“Safety is not the brake pedal; it’s the gas pedal,” Nima Badiey told the audience at Commvault’s SHIFT 2026 conference, borrowing a frequently used line from CrowdStrike CEO George Kurtz.
Badiey, who leads strategic AI product partnerships at NVIDIA, argued during the Driving Enterprise Value with AI panel that safety is what enables enterprises to move faster on AI, not what holds them back. Throughout the panel, Badiey and executives from Microsoft and Databricks described what the ability to press that gas pedal looks like: model explainability, cost governance, identity controls, and secure runtimes as the disciplines that determine how much authority an enterprise can grant to its AI agents as well as how much value the organization can expect back.
Pranay Ahlawat, panel moderator and Commvault’s chief technology and AI officer, said enterprises will soon operate alongside 2 billion to 3 billion agents, and that eighty percent of organizations say they have already encountered risky agent behavior. The most AI-mature companies, he added, generate 2.4x more revenue than the bottom half.
Effective agentic governance drives security and value
Stephen Orban, senior vice president of product ecosystem at Databricks, said Databricks ’ customers are seeking answers to those core tenets laid out by Badiey, such as which employees can use which models, what data they give those models, where that data goes, and what the agents do with it. His answer? Explainability through traces and lineage: records of what agents did and how they reasoned as they acted on behalf of employees or customers.
Standard and Poor’s, a Databricks partner, is building agents into its product offerings and has product managers who “actually love those traces,” Orban said, because they show what questions customers ask, where agents get tripped up, and which queries need fixing to produce better answers. The Standard and Poor’s team, he explained, runs the traces as a continuous feedback loop.
That’s an example of how “security-as-enabler” can play out. The instrumentation a CISO requires for accountability also enables product managers to make their commercial AI product better. One investment produces oversight and product intelligence.
To keep costs in line, Orban said enterprises can set AI budgets at the user, team, and agent level. For instance, a software engineer might get $1,000 a day, he said, while a finance analyst gets $100. He also described a graceful fallback as when a user hits their cost threshold, rather than blocking access, the work routes from the frontier model to a cheaper open-weight model.
Governance designed that way does not gate people out of AI. It keeps them working inside organizational limits.
Identity security is central to agentic AI security
Badiey said most of the tooling needed to guardrail AI already exists. For instance, role-based access controls live inside data platforms, identity systems govern access, and safety, he said, is “first and foremost an engineering problem” that gets built into tools rather than bolted on.
Another point Badiey made is that no company should try to build AI governance and security alone. NVIDIA launched the Open Agent Safety Platform, he said, and contributed OpenShell, a secure runtime for agents, with bindings for NVIDIA Sentry that offload agents into a secure hardware environment on the company’s Bluefield data processing units. “We don’t think one company has all the answers,” Badiey said. “Don’t try to go out alone.”
Aung Oo, vice president of product management for Azure Storage at Microsoft, asked rhetorically whether today’s identity and recovery architectures are ready for agents with real permissions. “I personally don’t think we are ready,” he answered.
The central challenge is borrowed identity. Such as letting an agent run in production on an employee’s credentials, Oo said. That, he added, is like “sharing my username and password with my wife. A lot of things can happen. Not a good idea.” He described seeing agents prototyped on their creators’ credentials go on to do “all the crazy things” with every action logged under the employee’s name.
Oo suggests treating agents as a new class of employee: a distinct identity, a named human sponsor accountable for access and lifecycle, an assumption of breach, backed-up audit trails, protected data, and what he called a “big, big red switch” for when behavior turns abnormal.
While Badiey said the components exist, Oo added that the assembled architectures lag. Closing that gap is the immediate enterprise agentic AI security challenge: the controls exist, and the work is wiring them together.
Enterprises need to gear up to effectively manage the agentic “army” heading their way
Commvault’s Ahlawat described the framework that he said explains the 2.4x gap between agentic leaders and laggards. Citing a BCG model, Ahlawat put 10% of AI transformation in algorithms, 20% in surrounding technology, data infrastructure, and resilience, and 70% in change management and organization. “Get the 70% right,” he said to the room. Formula 1 cars, he added, use the best brakes in the business: that is what enables them to go so fast.
Enterprises better get their brakes installed and checked regularly if Databricks ’ Orban’s prediction proves accurate: A year from now, he said, thriving companies will have every employee orchestrating an army of agents that produces more code and content in twelve months than the company created in its entire prior history.
If anything close to that prediction comes to fruition, the differentiator will not be model access, as every enterprise can buy the same intelligence at roughly the same price, as Ahlawat put it. The differentiator will be how much proprietary data and operational authority an organization can confidently place behind those agents. The enterprises that have the best “brakes,” or can trace, budget, sponsor, and shut down agents, can reach greater agentic speeds through granting agents more access, more tasks, and more autonomy than a competitor that hasn’t gotten as far down their agentic AI governance buildouts.
Brakes and throttle are separate systems, and any car, Formula 1 or not, only goes as fast as both allow.
