Why Trust (Not Intelligence) Will Decide the Winners of Enterprise AI

As enterprises move AI agents from experiments into production, the biggest challenge may not be model intelligence, but whether organizations can trust those systems with real authority. In this interview, Descope co-founder Rishi Bhargava explains why identity, permissioning, and accountability will determine which companies can safely scale agentic AI, and which will be forced to keep it confined to the sandbox.

Everyone seems to be talking about AI agents right now. Why has identity suddenly taken over the conversation?

Every frontier model is now good enough for the workflows most companies actually want to automate. But the gap between the first and third-best models matters far less than the gap between an agent you trust and one you don’t.These are the questions I hear from teams trying to take agents live: “If this agent updates a record, opens a ticket, or moves money, can I reliably say who authorized it, what it was allowed to do, and what it actually did?” Those all trust questions, and trust questions are fundamentally identity questions. 

The intelligence of the model doesn’t change the answers to a single one of them.

Traditional identity systems have worked for decades. Why don't they work for AI agents?

Every identity system we’ve used for the last two decades puts a credential into one of two buckets: human and machine.

Human credentials assume a person is present. Machine credentials assume the software is predictable. Agents don’t fit in either bucket.

They act without a person present, and they’re non-deterministic. They decide at runtime which systems to touch. If you give them user credentials, they inherit everything that particular user can do. Machine credentials like service accounts and API keys aren’t granular or flexible enough for agentic workloads, and you can’t tell which user it was working for.

Both fail with agentic workloads, just in different directions.

What's the biggest security mistake you're seeing organizations make?

Treating access as something you grant once and never revisit. Organizations are moving fast and assuming they'll clean things up later. They rarely do.

Traditional machine workloads accumulate identity debt slowly, but AI agents aren’t slow. A team can spin up a fleet of dozens in an afternoon, with each agent touching a different handful of systems, and every one of them gets permissions someone granted quickly and will never review.

Most organizations can’t produce a list of all their agents and their permissions, and you can’t govern what you can’t see. AI didn't invent that problem, but it’s making it happen faster, 

What do most people get wrong when they think about AI governance?

People should stop equating governance with policies, frameworks, or compliance checklists. Those things matter, but by themselves, they don't actually stop anything.

Make it really practical. If your security team can't answer, "Who asked this agent to do this?" within a couple of minutes, you've probably got a governance problem.

And then I'd ask a second question: What else could that agent have done? That's usually where things get uncomfortable.

Good governance isn't just knowing that an action happened. It's knowing who delegated it, what permissions were used, whether those permissions were appropriate, and being able to prove all of that after the fact. 

Let's say I'm a CIO who's about to roll out AI agents across my company. What's the first piece of advice you'd give me?

Start by asking exactly what this agent needs to do. Not what it might need to do someday, but what this specific task requires today.

Scope everything from that answer. The agent gets its own first-class identity, not a borrowed one. It gets the narrowest possible permissions that let it finish the task, and those permissions die when the task is done. Every action it takes carries a record of what was requested, what was granted, and what actually happened.

The instinct is to grant broad access now and tighten it later because that’s fast. Resist it. Enterprises spent the last two decades accumulating identity debt from that anti-pattern, one reasonable-sounding exception at a time. Agents will run up the same debt in months, not years.

Where do you think enterprise AI is headed over the next few years?

Models are increasingly becoming infrastructure. Enterprises will swap them the way they swap around cloud regions: for cost reasons, performance gains, compliance, whatever their team decides. That’s a healthy practice, but it only works if trust doesn’t live inside the model.

The organizations that get this right will build their controls on a model-agnostic agentic identity layer, where it doesn’t matter whether Claude Fable or GPT-5.5 Instant is doing the reasoning. 

What matters is that every action is authorized, scoped, and attributable. Then a more performant model is a drop-in instead of a fresh security audit. 

Eventually, everyone’s going to have smart AI. The real gap is between companies that can safely hand agents real work in production, and those that have to keep it caged in a sandbox.