AI Governance Cannot Be Solved by Blocking Models, Says Kimchi Core Contributor

Enterprises cannot govern AI by simply blocking access to models, as employees may resort to using personal accounts and create greater data and security risks, according to Kunal Das, a core contributor to Kimchi, an open-source, multi-model AI governance platform.

The bigger challenge is that AI does not behave like conventional enterprise software. The same prompt can produce different answers, making it difficult to predict outcomes and govern the technology through traditional controls, Das said.

Instead, enterprises need to build governance into the infrastructure layer, particularly the routing layer between an employee's device and an AI model. This is where organisations can control which models are accessed, how much they cost, what data is sent to them and which actions require human intervention.

Das, Developer Advocate for APAC at Cast AI and a core contributor to Kimchi, made the case during his talk on ‘How to Govern AI, Even If It's Hard to Predict’, at Cypher 2026. 

The economics of AI use are already making this a governance problem. Das said his team of about 200 engineers spent nearly $200,000 on Claude Code over a year without establishing a clear link between AI usage and productivity.

“More AI use” does not necessarily translate into more output, he said.

Match the task to the model

Das argued that enterprises should stop treating access to the most capable model as the default. Most routine tasks, such as fixing bugs or summarising PDFs, do not require a flagship model.

He advocated a “matching task to model” approach, where AI workloads are routed to models based on their complexity, cost and requirements.

Das cited a workshop involving 100 people in which participants built projects over three hours for about $54 when their requests were first routed through a cheaper-model system. Running the same work exclusively on a flagship model would have cost about $3,800, he said.

“We use a very small cheaper model to decide which model is best for that prompt. That changes the game,” Das said.

This kind of routing can also allow enterprises to control AI spending before costs accumulate, rather than auditing usage after the fact.

Four layers of AI governance

Das outlined four areas enterprises should focus on: spending, perimeter, resilience and control. He said the approach puts governance below the model-output layer. 

Instead of trying to determine whether every AI response is correct, enterprises can control the infrastructure through which models receive data, execute tasks and return results.

Das compared the shift to the early days of cloud computing, when organisations adopted large amounts of compute before developing mature systems for tracking and controlling usage.

Kimchi is also working on a “teleport” feature that can move long-running agent tasks to a temporary cloud sandbox, allowing work to continue without being tied to a developer's laptop, Das said.

Governance cannot fix everything

Das also acknowledged the limits of AI governance platforms. Review and monitoring tools can provide visibility into AI usage, but they cannot automatically fix poorly written code or prompts.

Data already sent to an external AI provider presents an even harder problem. Once sensitive information has been shared, an organisation may not be able to recover or control how the provider subsequently handles it.

“We can't fix all those things,” Das said.

The implication is that AI governance cannot depend entirely on the model itself. Enterprises need controls around the models that determine where data goes, which model handles a task, how much the task costs, and when a human must take over.

“You cannot predict AI,” Das said. “But we can make sure that the decision that we're taking is properly covered and we have all the things in place to track it.”