Skip to content
Fast Company Executive Board

The Fast Company Executive Board is a private, fee-based network of influential leaders, experts, executives, and entrepreneurs who share their insights with our audience.

What are we really prepared to let AI do?

Four questions to help decide how and when to give it authority.

[Photo: Getty Images]

Profile photo of Vinit Sahni

Founder · governr.ai

For the last few years, the question around artificial intelligence has been simple: What can it do?

Every few months the answer becomes more impressive. AI can code, analyze, research, reason, use tools, and increasingly string those capabilities together to complete real work. But as agents become more capable, the question that matters most for business leaders is changing. It is no longer what AI can do. It is what we are really prepared to let it do.

Robert F. Smith, chairman and CEO of Vista, recently described AI as moving from a technology investment to an operating discipline. I think he is right. Once AI becomes part of the operating fabric of a business, the questions change character entirely. Letting AI analyze a portfolio is one thing. Letting it see the live portfolio is another. Letting it change something, communicate externally, or act without asking first is quite different.

The technology can already do much of this. The harder question is whether the organization is prepared to let it.

THE NEW CONSTRAINT IS PERMISSION

Imagine an agent that could save a team hundreds of hours a month. Technically it works. But to deliver its full value it needs access to customer information, internal knowledge, perhaps a financial system, and perhaps the ability to take action. Who says yes to those requests? And what exactly are they saying yes to?

What happens if six weeks later somebody changes the underlying model, or connects another data source, or gives the agent code-write access? Is the original approval still valid?

At governr.ai, we saw this ourselves earlier this year. At one financial institution, we found an agent still running with live system credentials months after the person who built it had left. It was doing useful work. Nobody had approved its continued operation, because nobody knew it was still operating.

Companies already have substantial infrastructure around AI: security checks, identity confirmation, evaluation protocols, and other guardrail systems. They answer important questions about vulnerability, access, and behavior. All of those questions matter. But there is a decision sitting above those systems: Given everything we know, should this AI system be allowed to operate in this way? And a harder second question: Is that still true today?

AI NEEDS A CURRENT AUTHORITY TO OPERATE

Companies are used to thinking about approval as an event. A project is reviewed. Someone signs off on it. It goes into production. That model breaks down when the thing being approved changes continuously. Someone may add a dataset, a connector, or a new permission. An approval that was entirely sensible three months ago tells you very little about whether that same AI should still be doing the same job now.

For consequential AI systems, companies will need something closer to a current authority to operate. Not an approval buried in a spreadsheet from six months ago, but a current answer: Yes, this AI is allowed to do this work, with this data, these permissions, and these controls, today.

That sounds like more control. Counterintuitively, it produces more adoption, not less. Companies that gain the most from AI will not be the ones with the fewest controls. They will be the ones that understand what their AI is doing, and hold the boundaries around it, well enough to be comfortable giving machines considerably more responsibility.

FOUR QUESTIONS TO ASK FIRST

Rather than beginning with a list of prohibited use cases, executives could start with four questions for every consequential AI deployment.

1. What valuable work do we want AI to take over? Start with the economic outcome, not the technology. Where would greater autonomy materially improve productivity, capacity, speed, or service?

2. What exactly are we prepared to authorize it to do? Reading information, generating a recommendation, communicating externally, and taking an irreversible action are very different things. Make the boundary explicit.

3. What conditions make us comfortable granting that authority? Think about which data the AI system can use, which systems it can touch, how much autonomy it has, and where human approval still sits.

4. What change would cause us to reconsider? This is the question most organizations skip but shouldn’t. A new model, a new dataset, an additional permission, or a new external connection can alter the risk completely. Approval should carry the conditions under which it is automatically revisited.

These are operating questions, not technology questions. And, increasingly, they are questions for the board of directors. That’s because AI is moving from helping people do the work to being entrusted with the work itself.

The competitive question is no longer how capable your AI can become. It is how much valuable work you can confidently delegate to it without losing control.

Vinit Sahni is the founder of governr.ai.