A lot of the conversation around agentic AI is focused on autonomy: How much work can we hand off, how quickly can agents make decisions, and how far can automation go? Where should humans remain involved? ASAPP Chief Architect Nirmal Mukhi has a strong perspective about all of these questions.

Nirmal has spent the last several years helping enterprises deploy AI into customer-facing environments, and one theme keeps surfacing. The challenge isn’t getting AI to make decisions. It’s deciding which decisions carry enough financial, operational, or customer risk that human judgment still matters.

Nirmal’s perspective? AI should handle the work, while humans handle the consequences.

M.R. Rangaswami: Where do you think human judgment still creates the most value?

Nirmal Mukhi: I don’t actually start by asking where humans need to stay involved. I start by asking what happens if this decision turns out to be wrong.

That changes the conversation pretty quickly.

If we’re talking about summarizing a customer conversation or recommending the next best action, that’s usually an easy discussion. The cost of getting it wrong is relatively low, and AI is incredibly good at those kinds of repetitive tasks. Most people would rather spend their time solving a customer’s problem than digging through notes or copying information between systems.

Where I see people slow down is when a decision starts affecting a person’s finances, their relationship with a company, or the reputation of the business itself.

Imagine explaining to a customer that an AI denied their refund after months of back-and-forth. Or explaining to an auditor why an automated system approved an exception nobody ever looked at. Those situations feel different because they are different. At that point, you’re not just talking about efficiency anymore. You’re talking about ownership. 

M.R.: What’s the biggest misconception you’re seeing as organizations rush to deploy AI agents?

Nirmal: I’ve had a version of the same conversation with a lot of enterprise leaders over the past year. It usually starts with, “Our AI can do X now.” Maybe it’s resolving customer requests, making recommendations, or handling more of a workflow without someone stepping in.

That’s the problem – organizations are deploying AI as an alternative to human effort. That is appropriate but also limited.

Most business processes are complex and you cannot entirely subtract humans from them. So any AI deployment, instead of thinking “how do automate more” needs to instead start at “how do I redesign this so that humans + AI can do this better”. That is where you use the strengths of each – AI to scale efficiently, human to apply judgement where appropriate – and use the best of each.

M.R.: If you could leave enterprise leaders with one idea as they expand their use of AI, what would it be?

Nirmal: I’d probably tell them to stop treating autonomy like the finish line.

It’s easy to assume that every new capability should lead to another layer of automation. Sometimes that’s exactly the right answer. Sometimes it isn’t.

The companies I think are getting this right aren’t necessarily the ones automating the most work. They’re the ones thinking about the big picture – which centers on trust and efficiency – you can have your cake and eat it too. 

“Trust AND efficiency” sounds less exciting than talking about fully autonomous agents, but in practice it’s what allows organizations to move faster without creating problems they’ll have to untangle later.

Five years from now, I don’t think anyone is going to remember which company automated 82% of a workflow instead of 78%. They’ll remember which companies built systems people actually trusted.

To me, that’s the bigger challenge. AI is going to keep getting more capable. The question is whether our judgment about where to use it matures just as quickly.

M.R. Rangaswami is the Co-Founder of Sandhill.com