AI Governance in the Age of Agentic AI: Building Safe, Responsible Enterprise AI | Jamie Campbell

Sanjay Puri (0:20)

Thank you. The platform decides to call a tool, maybe pull a policy record or trigger a workflow. What's actually standing between that decision and it happening?

Jamie Campbell (0:32)

Agent, just doing that. And I think this is the kind of control harness that we have within our platform, where the noted behavior of that agent, you know, we are actually saying this is the intended behavior. Is the agent actually accessing that data? How is it accessing it?

Sanjay Puri (0:58)

I'm your host, Sanjay Puri. Today, I'm joined by Jamie Campbell, Head of AI at Manulife Canada.

Jamie has spent his career rebuilding operating models under pressure, R&D transformation at Roche, turning around a private equity-backed healthcare consultancy at Mordic Healthcare Group, and now leading AI for Canada's largest insurer, Manulife's Automated Underwriting Engine Mod, with deciding the majority of eligible life insurance applications in minutes. Jamie, welcome to the Chief AI Officer Podcast.

Jamie Campbell (1:30)

Thanks, Sanjay. Great to be with you.

Sanjay Puri (1:33)

Jamie, we have a global audience largely made up of your peers, chief AI officers, CIOs, CTOs, enterprise leaders, and entrepreneurs within AI. Just very briefly, what is Manulife?

Jamie Campbell (1:47)

Thanks, Sanjay. Yes, Manulife Global Organization, headquartered in Toronto, Canada.

Origins were very much in the life insurance business, but global wealth and asset management insurance company, operating essentially internationally.

Sanjay Puri (2:08)

So, Jamie, you've now rebuilt operating models under very different regulatory regimes. You know, Pharma, R&D at Roche, which is, you know, you talk about heavily regulated FDA, healthcare consulting at the Nordic Healthcare Group, and now insurance underwriting. Is there one piece of commonality that kind of moves over from one, you know, regulatory framework to a framework to another?

Jamie Campbell (2:40)

Yeah, it's a great question, Sanjay. And oftentimes, what I hear is that, oh, they're very uniquely different.

And I think what you're calling out is that there's a lot of commonalities in a highly regulated environment. I think, you know, the best way I like to put it is that you have to be very mindful and conscientious to build those regulations into your designs and your models. For those who are out there who are, you know, just wanting to build AI for the sake of AI in a highly regulated environment, you have to be very kind of structured and aware externally of what, you know, those safeguards and regulations are. And I think that's the most important part in the commonality between those experiences.

Sanjay Puri (3:34)

Jamie, you described governance as the operating system for AI at Manulife, you know, built in from day one. Has there been a time it didn't work or a guardrail that failed or produced a sense of safety? Tell your peers a little bit.

Jamie Campbell (3:51)

The way, Sanjay, I like to view governance, if we weren't having issues in detecting things, that means the governance system isn't actually working. So I like to view a governance system as ones where you are putting the appropriate safeguards in to proactively identify the issues and the challenges in the models that we're developing. And as you said, it's all over the news right now in terms of issues. And I think that's an effective governance model. If it allows you to put the appropriate safeguards, gain and identify the right evidence to be able to make meaningful decisions and mitigate risk, that in my mind is an effective governance model.

Sanjay Puri (4:43)

Jamie, when you restrict the model to ignore everything you've learned, except this defined document set, how do you actually verify if it's complying versus just trusting the prompt in some shape or form?

Jamie Campbell (4:57)

I think that the way I respond to that is, do I ever trust a prompt? The way I view it, and to my colleagues out there, I would say and propose that a prompt is just an instruction, and it is in our responsibility, as humans developing these models, and is to validate what is the data set? What is the data output that we're looking and expecting to have?

I think that's really important is that for those out there, you really have to understand what data is being accessed, how that data is being used and the outputs, because that then gives meaning and structure to responsible use of the models, and it kind of puts the safeguard around the prompt, if you will, so you're not just trusting the prompt.

Sanjay Puri (5:58)

Mod is currently deciding the majority of eligible life insurance applications that come to Manulife.

Jamie Campbell (6:06)

No, I think for those out there, Mod, that's our Manulife's automated underwriting decision engine, Sanjay, and it's used to auto-approve straightforward individual life insurance applications.