Companies are giving AI agents greater responsibility for business decisions while leaving some of the rules behind those decisions unresolved. Even a basic question about the number of active customers can produce different answers. The sales department might count customers who continue using a product, while finance might count those who generated revenue during the quarter. An agent who chooses the wrong one can calculate the total correctly and still mislead the person asking.

The risk lies in how convincing that answer can look. A precise customer count gives little indication that the agent has made an assumption about how the company measures its business.

Alation is betting that closing the gap between enterprise data and business understanding will create a larger role for its software. The company, which built its reputation helping organizations catalog and understand their data, announced six new or expanded products for its Alation Intelligence Operating System, or AIOS, at its revAlation conference. The open platform, launched in July 2026, connects agents to enterprise data and the business context they need to interpret it. Alation wants to govern that knowledge and use agents’ mistakes to improve the guidance they receive.

Positioning AIOS as an “intelligence operating system” signals a bid to extend that business into governing how AI agents interpret and use enterprise information. However, the ambition also brings Alation into closer competition with Snowflake and Databricks , which build tools to help AI agents interpret company data correctly and follow rules about what they can access or do with it.

Satyen Sangani, Alation’s co-founder and CEO, argues that companies should resist treating an AI model’s growing capabilities as a reason to give it more authority. He advocates “a start point of zero trust,” with companies establishing where AI proves useful before expanding its responsibilities.

“I don’t approach AI from a maximalist perspective, and I think many of our customers don’t either,” he told me in an exclusive interview. “There’s a lot of wasteful output from these models. They’re not always concise, and they’re not always useful. You should start by asking whether you have a clear sense of what changes when you introduce AI into the work. But when it comes to letting AI make the final decision, I think that’s ultimately where human judgment comes in.”

The uncertainty extends to executive oversight. Grant Thornton’s 2026 AI Impact Survey of 950 c-suite leaders found that 78% lacked strong confidence they could pass an independent AI governance audit within 90 days.

AI Agents Need More Than A Single Source Of Truth

Alation’s Semantic Model Mastering aims to put approved business definitions directly into the platforms where agents work. It brings definitions from Snowflake and Databricks into Alation’s catalog, where owners can enrich and approve them. The company says it then synchronizes those definitions back to both platforms. The process gives companies a way to coordinate meaning across systems, although deciding what that meaning should be remains a human responsibility.

That distinction becomes consequential when two departments disagree. Forcing them to adopt one definition could erase information an agent needs to interpret the business correctly.

Sangani points to “on-time in full,” a supply chain measure of whether an order arrives on schedule and in the expected quantity. Measuring timeliness at delivery can produce a different result from measuring it when the recipient records receipt.

“Ultimately, the real world arbitrates that decision. We’ve had customers who have different definitions of on-time in full,” he says. “There are nuances in these calculations, and it is legitimate that there might be different ways of calculating it.” An agent needs to understand which calculation serves the task at hand. “We as a company can simply highlight that these things are similar, or that they’re the same and should be consolidated. But ultimately, the consolidation decision is a decision for the customer to make,” he explained.

Ontologies extend that effort from individual definitions into business processes, expressing operating rules in machine-readable models that subject-matter experts review. Governed Collections addresses the documents behind those processes, giving agents access to policies and operating procedures through references that follow updates at the source.

The Gap Between Monitoring And Controlling Agents

Alation’s AI Governance aims to show companies when problems in their data put an agent’s work at risk. It builds a shared registry of agents and models through six connectors covering Snowflake Cortex, Databricks MLflow, Microsoft Copilot Studio, Microsoft Foundry, Amazon SageMaker and Amazon Bedrock, then maps the data each registered agent relies on. If a dataset used to prepare a report fails a quality check, Alation says the system updates the agent’s risk assessment. That could warn the team that the report rests on unreliable information. Acting on that warning requires another decision, however, about whether to pause the agent’s work and which system has the authority to intervene.

“If you had a long enough context window, you could give the agent every iota of knowledge it would need,” Sangani says. “If you made the context that complicated, then it would get confused. And so a big part of what we do is that we give people observability on the outputs.”

By observability, he means showing users which tools an agent called and what information it consulted to produce an answer. The platform then helps identify the source of a mistake and recommends where to record the correction. Likewise, Console and Intelligent Feeds tools bring that oversight closer to employees. Console lets users describe what they need and routes the request to the appropriate AIOS capability. Intelligent Feeds combines recurring business metrics with relevant written context, allowing employees to ask follow-up questions about the results.

“Sometimes you’ll just let the agent do whatever it’s supposed to do because you can just automatically rubber-stamp it because you have high confidence,” he says. “And in other cases, you might want to say, no, that may not be the right suggestion. But the feed is what gives you the ability to arbitrate between automated action and human-in-the-loop style action.”

Sangani said customers can approve changes and restore an earlier trusted version. He did not specify how restoring that version in Alation would affect connected systems that had consumed the update.

A Crowded Race To Govern Enterprise AI

Snowflake and Databricks offer their own business definitions and governance. Likewise, Collibra’s AI Command Center includes an agent registry and a dynamic trust score, and Atlan offers Active Ontology and tools that generate context for AI.

Alation claims that its advantage lies in testing agents against the business definitions they use, then using incorrect answers to identify gaps in that guidance. The company says it continuously updates an agent’s risk assessment as the quality of its source data or the policies governing that data change.

Sangani says vendors make similar promises and offer features that sound interchangeable, putting greater pressure on them to prove their value in practice. He argues that customers should judge the technology by whether people actually use it and whether it improves their work.

For Alation, that case becomes stronger when a company needs consistent business definitions across several platforms or wants to change providers without recreating the knowledge its agents depend on. “All of the definitions are available via API, and you can certainly use batch capabilities to upload and download them,” Sangani explained. “So we try to make it pretty clear and easy to move knowledge.”

He notes, however, that no universal open knowledge standard exists. Exporting a definition does not guarantee that another system will interpret it identically.

AIOS Faces The Enterprise AI ROI Test

The company says five customers have used AI Governance, while more than 20 use Semantic Model Mastering. Each of the four early-access products has three to five design partners. All six fit within existing pricing without requiring new contracts, although customers face the work of putting them to use.

An unnamed video game company runs more than 30 data products, reusable packages of data and business context that support dozens of agents, according to Alation. Euromonitor provides a named example of broader AIOS adoption. Alation says its customer-facing natural-language product generates tens of thousands of interactions.

Sangani argues that the measurement should begin before a customer buys into the technology.

“We have to engage with the customer very specifically to define what the before and after state happens to be,” he says. He applies similar scrutiny to the models Alation uses internally. “I don’t think that you need the frontier models to do most tasks,” he says. “We try to be very thoughtful about spend. We try to be super thoughtful about model selection. We’re doing that even more and more because I think it can get super expensive and it can also prove that there’s not a lot of ROI if you do it wrong.”

Alation’s customers will have to settle disagreements about what their data means and take responsibility when definitions change. Governing the information behind an answer can strengthen the case for trusting it, but the responsibility for deciding whether to act on it remains with the business.