At COLLIDE 2026, NexusOne CEO Anu Jain argued that AI has no control plane yet, leaving agents loose in brownfield data estates. He laid out three questions every leader should answer, the case for one identity across all systems, and why companies should own their ontologies.
Key takeaways
- Three questions leaders can't answer. Jain said few can say what data AI touches, prove it stayed within the rules, or state AI's full cost.
- AI lacks a control plane. Compute got Kubernetes and data got a patchwork of catalogs, but AI has nothing comparable yet, he said.
- Agents need context and boundaries. Open context with open boundaries is "a toddler with a knife" in the data estate.
- One identity for every actor. Jain called for a single identity and security model across people, services and agents in every system.
- Own your intelligence. A company's ontology is how its business thinks, and Jain said it shouldn't be handed to vendors.
Enterprises have pointed AI agents at nearly everything in their data estate, from customer contracts to pricing and risk data, without a way to see or control what those agents touch. That was the warning from Anu Jain, CEO and Founder of NexusOne, in a morning keynote at COLLIDE Data + AI Conference 2026, held October 1 at the Sandy Springs Performing Arts Center in Atlanta.
Jain, whose career has included MicroStrategy, Deloitte, the original IBM Watson team and Teradata, called the talk partly "an apology tour" for the fragmented data stacks his industry helped build.
What questions should leaders be able to answer about their AI?
For years, Jain said, AI was asked to answer: summarize a document, write some code, respond to a question, with a human deciding what happened next. Now AI is being asked to act, querying systems and executing workflows.
That shift means leaders should be able to answer three questions, he said. What data does our AI touch when it answers a question? Can we prove that data stayed inside our rules, for example if a regulator asked where PII went? And how much does AI actually cost, counting context, security, data and infrastructure, not just tokens?
When he asks chief data officers and chief AI officers those questions, "the room actually gets quite quiet," Jain said, because nothing in today's data stack was built to answer them. He cited a survey by Enterprise Management Associates of 202 large data companies in the U.S., Canada, the U.K. and France. Over half had spent more on AI than projected, and one in four had true data sovereignty, meaning retrieved data stays fully inside their own environment.
Why does AI need a control plane?
Jain traced the pattern through past technology waves. When compute got messy, Kubernetes became its control plane. When data got messy, companies got a patchwork of catalogs and governance tools, one per product. "When AI's gotten messy, we don't actually have a control plane yet," he said.
The same survey found organizations use more than eight tools on average to govern their data estate, and updating a business rule usually means changing it in more than two systems, from semantic layers to stored procedures to ETL. Add decades of on-premises systems, mainframes, multiple clouds and acquisitions, and "every enterprise stack now is a brownfield," he said, including at digital natives.
What happens when AI agents get open context and open boundaries?
Jain illustrated the problem with a specialty insurer NexusOne works with, which he described as about five years old with about $1 billion in revenue. A claims adjuster checking a policy start date may find it unclear in the policy system, open in the claims system, lapsed in billing and flagged in the fraud system, with key details buried in unindexed documents and an old email. The adjuster is the company's "shock absorber" who knows what to trust, he said, and companies now expect an agent to do the same.
"AI needs complete context and hard boundaries," Jain said. Most companies instead give agents open context and open boundaries. "There is a toddler with a knife running around your data estate," he said.
His answer is one identity for every person, service and agent across the whole estate. Platforms like Snowflake and Databricks provide context and security inside their own environments, he said, but most enterprises are heterogeneous and need one context layer and one security model across all of it.
How does a data and AI control plane work in practice?
Jain described NexusOne as a data and AI control plane built around four areas: sovereign infrastructure that can run in a data center, cloud or air-gapped environment; an integrated stack that brings the engine to the data; one identity layer; and intelligence the customer owns.
In the claims example, the agent first takes on the identity of the person invoking it. It then routes each part of a question to the right tool: a database for deterministic queries without spending tokens, a machine learning model that returns the same answer every time, a deterministic business rules engine, and a small internal or frontier model only when needed. "Rules are projections on the data," he said. Every step is recorded with lineage, including attempts to reach blocked data such as PII.
Why should enterprises own their ontology?
Jain argued that a company's ontology, such as how billing and underwriting each decide whether a policy is in force, is "the way your business thinks today." That logic often lives in someone's head, a dashboard's semantic layer or a stored procedure. He called it sovereign intelligence and said companies shouldn't give it away to vendors.
NexusOne's Auto LLM feature lets a business user pick a domain, choose an open-weight model and train it on internal documents and catalog data, Jain said. For the insurance client, it took under six minutes on a couple of GPUs.
He closed with three principles for software in the AI era: open, composable and sovereign. He also described what NexusOne calls a 5-5-5 method, aiming for setup in five minutes, a deployed environment in five hours, a first governed workload in five days and enterprise production in five weeks. He said NexusOne serves as the data and AI control plane at three of the top 25 banks and the largest telco in Africa.
See more enterprise AI architecture at COLLIDE 2027
COLLIDE, the largest AI conference in the Southern U.S., returns to the Sandy Springs Performing Arts Center in Atlanta on Wednesday, October 13, 2027, bringing together more than 1,000 enterprise data and AI leaders for keynotes like this one. Between conferences, Data Science Connect's free webinars dig deeper into AI governance and data architecture.
Frequently asked questions
What is a data and AI control plane?
As Jain described it, a layer across the whole data estate that gives every person, service and agent one identity, routes requests to the right engines, and records every step with lineage and audit trails.
Why do AI agents need hard boundaries?
Jain said agents need complete context to answer correctly but should only see what the person they act for is allowed to see, with blocked access attempts logged.
What is sovereign intelligence?
Jain's term for a company's ontologies, semantics and business processes, which he said enterprises should own rather than embed in other vendors' products.
How much do enterprises overspend on AI?
Jain cited an Enterprise Management Associates survey of 202 large data companies in which over half had spent more on AI than projected.
Watch: Keynote: The Missing Control Plane Your AI Needs Before It Scales on YouTube
Event: COLLIDE AI Conference · More coverage: COLLIDE AI Conference insights
