How leaders can scale AI by trading control for trust (Q&A)

There’s a quiet standoff happening in enterprise technology right now. On one side, we have artificial intelligence models that are staggeringly capable. On the other side, organizations are hesitating to deploy AI at scale. For years, the tech industry assumed that building smarter models was the only way to achieve true AI automation. Today, the reality is different: the intelligence is already here, but organizations are struggling to confidently put it into production.

The hesitation makes sense. Trust is built in stages, and transitioning responsibilities to agents isn't about a leap of faith. To reach that next level of cloud scale, leaders need a new framework for defining trust and scope.

To understand more, we sat down with Michael Gerstenhaber, VP of Product Management for Gemini Enterprise, to discuss why the future of AI is about defining safe boundaries. What follows is an edited transcript of our conversation.

Q: It seems like the industry is always waiting for the next major model update. Is raw intelligence still the main hurdle for enterprise AI?

Michael Gerstenhaber: Actually, intelligence is no longer the boundary condition. The technology is already capable of doing the work; the issue is that people don’t feel safe putting it into production yet. We tend to think of AI adoption as a binary switch: either it works or it doesn't. But it’s actually about trust. We haven't yet credentialed non-human agents to do high-stakes work without us watching. The intelligence is there, but the "scope" of trust is still very small. The tech is capable of doing automation, people just don’t feel safe putting it into production. I want to unlock the ability to use the Cloud at scale.

Q: Why are certain areas, like software engineering, adopting AI so much faster than others?

Michael Gerstenhaber: It all comes down to finding patterns and managing risk. Software engineering moved incredibly quickly because AI fits nicely into its existing lifecycle. In other industries, it takes time to put these patterns into production. Engineers need to figure out what a non-risky scope looks like in their specific domain: a boundary where they are willing to audit the AI's work instead of controlling every step. Once you do that, it just becomes a standard software problem.

Q: What does oversight look like in a world where we want AI to operate at scale?

Michael Gerstenhaber: You really only have two options: keep a human in the loop, or define a scope where you completely trust the agent. We already accept this kind of trust in other areas of tech. For example, logging into your computer is a common pattern that doesn't require manual oversight every single time you put your password in. We need to reach that same level of trust with AI agents.

I’m obsessed with the concept of trust and scope. The technology is highly capable of automating any work done on a computer, but people just don't feel safe putting it into production yet. Right now, agents operate in a mode where you chat with them and wait for a response. I want to decouple those tasks and help people trust AI enough to take them out of the loop entirely, which will unlock the ability to use the cloud at massive, infinite scale

Q: Is this shift specific to coding, or will it apply to other industries?

Michael Gerstenhaber: It applies everywhere, but it requires domain expertise to define the scope.

For example, a scientist at a pharma company might not be a coder, but they can use agents to run statistical analysis for compound discovery. A lawyer using Thomson Reuters’ CoCounsel2.0, their professional-grade gen AI assistant, can delegate the analysis of a contract. In all these cases, the human is still the expert. You have to be an expert to know if the agent did the job well. But instead of doing the grunt work, you are defining the scope and verifying the results.

I’m seeing this everywhere. Estee Lauder and Jo Malone introduced their AI Scent Advisor, which leverages Gemini Enterprise Agent Platform to interpret natural language responses and map them to Jo Malone London’s olfactory data attributes to generate bespoke, expert-informed fragrance recommendations. And Honeywell uses Gemini Enterprise Agent Platform to power its Smart Shopping Platform, which helps shoppers easily locate desired products, compare similar items and quickly find relevant substitutions when products are unavailable, making in-store shopping more efficient and enjoyable.

To learn more about the latest in Agent Platform, check out our latest recap.

Q: What’s your remit in making that happen?

Michael Gerstenhaber: The intelligence is already here. My job is to be a good steward of technology worldwide, help people access these novel patterns, and solve all the other complex enterprise concerns. That means helping organizations figure out how to assess and delegate trust, manage governance, and maintain context awareness. It all boils down to scope and trust.

Q: What’s the payoff of getting this scope and trust equation right?

Michael Gerstenhaber: Infinite scale. If you can find a scope where you trust the agent enough to take the human out of the loop, you can scale infinitely in the cloud. That’s when the economics of the cloud really change. We move from one person solving one problem to distributing intelligence to solve thousands of problems simultaneously.

Q: Do you have parting advice for leaders who are hesitant to let go of control?

Michael Gerstenhaber: Don't wait for a model that never makes a mistake, because that’s not how software works, and it’s not how humans work. Instead, focus on defining the boundary. Ask yourself: What is a non-risky scope in my business that I can hand over to an agent today? Start small, audit the results, and as you gain confidence, widen the aperture. That’s how you prepare your organization for an agentic future.