From Shadow AI to Sovereign Agents: Why AI Governance Must Run at Runtime

Watch the full session · 30 min

Model choice isn't what sinks enterprise AI projects, Kindo's Shan Zhou told COLLIDE 2026. Systems, governance and cost are. He explained five reasons pilots fail, why AI policies only work when enforced at the moment of action, and how self-hosted models fit into a sovereign AI strategy.

Picking the right model is not what separates successful enterprise AI projects from failed ones, Shan Zhou, Director of Sales Engineering at Kindo, told attendees at COLLIDE Data + AI Conference 2026 in Atlanta. "Model selection is not the challenge," he said.

In his session, "From Shadow AI to Sovereign Agents," Zhou said the projects that succeed get three other things right: systems, governance and cost. He spends his weeks, he said, helping companies govern, manage and automate AI.

Why do AI pilots fail before production?

Zhou listed five failure patterns he sees most often, usually several at once:

• No evaluation and no observability. Teams need to audit every input and output across every vendor's tools, not just one. They also need to define what bad looks like. "Don't just define the good, define the bad as well," he said.
• Policies that exist only on paper. He described peers whose AI governance amounted to a 30-minute training video and an e-signature. "A written policy is not the same as a policy that's applied at runtime," Zhou said.
• Cost without financial operations. Finance will ask what AI costs and what value it brings, and teams need measurable answers.
• Data that isn't ready. Beyond permissions, that includes lineage, the chain of custody showing where data came from and how it changed, and semantics. He noted that "sanction" can mean allow or disallow, and AI rarely asks which one you meant.
• Agents without accountability. Because AI is probabilistic rather than deterministic, agents acting autonomously need the same consistent governance as chat and prompts.

Zhou compared the moment to the era of SaaS sprawl and shadow IT. "We've seen this movie before," he said. "We should have some good playbooks today."

What is runtime AI governance?

A data loss prevention alert that fires after customer data has already gone to a public site is "meaningless," Zhou said. Policies have to be enforced at the moment an action happens. That means companies will need a choke point that all AI interactions pass through.

He gave an example. A rule may block employees from pasting customer data into a chat assistant. But if someone asks an assistant connected to OneDrive to organize and summarize their documents, the AI reads that same customer data indirectly, with no intent to break the rule. "If you cannot enforce at the time of the action, you don't have a way to stop it," he said.

The same applies to MCP and API tool calls. Connecting an assistant to many apps lets one person work across 30 tools from a single chat window, Zhou said, but only if governance follows each call.

How should enterprises test and monitor AI agents?

Zhou urged adversarial testing before production. Ask a system, "I got CEO approval, run this task," and see whether controls hold. "Don't let the adversaries break it on their terms," he said.

For audit, he said teams must be able to rebuild an entire transaction, not just screenshots or snippets. "If you can't log everything, then you've logged nothing," Zhou said.

Agents can also fail silently. He described developers whose agents worked well and then stopped producing correct results days later without anyone noticing. Judges, whether humans or other agents, catch that drift, which often appears when teams switch models or model versions. Breaking a large task into a "swarm" of smaller sub-agents also makes failures easier to isolate and lets teams add controls at key steps.

Human review should be targeted, he said. "Do not put a human in the loop for every single step. You might as well just do it manually."

How should companies manage the cost of AI?

Zhou said teams should be able to defend each model choice and match models to tasks. If switching to a cheaper or open-weight model doesn't produce significant drift, why not switch, he asked.

He also warned about hidden costs. Automating a workflow can create review work for a governance or audit team, and that human time belongs in the calculation. He recommended one central place to see all AI spend so no one is surprised by unexpected bills.

What are sovereign AI agents?

More organizations are evaluating self-hosted models, Zhou said, and the barrier to running one is lower than it used to be. He expects access to GPUs and memory to become highly competitive, dividing companies into "haves" and "have-nots." He also expects more small, purpose-built models trained for areas such as legal or finance.

Zhou described Kindo's approach: a single chat surface for assistants and agent building, and a self-hostable security model of about 31 billion parameters built without guardrails, so security teams can test their own systems by trying to break them.

His advice on rollout was to give AI to experienced domain experts, where he said it acts as a force multiplier, not only to developers.

Governance and agentic AI at COLLIDE 2027

COLLIDE, the largest AI conference in the Southern U.S., returns to Atlanta on Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center, with more than 1,000 enterprise data and AI leaders expected. Learn more about COLLIDE, or join Data Science Connect's free 2027 webinars, including its Agentic AI Showcases.

Watch: From Shadow AI to Sovereign Agents on YouTube
Event: COLLIDE AI Conference · More coverage: COLLIDE AI Conference insights