Airrived's Anurag Gurtu on real AI agents vs expensive autocomplete

Airrived, the company behind an enterprise Agentic OS and sovereign AI platform, was named a Cool Vendor in Gartner's Coolest Vendor Innovations in Agentic AI Security report earlier this month. 
 

Its platform lets business and domain experts build, deploy and govern autonomous AI agents across cybersecurity, IT and enterprise operations, whether in the cloud, on premises or in fully air-gapped environments.

Anurag Gurtu, Airrived's co-founder and CEO, argues that much of what the industry calls agentic AI is nothing of the sort. He told The Datatech Times how to tell the difference, why kill switches miss the point, and what boards should be protecting instead of patents.

Does it complete work, or describe it?

Gurtu's test for a genuine agent is deliberately demanding. “Ask it to do something that requires three systems and a decision it wasn't pre-programmed for. A real agent executes, observes the result, and adapts. An LLM prompt in a suit generates a recommendation and waits for a human to do the work.”

The tell, he says, is simple. “Does it complete work, or does it describe work? If a human is still the last mile, you don't have an agent. You have an expensive autocomplete.”

A fire extinguisher is not a fireproof building

That view shapes his scepticism about proposed AI kill switches, which he sees as a fix designed for the wrong moment. “A kill switch is what you build when you've already lost control. It stops damage after it happens.”

Real governance, in his view, means real-time guardrails that check every agent action against its authorisation boundary before it runs, “not a log you read after the breach. The analogy is simple: a fire extinguisher is not a substitute for building a fireproof building. Stop designing for the emergency. Design so the emergency can't happen.”

Renting intelligence versus owning it

On the choice between closed-model APIs and open-weight models, Gurtu reduces the question to one he says every buyer should put to their vendor. “When we stop paying you, what do we keep? The answer for every closed-model API is the same: nothing. The outputs disappear. The intelligence disappears. Your competitor paid the same price and got the same outputs.”

Fine-tuning an open-weight model on proprietary data, he argues, produces something a competitor cannot buy: “a reasoning system that thinks like your organization, runs in your infrastructure, and compounds institutional intelligence every time it executes. That's not a data privacy argument. That's an EBITDA argument. One builds an asset. The other funds a cost center with no exit.”

The patent expires, the machine compounds

If intellectual property is losing its force as a moat, Gurtu has a clear order of priorities for CFOs and boards. “First, the data: your operational history, your decision library, your domain knowledge. Nobody can subpoena it into a competitor's training run.”

Second comes the architecture, how a company has fine-tuned its models and structured its agentic workflows, which he calls “a genuine trade secret in the practical sense” because “replicating it requires replicating years of operational execution”. Third is the governance infrastructure: “the audit trails, the guardrails, the explainability framework that makes your AI trustworthy enough to act autonomously. The patent expires. The machine compounds. Boards should be funding the machine.”

Measure closure rate, not adoption

For a Fortune 500 technology leader starting today, Gurtu's advice is to change the metric before anything else. “Stop measuring AI by adoption and start measuring it by closure rate.” By that he means the share of workflows entering an AI system that the system completes without a human touching them between steps.

“Most enterprises are running AI that advises and humans that execute. That's not transformation, that's a faster briefing. The moment you measure closure rate, you immediately see where you have genuine agents and where you have LLM prompts in expensive suits. Fix the metric first. The architecture conversation follows automatically.”