Many vendors in enterprise data are now selling context, Alation, Snowflake, and Databricks among them.¹ The layer that tells an agent what a metric means, where the number came from, and which definition the business actually uses has become the most contested position in the category.
What is far harder to find in anyone’s public material is a portable record of what an agent was refused, by whom, and when.
The result is an uneven kind of readiness. Grant Thornton found that organizations deploying AI often cannot show how decisions are made or who is accountable for the outcome.³ An organization in that position has instrumented half the problem. In our view, the half it skipped is the half that separates organizations that can defend their AI from those that cannot: the organizations pulling ahead can explain how their AI reaches decisions, name who owns the results, and describe what happens when it fails.³
Forbes put the authority question in print
The sharpest version of this critique came from an analyst and Forbes contributor¹² who had just written about Alation’s own agent governance products.
In September, Victor Dey published an exclusive interview in Forbes on the six products Alation added to AIOS at revAlation, from Semantic Model Mastering to an AI governance registry that connects to agents across Snowflake Cortex, Databricks MLflow, Microsoft Copilot Studio, Microsoft Foundry, Amazon SageMaker and Amazon Bedrock.⁴ It was not a soft story.
Dey opened it on a problem that sounds too small to matter. Ask how many active customers the business has, and two departments can give different answers, each defensible. Sales might count the people still using the product. Finance might count the people who generated revenue in the quarter. An agent picks one of them, calculates it perfectly, and hands back a number whose precision hides the assumption underneath it.⁵
What Dey’s reporting kept circling was the question one layer up. Alation says its registry updates an agent’s risk assessment when a dataset behind a report fails a quality check. Acting on that warning is a separate decision, as Forbes put it, about whether to pause the agent’s work and which system has the authority to intervene.⁶
He made the broader version of that argument when he joined AI Radicals (Season 4, Episode 11).¹¹
Context tells an agent what the data means. It says nothing about what it may do.
Dey agrees that context is the backbone of a working agent, which is what makes his objection to stopping there worth taking seriously.
“If you look at it and if you want to understand how an agent works, it’s all about context… It needs to have the right understanding from the right data and from a clean source and the right amount of data,” he told host Satyen Sangani. He anchors it in a basic rule every engineer and data scientist already knows: “garbage in and garbage out.”¹¹
Then he complicates it: “What I think is more important than context, or equally important as context, is also permissions,” Dey said. “What data did it touch? What points did it touch? Was it denied, or was it told not to perform certain actions? If not, who did it? At what moment?”¹¹
The evidence many enterprises may struggle to produce: a record of refusal
Companies rarely tell the story of an agent that was told no. That silence reflects a wider problem: many organizations aren’t confident they can show how they govern their AI.
“We always talk about what agents have achieved…" says Dey. "But I want to hear about times where their agentic implementations failed, or a particular agent was denied doing a particular task,” he said. “Who denied it? What happened behind it?”¹¹
That is an analyst’s question, but it is also an auditor’s. In our view, a serious review does not only ask what a system did. It asks what the system was prevented from doing, and who held the authority to prevent it. The closest signals point one way: Grant Thornton’s 2026 AI Impact Survey of 950 C-suite and senior business leaders found that 78% lacked strong confidence they could pass an independent AI governance audit within 90 days.⁷ The same survey found that only 20% of organizations have a tested incident response plan for when AI fails.⁸
Dey is not an alarmist about the schedule. He thinks many enterprises are still working through pilots, that mid-size companies and SMBs are not yet in these vendors’ consideration set, and that it will be another eighteen months to two years before anyone can see real adoption numbers. What he does think is already happening is proliferation: “Agents are going berserk, and agents are growing,” he said.¹¹
Those two observations point in the same direction. The agent population is expanding now, and the reckoning that measures it arrives later. This makes the present the cheapest moment to start keeping the record.
Lineage can tell you where an answer came from. A record of what an agent tried and was refused is a different artifact.
Monitoring an agent is not the same as being able to stop one
Observability tells you what an agent did after it acted, which is a different capability from the authority to prevent it from acting. Today, many architectures have the first without the second.
Sangani put the open questions on the table himself: “Which agents are part of the control plane? When are they not a part of the control plane? How do I even know what the agent is going to do? How do I control for what controls I give it?”¹¹
Dey’s answer puts four capabilities on equal footing. “But there’s also the equal importance of permissions and guardrailing and observability and transparency for what the agent touched, what it did, why it did it, and on what data,” he said.¹¹
The architectural problem underneath this is plain once you draw it. Alation’s registry can flag that a report rests on failing data across six connected platforms. Whether to pause the agent, and which system has the authority to do it, is a separate decision.⁶ A warning crosses platform boundaries easily. Authority does not.
Forbes noted that who holds the authority to intervene remains an open question.⁶ Sangani told Forbes that no universal open knowledge standard exists, so exporting a definition does not guarantee that another system will interpret it identically.⁹ If that is true of meaning, it is plausibly harder for authority, which may be one reason the permissions half stays under-built.
Why context gets the attention, and permissions don’t
The imbalanced focus on context over permissions is not an oversight. It is the predictable result of a demand curve that activates only after failure, and of standards that do not yet carry authority across systems.
Start with the demand curve. Dey illustrates it with his own writing. Say he covers a data governance story. When something goes wrong later, “I see people coming back and revisiting that story from two weeks ago, and all of a sudden it has forty thousand more views,” he said.¹¹ Permissions become urgent in the week after an incident and theoretical again by the following quarter.
Finally, portability. Sangani’s point to Forbes, that no universal open knowledge standard exists, matters here.⁹ If that is true for meaning, it is worse for authority. A permission that does not travel between systems is a preference. Teams building toward a data products marketplace can run into this the moment the same product serves agents on two platforms.
Four questions worth answering before the incident
A data leader does not need a new maturity model to close this gap, but does need to be able to answer four questions about the agents already running: Which agents currently touch governed data? What is each one allowed to do? Who approved each action? Where are refusals recorded, and in what form? Which system has the authority to stop an agent mid-task?
Working through those questions can produce a shorter list of trusted agents and a longer list of unowned ones, which is the point. It also matches how Sangani argues enterprises should expand an agent’s remit in the first place, starting from what he called “a start point of zero trust” and widening scope as the agent earns it, rather than widening it because the underlying model improved.¹⁰ Dey arrives at the same place from the other side of the table, and for a plainer reason: “Models are powerful, but not that powerful,” he said.¹¹ Defining critical data elements and a clear data product operating model is what makes those four questions answerable instead of rhetorical.
Permissions are following governance’s own trajectory
Dey says data governance was under-covered until incidents made it urgent, and permissions may be on the same path now.
He says he saw the same curve in quantum computing, which he started covering four or five years ago, when only a few authors did. “Even data orchestration or data governance… I think it was unexplored before, but now when incidents happen… people started talking about it more,” he said.¹¹
The enterprises that build a denial record before they need one are the enterprises that will not be assembling it under audit. Everyone else may end up reconstructing it from logs, under time pressure, for a regulator.
Dey’s framing of the pair is the plainest anyone has offered. “It’s basically like telling a child… that you know this, but you can’t do this no matter what happens.”¹¹
Both halves, or neither.
Curious how governed context and trusted data can support your agent strategy? Book a demo with us today.
SOURCES & NOTES
Every external claim on this page is independently verifiable. The public sources are listed here.
Snowflake and Databricks offer their own business definitions and governance; Atlan offers Active Ontology and context tools; Alation’s AIOS connects agents to enterprise data and business context. — Forbes, Sept 17, 2026.
A vendor says its runtime capabilities enforce policies where agents run and block violating actions before they happen. — Vendor press release, Oct 5, 2026 (accessed Oct 19, 2026).
Organizations deploying AI often cannot show how decisions are made or who is accountable; leading organizations can show how decisions are made, who owns outcomes, and what happens when something goes wrong. — Grant Thornton, 2026 AI Impact Survey Report.
Exclusive interview; six new or expanded AIOS products announced at revAlation; AI governance registry with six connectors (Snowflake Cortex, Databricks MLflow, Microsoft Copilot Studio, Microsoft Foundry, Amazon SageMaker, Amazon Bedrock). — Forbes, Sept 17, 2026.
Active-customers example (sales vs. finance definitions). — Forbes, Sept 17, 2026.
Alation says a failed quality check updates an agent’s risk assessment; acting on the warning requires a separate decision about pausing the agent and which system has the authority to intervene. — Forbes, Sept 17, 2026.
78% of executives lack strong confidence they could pass an independent AI governance audit within 90 days; survey of 950 business leaders (C-suite and leaders reporting directly to the C-suite), fielded Feb 23–Mar 18, 2026. — Grant Thornton, 2026 AI Impact Survey Report.
Only 20% of organizations have a tested AI incident response plan. — Grant Thornton, 2026 AI Impact Survey Report.
Sangani: no universal open knowledge standard exists; exporting a definition does not guarantee another system will interpret it identically. — Forbes, Sept 17, 2026.
Sangani advocates “a start point of zero trust” for expanding AI’s responsibilities. — Forbes, Sept 17, 2026.
Quotations and paraphrases from Victor Dey and Satyen Sangani, AI Radicals, Season 4, Episode 11. Sections: “Is context for AI agents a passing buzzword?”; “Are permissions as important as context for AI agents?”; “The AI agent failure stories journalists should be covering”; “Why do AI agents need governance, observability and guardrails?”; “The case for an enterprise AI agent control plane, and the evidence so far”; “Why are so many enterprises still stuck in AI pilots?”; “What should enterprises fix first to get value from AI agents?”; “Why do AI and data topics only get attention after an incident?”; “How incidents and AI chatbot traffic change who reads governance coverage.” — Alation.
https://www.alation.com/podcast/episodes/ai-agent-control-plane-victor-dey/
Victor Dey is a Forbes contributor and an analyst and writer covering AI and emerging tech. — Forbes, Sept 17, 2026.
- AI
- Active Data Governance
- Data Catalog
- Data Governance
- Data Products
- Digital Transformation
Keep reading
More from the data desk
