Aug 6, 2026 · 9 min · 4 segments
Would you trust AI at only 80% accuracy? What about for your enterprise? KPMG Chief Digital Officer Kelle Fontenot explains why AI success cannot be measured by headcount reduction alone - and how…
Kelle FontenotGuest
Chris BrandtHost
[gentle music] We're gonna get an 80% success rate on some of these agents, but we gotta figure out how we mitigate that 20% that weren't successful once we roll them into production.

I love what you were saying about, like, changing your success criteria for the new world, because now we're entering into non-deterministic systems, so that, like, no mistakes criteria really doesn't align well to what's happening now in technology.

I think that's a really insightful piece that, you know, probably more companies need to start rethinking the way that they're determining success.

The hard part is there's so much of what people will use these capabilities for where you don't wanna have a mistake, and so then you've got to ensure that everything behind the agents is working appropriately, right? That's why some part of having the data right really matters and, and having, like, not just rows and columns data, but truthfully, like, the content.

So if I type in a prompt that says, "What is the KPMG PTO policy?" We don't have a PTO policy anymore in the US.

Well, so guess what? If I just type, if I type it in wrong, I get last year's policy.

So then you've got to put a, a huge amount of effort around what is this year's policy.

And, and so that whole knowledge management around what you have out there in, like, legacy SharePoint folders or whatever they are, you've got to modernize all that.

And so it takes a lot of work to even be at 80%, 'cause you don't want somebody thinking that they have two, three weeks when they actually have flexible.

That's just one simple example, because if you ask the wrong question, you get an inappropriate answer.

And I think that's one of the big problems with AI overall is, is the data pruning problem.

And because it's so new, we haven't had to do the level of pruning that we will in, you know, two, three years.

And so I think this is a building and interesting problem that's, that, that we've got to turn our attention to pretty quickly.

[gentle music] We're gonna get an 80% success rate on some of these agents, but we gotta figure out how we mitigate that 20% that weren't successful once we roll them into production.

I love what you were saying about, like, changing your success criteria for the new world, because now we're entering into non-deterministic systems, so that, like, no mistakes criteria really doesn't align well to what's happening now in technology.

I think that's a really insightful piece that, you know, probably more companies need to start rethinking the way that they're determining success.

The hard part is there's so much of what people will use these capabilities for where you don't wanna have a mistake, and so then you've got to ensure that everything behind the agents is working appropriately, right? That's why some part of having the data right really matters and, and having, like, not just rows and columns data, but truthfully, like, the content.

So if I type in a prompt that says, "What is the KPMG PTO policy?" We don't have a PTO policy anymore in the US.

Well, so guess what? If I just type, if I type it in wrong, I get last year's policy.

So then you've got to put a, a huge amount of effort around what is this year's policy.

And, and so that whole knowledge management around what you have out there in, like, legacy SharePoint folders or whatever they are, you've got to modernize all that.

And so it takes a lot of work to even be at 80%, 'cause you don't want somebody thinking that they have two, three weeks when they actually have flexible.

That's just one simple example, because if you ask the wrong question, you get an inappropriate answer.

And I think that's one of the big problems with AI overall is, is the data pruning problem.

And because it's so new, we haven't had to do the level of pruning that we will in, you know, two, three years.

And so I think this is a building and interesting problem that's, that, that we've got to turn our attention to pretty quickly.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.