No Joke: A Big Consulting Firm Thinks It Can Prevent the Inevitable
April 1, 2026
Another dinobaby post. No AI unless it is an image. This dinobaby is not Grandma Moses, just Grandpa Arnold.
I read a darned amusing post title “How Big Four Firm KPMG Is Protecting Itself from AI Agents Going Rogue.” The idea is that an outfit that hires mostly MBA and CPA type financial types who are, by definition, the best of the best will take steps today against tomorrow’s problems. That sounds like a very pragmatic approach under certain conditions. I will come back to the “certain conditions” idea is a moment.
Thanks, Venice.ai. I liked the message that the image was in violation of your guidelines. Good enough.
But first, let’s look at this passage from the cited write up (which I assume was not written by AI like some consulting firms’, law firms’, and government agencies’ content is. Consider this statement:
But as these autonomous systems become embedded in workflows, so too does a sense of unease about their unpredictability and the risks they pose to businesses.
Sane enough. Autonomous smart software, vibe coded agentic workflows, and humans who may not know exactly how those gradient descents work should engender unease. I would choose a different word; for example, what happens when smart software gets a target of a kinetic weapon wrong? Answer: Dead kids. But that’s not going to worry the best of the best in a big time consulting firm.
KPMG has created a multifaceted framework to protect against worst-case scenarios for both clients and its own employees.
The source of this forward leaning statement is Sam Gloede, Trusted AI leader at KPMG. Is Sam “trusted” or is the AI? It must be Sam because. Sam has come up with a method to make sure that no KPMG agentic system goes off the rails. Stated another way, KPMG does not want the business equivalent of dead kids.
What’s the method? That’s easy an MBA type “set of controls.” Sam points out what I call a woulda coulda shoulda approach:
Agents should only interact with the systems and data they strictly need, limiting the potential impact of errors.
In my lingo, this is similar to the fellow Louis Slotin who took controls seriously. Then he dropped his screwdriver. Plutonium was little understood in the mid 1940s. Is that something that one might say about agentic software today. But KPMG is definitely not Louis Slotin. KPMG logs. It uses red teams. It monitors with humans and smart software. Los Alamos National Lab in 1946 did the same thing. Then zap.
Sam allegedly said:
“It’s not about scrutinizing people’s behaviors for performance and alignment,” said Gloede. “It’s the ability to just always be monitoring your technology ecosystem.”
I like that “technology ecosystem” for two reasons: [a] KPMG acknowledges that the agents are operating independently just like Louis Slotin and [b] it includes a categorical affirmative; specifically, “always.” Yep, always. That means never, ever, no, and nope. I am not sure I am a believer in categorical affirmatives when probabilistic systems and MBAs / CPAs are on duty and making checklists and designing procedures.
But there are several other issues that this article caused me to consider:
- This is a content marketing write up designed to reassure existing clients and make prospects believe that KPMG has its circus organized and on the AI train
- KPMG apparently has control over the AI models it uses. I am confident that KPMG believes this to be true, but I am not sure than the memo has been received at the frontier AI firms and if received, even opened.
- Smart software can demonstrate interesting characteristics; for example, hallucinations, mistakes, and an inability to identify appropriate context for certain prompts/actions. KPMG obviously is confident that its MBA/CPAs have the system and method to cope with these known behaviors. But what about known unknows of upgraded smart software.
KPMG’s confidence strikes me like the race trach tout in Jack Benny’s radio program. The fellow sounded quite confident in his predictions. How did those work out? Well, the race tract tout was not a professional gambler. He did game shows. That confidence for predicting the winner of a horse race was less robust than Sheldon Leonard’s character.
I want to point out that this allegedly happened:
An experimental AI bot designed as a virtual assistant escaped its closed test server and began unauthorized cryptocurrency mining on Alibaba servers.
KPMG asserts that it has the matter encapsulated in its procedures, checklists, and processes. I assume that means the Chinese wizards flubbed the bounce pass. Whom do I believe? Gee, I believe everything I read on the Internet whether written by a marketer, a “real” news outfit, or an AI system. I really believe blue chip consulting firms. Don’t you?
Net net: Some people want AI to work but beneath the surface is a real concern that the smart software can cause problems, big problems. Like the kids.
Stephen E Arnold, April 1, 2026
Comments
One Response to “No Joke: A Big Consulting Firm Thinks It Can Prevent the Inevitable”
Got something to say?
-
- Subscribe to Beyond Search
Feature archive
News archive
-
?????? Haval – ?????? ? ??? ?? ??????? ?????? ????????????. ??????? ????? ??????? ????? ?? ????? ?????? ??????????? ????? ????? ?????? ? ???!
[url=https://havalpromsk11.ru]?????? ??????????? ?????[/url]
????? ?????? – [url=https://havalpromsk11.ru/]https://www.havalpromsk11.ru/[/url]