For the past several years, the enterprise AI race has been framed as a contest for model superiority. Companies compared benchmarks, invested in infrastructure, and rushed to put the newest foundation models into production. That made sense when access to capable AI was scarce, but it makes less sense now that model performance is rapidly improving, agentic tools are multiplying, and AI is moving into the workflows where businesses make decisions.

The next differentiator will be trust, demonstrated through the ability to show where information came from, how it was governed, why an AI system produced a particular answer and whether that answer can be defended when the stakes are highest.

Recent events are making this shift harder to ignore. The industry’s loudest debate right now is whether frontier labs should coordinate a slowdown in response to a wave of AI safety warnings. But a more subtle and immediate shift is already underway inside enterprises. Companies are pulling back on frontier models over questions of trust and sovereignty, not over questions of capability. The Information reported that corporate customers of Anthropic and OpenAI have started restricting or refusing to use the labs’ most advanced models over unresolved questions about how their data is retained and used.

The market’s response has been to demand AI sovereignty: the ability to maintain ownership and control over their data, systems, and the outputs they produce, including where that data resides and whose laws govern access to it. In many cases, that calls for a smaller, fine-tuned, self-hosted model that keeps data where it lives, whether on-prem, in the cloud, air-gapped or at the edge. Wharton’s Ethan Mollick, who studies enterprise AI adoption, recently flagged that maintaining a model that keeps pace with general-purpose frontier models is expensive and gets harder every time AI capability jumps again. And on its own, owning the stack solves a jurisdiction problem, not a trust problem: whether an organization can explain and defend what its AI system produced, regardless of whose infrastructure it runs on.

Even a highly capable model can produce poor outcomes when it operates on fragmented, stale or poorly governed information. In customer service, that can mean a confidently wrong answer. In financial services, healthcare, defense, or other regulated environments, the same failure can become a compliance, operational, or reputational problem. The enterprise question is changing from “Which model is best?” to “What information can we safely let this system rely on, and can we prove why?” Adding a policy layer around a model is insufficient. Organizations need controls that travel with the data itself: who can access it, when it was updated, what changes were applied, and whether the output matches the evidence for the decision at hand. Without that context, an answer can look precise while remaining impossible to verify.

According to a recent survey, from my company, Prosper Insights & Analytics , 53.2% of executives and business owners say they already use generative AI, compared with 37.9% of employees. That gap suggests leadership is moving quickly to capture AI’s upside.

But the same executives are not blind to the risks: 40.2% are concerned AI can provide wrong information, 39.3% say it needs human oversight, and 32.0% want more disclosure and transparency around the data it uses. Adoption and skepticism are rising at the same time.

That combination should change how companies approach AI. If leaders want employees, customers and regulators to trust AI-assisted decisions, governance cannot begin at the moment a model generates an answer. It has to begin earlier, before information ever enters the system.

Pat Condo, CEO of Seekr, puts it this way: “For AI to be trusted, you have to be able to account for four things behind every answer: the data it learned from, the context it was given, the model that produced it, and the actions an agent took on it. Most AI systems, frontier models included, cover one or two of these at best. What they offer is data traceability, which amounts to ‘I think the answer came from this source,’ and models often hallucinate about that too. That doesn’t go far enough to explain an answer. Provenance and lineage have to run through all four: where the information came from, and how it was transformed at every step. When they do, you get explainability. And when that explainability stays with your data, under the laws of the nation that governs it, you get sovereignty. Residency alone without explainability isn’t sovereignty.”

Independent research reinforces this. McKinsey’s 2025 State of AI survey found that 51% of respondents at organizations using AI had experienced at least one negative consequence from its use, with inaccuracy the most frequently reported problem. Explainability was among the risks organizations reported experiencing yet was not among the most commonly mitigated. Deployment is outpacing the mechanisms needed to understand and defend outcomes.

The technical evidence is just as sobering. The Stanford HAI 2026 AI Index reported hallucination rates ranging from 22% to 94% across 26 leading models on one new benchmark. Model capability alone does not eliminate the need for validation, and the need grows as systems become more autonomous.

Prosper’s findings on agentic AI make the same point. Among executives and business owners, 17.5% say they already use agentic AI, more than double the 6.6% of employees who say the same.

Yet only 33.8% of executives call agentic AI a good idea, while 35.7% are unsure and 30.5% say no. Even among early adopters, confidence has not caught up with capability.

Condo argues that the industry has the trust question backwards: “Don’t ask users to trust AI. Give them the evidence to verify it. Too much of AI today runs on assumed trust: the vendor says it works, the answer sounds confident, and everyone moves on. A company doesn’t go public without an audit, and AI shouldn’t be making high-stakes decisions without that same level of scrutiny and accountability. The organizations that can show their work, challenge and correct results, will put AI into the workflows that matter. Everyone else stays stuck in pilots, because ‘the model said so’ doesn’t hold up with a regulator, a customer or a board.”

That matters because the model layer is unlikely to stand still. Enterprises will use multiple models, swap providers, deploy smaller specialized systems and introduce more autonomous agents. A strategy tied to one model will age quickly. A strategy centered on trusted information will travel with the business.

For leaders, that means treating governed data as a strategic asset. Identify which sources are authoritative for high-impact workflows. Make provenance and lineage visible. Establish evaluation criteria before deployment, not after a failure. Define where human review is required. And keep an evidence trail that shows not only what an AI system said, but what information shaped the answer.

Enterprise AI’s first phase rewarded whoever had access to the best models. The next will reward explainability: the ability to show, on demand, the evidence and reasoning behind any model’s output. Doing that foundational work now, before a high-stakes failure forces the question, should be the priority for any organization serious about scaling AI responsibly.

Disclosure: The consumer sentiment study referenced above was conducted by my company, Prosper Insights & Analytics . This is the same dataset used by the National Retail Federation, and available from Amazon Web Services, Databricks, and the London Stock Exchange Group for economic benchmarking.