By Michael Oiknine
A newsletter from Foundation Capital landed in inboxes this week that every enterprise AI vendor should read carefully.
Ashu Garg pulled together two dinners — one with the co-founder of Databricks and the President of Technology at Visa, another with CIOs from Zuora, Asana, and BlackLine — and asked a simple question: what does it actually take to win enterprise AI deals in 2026?
The answers were more candid than most vendor content you’ll read this year. And for anyone selling into regulated industries, one theme dominated: governance is the next competitive moat.
Most Enterprises Are Stuck in AI Sprawl
Arsalan Tavakoli of Databricks described four stages of enterprise AI maturity.
- Stage one: not using AI in any meaningful way.
- Stage two: sprawl — dozens of pilots running with no clear strategy.
- Stage three: measurable productivity gains.
- Stage four: process redesign, where work is rethought from the ground up around what AI makes possible.
His assessment? The vast majority of enterprises are still at stage two. They have activity, not outcomes.
That tracks with what we see in financial services, insurance, and telecom. There is no shortage of AI experimentation. What’s missing is a clear path from pilot to production — particularly in customer-facing workflows where every interaction carries compliance risk, audit exposure, and regulatory consequence.
The Governance Problem Is Hiding in Plain Sight
Here is the part of the Foundation Capital piece that should stop every enterprise AI leader cold.
An agent inside Databricks was told the system was at capacity and to delete something. It found the nearest thing it had permissions to delete and brought the whole system down. It did exactly what it was told. What was missing was the judgment any human would have applied automatically.
Now transpose that story to a customer interaction in a regulated vertical. An AI agent handling a billing dispute at a utility. A virtual agent walking a customer through a claims process at an insurance carrier. A digital workflow guiding a telecom customer through a contract change. In each case, the agent is operating at the edge of compliance, where the cost of doing exactly what it was told — without human judgment — is not a system outage. It is a regulatory violation, a customer harm event, or an audit finding.
Saket Srivastava, CIO of Asana, framed the governance problem precisely: enterprises need to start thinking about the lifecycle of an agent the way they think about the lifecycle of a worker. When do you onboard them? How do you monitor them? How do you retire them? Most enterprises do not have answers yet.
The Model Layer Is Commoditizing. The Control Plane Is Not.
Rajat Taneja, President of Technology at Visa, made a point that deserves to be quoted directly: as the model layer commoditizes, governance becomes the differentiator. The value is moving into the control plane.
This is not a prediction. It is already happening. The underlying models — GPT, Claude, Gemini, Llama — are converging in capability. What separates a successful enterprise AI deployment from a failed one is not which model you chose. It is whether you built the infrastructure to govern what that model does when it touches real customers, real data, and real regulatory obligations.
That control plane problem is exactly what Callvu solves.
Legacy Architecture Was Not Built for This
Every enterprise application running today was designed around a human interface. A human read the screen, applied judgment, and took action within the boundaries of a defined process. Compliance was enforced because a human was in the loop.
When AI agents replace that interface, the architecture has to be rethought. The agent does not read screens. It does not apply contextual judgment about when a workflow is veering into legally ambiguous territory. It does not know that this particular customer interaction requires a specific disclosure, or that this step in the process cannot be skipped in a regulated state, or that the next action requires supervisor review before execution.
Callvu is built for exactly this moment. Our platform sits at the intersection of AI-driven customer workflows and the compliance obligations that govern them — providing the completion and control layer that ensures agents operating in regulated industries do not just do what they are told, but do what is permissible, auditable, and safe.
What This Means for Enterprise Buyers
If you are a CIO or VP of Digital in a regulated industry, the Ashu Garg piece offers useful guidance: pick a use case with bounded scope, measurable outcomes, and significant cognitive repetition. Customer service, dispute handling, and onboarding workflows check every box.
But running an AI agent on those workflows without a governance layer is where the pilot trap becomes a compliance trap. Pilots succeed in clean environments. Production is where real customers arrive with edge cases, where agents face decisions that require judgment, and where audit trails either exist or they do not.
The founders winning enterprise AI deals right now, according to the CIOs at those Foundation Capital dinners, are the ones who come in as partners — with a point of view, with a clear ROI thesis, and with the infrastructure to make the CIO look good when the regulator asks what controls are in place.
Michael Oiknine is CEO of Callvu, Inc. Callvu helps regulated enterprises deploy AI-driven customer workflows with the completion, compliance, and control infrastructure production demands. Learn more at www.callvu.com
