AI & TechnologyEnterprise AIControl Layer

Alation CEO Satyen Sangani Says AI’s Worst Failures Look Exactly Like the Right Answer

Alation CEO Satyen Sangani asserts enterprises don't have a model problem—they have a trust problem, as AI agents begin making decisions instead of simply answering questions.

Enterprise software has long operated on a simple principle: when something breaks, it leaves evidence. Failed transactions generate logs, deployments trigger alerts, and engineers can usually trace the source of the problem. AI agents, however, don’t work that way.

When an autonomous agent retrieves outdated customer data, applies an obsolete pricing policy or misinterprets a business rule, it produces an answer that sounds reasonable enough for an employee—or increasingly another AI agent—to act on. By the time anyone realizes something went wrong, the decision has already shaped the business outcome.

That, according to Alation CEO Satyen Sangani, is enterprise AI’s most dangerous failure mode. “Asking a question has become incredibly easy. Getting the correct answer is now the hard part,” he tells me. “The AI will almost always give you an answer. The challenge is knowing whether that answer is actually right.”

The platform recently unveiled the Alation Intelligence Operating System (AIOS), aiming to reposition itself from the company best known for enterprise data catalogs into what it describes as an operating system for enterprise intelligence. Rather than governing AI after an agent has completed a task, AIOS attempts to validate enterprise data, business context and governance policies while the agent is still deciding.

It’s the company’s biggest strategic shift since it pioneered the data catalog market—and it comes as Microsoft, Snowflake and Databricks compete to become the control layer for enterprise AI. The company argues enterprises don’t primarily have a model problem. They have a systems problem.

Why AI Agents Fail Differently Than Enterprise Software

Sangani traces the problem back to a fundamental change in how enterprises consume information. For decades, employees accessed business data through reports, dashboards and catalogs. Today, AI increasingly sits between enterprise data and the people making decisions.

“Historically, we created our data inside deterministic systems, typically databases, and delivered that information to people,” Sangani says. “Today, increasingly, the human isn’t opening Tableau or Power BI to analyze data anymore. They’re sitting in front of ChatGPT, Claude or Copilot and asking a question directly. The AI is now doing the work that a human analyst used to do.”

That shift, Sangani argues, exposes a deeper problem. AI behaves fundamentally differently.

“I think AI is actually a little bit more like data than it is like software,” he says. “Large language models are fundamentally stochastic systems. Sometimes that’s perfectly acceptable. If you ask where Abraham Lincoln was born, ‘United States,’ ‘US’ and ‘United States of America’ are all correct answers. But enterprise data doesn’t work like that. If you ask, ‘What were my revenues yesterday?’ and one answer is one, another is ten and another is one hundred, only one of those answers is correct. In enterprise environments, correctness isn’t subjective.”

He explained that enterprise AI typically breaks in three ways: agents consume stale or inaccurate data, apply outdated business context or business rules, or gradually drift as prompts, tools and surrounding systems evolve. None of those failures necessarily produces an error message. Each can quietly influence a business decision before anyone realizes something has gone wrong. Moreover, Sangani argues enterprises also can’t benchmark their way out of the problem.

“You might test a model against one hundred questions and conclude that it answered ninety-seven correctly,” he says. “But businesses don’t fail on the hundred questions you anticipated. They fail on the hundred-and-first question—the scenario your evaluations never covered. You can’t test infinitely. Eventually an AI system encounters something new, produces an answer that sounds confident and plausible, but is fundamentally wrong.”

The implication is that governance can no longer remain a postmortem exercise. Instead, organizations need systems that verify data, context and outputs while the agent is still reasoning.

Enterprise AI ROI: From Experiments to “Hard Yards”

Sangani was careful not to blame the frontier model companies themselves. “They should absolutely focus on benchmarks and model capability because that’s the business they’re in,” he says. “The challenge begins when enterprises assume those benchmarks automatically translate into reliable business decisions.”

Just a few months ago, the conversation centered on productivity gains and experimentation. Today, executives are asking a more fundamental question: “We’re spending millions, tens of millions, and in some cases hundreds of millions of dollars on AI. What are we actually getting back?”

“If Anthropic becomes a $100 billion company, they’re bigger than Oracle, SAP, IBM and Salesforce,” he says. “That money is coming from enterprise IT spend.” Bigger budgets alone won’t close the gap, he adds. “Simply increasing AI spending isn’t going to produce results on its own. Getting to ROI requires what I’d call the hard yards.”

Even if the world’s most capable frontier model became free tomorrow, Sangani asserts, enterprises would still face the same challenge. “Better models don’t eliminate those problems,” he says. “They simply make solving them more important.”

From Data Catalog to Intelligence Operating System

That conclusion forced Alation to rethink what kind of company it wanted to become. “We used to be a destination,” Sangani says. “People would come to our catalog to answer questions.” Now the front door belongs to whichever AI assistant an organization chooses. Rather than competing to own the user interface, Alation wants to become what Sangani describes as “the invisible layer underneath,” delivering trusted enterprise data behind them all.

The company’s acquisition of Numbers Station in May 2025 accelerated that transition, but Sangani rejects the suggestion that Alation simply reinvented itself to fit the AI boom. “We didn’t wake up one day and decide we needed an AI narrative,” he says. “We responded to a fundamental change in how enterprises consume information.”

The thinking also explains why Alation chose the term “Intelligence Operating System.” Sangani says the analogy is intentional. “In the same way an operating system lets you swap a monitor, keyboard or mouse, our operating system lets customers use any compute layer, any model, any storage layer or any semantic system.” The goal is to make governance independent of whichever frontier model happens to lead this quarter’s benchmarks.

Who Governs the AI Governance Layer?

Alation isn’t the only company trying to become enterprise AI’s control layer. Microsoft is embedding governance inside Copilot and Azure AI. Snowflake argues trusted AI begins with governed enterprise data inside its AI Data Cloud. Databricks has positioned Unity Catalog as the policy engine for enterprise AI. The industry largely agrees on the destination. The disagreement is over where governance belongs.

Alation is making the case for an independent layer that sits above models, clouds and enterprise platforms. That strategy makes sense if organizations continue running heterogeneous AI environments, mixing OpenAI, Anthropic and Google models across platforms like Snowflake and Databricks. It becomes harder to justify if governance eventually becomes another capability bundled into whichever ecosystem already owns the rest of the enterprise AI stack.

AIOS describes itself as an open, governed and self-improving operating system. But a governance layer that continuously adapts through feedback loops is itself an intelligent system. If that system evolves alongside changing data, policies and business context, how do enterprises verify that the control layer itself hasn’t drifted? At some point, the software responsible for validating AI decisions inherits the same risks it was designed to prevent.

For Sangani, governance should provide enterprises with the transparency and evidence needed to make informed decisions—not dictate those decisions. “The final decision about acceptable risk,” he says, “belongs to the enterprise deploying the technology—not the AI vendor and certainly not the AI itself.”

Control Layer Verdict

Sangani has identified one of enterprise AI’s biggest blind spots. AI’s most dangerous failures are the confident decisions built on stale data, outdated context or obsolete business logic. Runtime governance is a step in the right direction, and much of the industry is heading there. But every adaptive governance layer eventually becomes another intelligent system capable of drift. The AI industry has spent almost two years building referees for AI agents. It still needs a referee for the referee.

Author:

Related Articles

Back to top button