, concretely, when they talk about a governance or application layer sitting between a raw frontier model and an enterprise’s actual operations, since the phrase can sound abstract. In practice, this typically involves an intermediate system that mediates every request an AI agent makes: logging what data the agent accessed and what actions it took, enforcing permission boundaries so an agent can’t reach systems or records outside its assigned scope, and providing a structured way for a human reviewer to audit or roll back an agent’s actions after the fact. Some of these platforms also maintain what’s sometimes called an ontology — a structured map of an organization’s data, systems, and business logic — that an AI agent operates against, rather than working directly against raw, unstructured enterprise data with only a prompt as its instruction set.

The pitch behind this architecture is that it lets an enterprise swap out or upgrade the underlying model relatively freely, since the governance layer is designed to be model-agnostic, while retaining consistent control, logging, and compliance guarantees regardless of which frontier lab’s model happens to be doing the underlying reasoning at any given time. That model-agnostic positioning is also, not coincidentally, why companies selling this kind of layer have an incentive to frame direct enterprise reliance on any single frontier lab as risky — it strengthens the case for needing an intermediary in the first place.

How regulators and boards are starting to respond

This debate isn’t confined to earnings calls and cable news appearances. Corporate boards and risk committees at large enterprises have increasingly started asking pointed questions about AI vendor arrangements over the past several quarters, particularly around data usage rights, model training opt-outs, and incident response procedures if an AI agent takes an unauthorized or erroneous action with real business consequences. Some regulated industries, particularly financial services and healthcare, are further along in formalizing these questions into actual procurement requirements rather than informal due diligence, reflecting the reality that regulators overseeing those sectors have their own reasons to care about how AI vendor relationships are structured, independent of the public dispute between any two companies.

The bottom line

Agentic AI adoption is accelerating faster than the governance frameworks meant to make enterprises comfortable with it, and this quarter’s very public dispute over trust is really just that gap surfacing loudly rather than something new appearing out of nowhere. For any organization evaluating agentic AI right now, the practical lesson isn’t to take either side’s framing at face value, but to ask concretely: who controls this data, what happens when an agent makes a consequential mistake, and who is accountable for the answer.