Best of LinkedIn: AI & Agentic Systems CW 41/ 42

00:00:00: This episode is provided by Thomas Allgaier and Frennus, based on the most relevant LinkedIn posts about AI and agentic systems from CW-forty one in forty two.

00:00:10: Frenness supports ICT and tech providers with a ecosystem strategy by delivering independent vendor assessment build versus buy analysis an ecosystem intelligence that prevents expensive mistakes and positions the provider's competitively.

00:00:24: you can find more info in the description.

00:00:26: all right let's get ready to do it.

00:00:28: yeah Welcome to the deep dive.

00:00:30: Our mission today is just, you know cut through a lot of the hype.

00:00:33: We're looking at the most relevant trends across the professional landscape To figure out what's actually happening with AI and agentic systems right now.

00:00:41: Right because we are definitely moving out of that era Of pure unfiltered experimentation.

00:00:47: We've entered this new phase in an area of accountability really

00:00:51: Exactly.

00:00:52: And the overarching narrative we're seeing from professionals on the ground is that, The main constraints are getting actual value out of AI.

00:01:00: They aren't tech itself anymore!

00:01:02: No not at all...the bottlenecks are structural.

00:01:04: We talking broken operating models massive governance cues and just quality human judgement.

00:01:10: Yeah if you look what people actually deploying these systems or saying technology IS ready It's organizations totally lagging behind

00:01:19: Which a perfect place to start looking right at that enterprise level, where the rubber meets the road and why so many companies are just hitting a complete wall trying to scale this stuff.

00:01:28: Yeah!

00:01:28: So Nandan Mulakara brought up point I think is massive warning light for anyone in tech or digital transformation.

00:01:35: now he pointed out enterprises are risking total repeat of scaling failures we saw with RPA.

00:01:42: Oh robotic process automation.

00:01:44: yeah

00:01:45: That's brilliant historical parallel

00:01:48: Because back then, RPA just made broken processes move faster.

00:01:52: And now everyone is basically mounting this massive AI gun onto those exact same outdated operating models.

00:01:58: They're essentially just automating the waste.

00:02:01: But I do want to pushback on that a little bit like isn't speed kind of the entire point?

00:02:07: Of AI?

00:02:08: if you have a broken shopping cart but you strap A Ferrari engine into it Isn't that still better than walking?

00:02:13: well i mean It'll get you down the aisle faster, sure.

00:02:16: But you're going to crash into all the shelves right?

00:02:18: You're gonna tear this store apart...you just cannot automate a bad workflow and expect good results.

00:02:24: Asad Ansari highlighted this perfectly when he was looking at the UK civil service.

00:02:29: Oh!

00:02:29: The massive savings promises.

00:02:30: right

00:02:31: exactly it's this promise of AI saving something like thirty six billion pounds a year.

00:02:37: But his reality check was that those numbers are completely unrealistic if you don't tear down the foundational broken processes first.

00:02:45: Because AI is just a pure multiplier?

00:02:48: Yes, feed it a broken system and doesn't fix it – the chaos just multiplies!

00:02:54: So Atena Rahani actually brought great framework to this operating model problem.

00:02:59: She pointed out scaling agentic AI where the system takes action on its own requires fundamental shift in the operating itself.

00:03:07: Right, because it's no longer just a question of whether the tech can act.

00:03:10: Exactly!

00:03:11: It is about deciding which specific outcomes get.

00:03:13: what she called bounded autonomy

00:03:16: Like

00:03:16: where does an agent have free reign?

00:03:18: And where do human judgments still need to step in and shape their path?

00:03:21: If you skip that restructuring part things gets messy fast.

00:03:25: Gabriel Millian actually broke down some hard data on this from Airstreet Capital state-of-AI report.

00:03:30: What did they find?

00:03:32: They looked at the OS world.

00:03:33: two point O test Benchmark for real computer work, like navigating web browsers using desktop apps.

00:03:41: The messy everyday enterprise stuff?

00:03:44: Exactly and a top AI model Claude Opus V. it got partial credit for seventy seven point seven percent of the steps

00:03:53: which sounds pretty incredible on paper.

00:03:55: almost eighty percent Of the steps right.

00:03:57: that sounds Like A massive productivity boost.

00:04:00: It does.

00:04:00: until you look at the actual finish rate It only finished forty-four point three percent of the actual tasks.

00:04:07: Oh wow!

00:04:08: Less than half,

00:04:09: right?

00:04:10: Because a high score on these steps isn't the finish job.

00:04:13: in an enterprise partial credit is basically zero credit.

00:04:17: Yeah I mean if an agent drafts a perfect contract but then gets confused by the final eSignature API The contract doesn't go out.

00:04:25: A single missed approval leaves the whole thing undone.

00:04:28: Exactly So.

00:04:29: the technology's the easy part

00:04:33: which perfectly ties into ERA Agusabt-Sien's point.

00:04:36: He made this argument that when AI actually does save time on a project, you shouldn't just convert those savings in to cheaper bill for the client.

00:04:44: Right!

00:04:44: The instinct is always to pocket the savings.

00:04:47: Yeah but he was saying that savetime has to be reinvested into process.

00:04:51: You need it quality checks business validation and change management.

00:04:55: So what all of us mean as leader?

00:04:57: It means we have to redesign workflow around human oversight not machines output

00:05:02: Absolutely.

00:05:04: And if fixing the process is step one, then Step two... ...is figuring out how you govern all these new workflows without just grinding innovation to a total halt.

00:05:13: Oh!

00:05:13: The governance nightmare?

00:05:14: Yeah.

00:05:15: Joseph Abraham highlighted what he calls the Q problem.

00:05:19: When you go into an enterprise and ask the AI team what's slowing them down they usually blame model hallucinating or bad data.

00:05:26: But it's not that data, isn't it?

00:05:28: No It's the queue.

00:05:29: People are submitting use cases by email and they just sit there for eighteen days.

00:05:35: And a basic internal chatbot gets reviewed with the exact same depth as a critical financial model.

00:05:40: It's like installing speed bumps on a major highway, people don't stop driving They just detour through residential neighborhoods where it is way more dangerous.

00:05:47: Exactly

00:05:48: Slow governance doesn't reduce risk

00:05:53: literally how shadow AI happens.

00:05:55: Employees have deadlines, so they just quietly use unapproved tools.

00:05:59: and by the way if you want to keep navigating these enterprise detours with us make sure your subscribed to The Deep Dive So You Don't Miss Future Editions.

00:06:06: We Cover A Lot Of These Roadblocks.

00:06:09: And on the regulatory side of that roadblock Christian Bean analyzed a new EU AI Act standard EN- one eight two eight six.

00:06:18: Okay That sounds extremely dense

00:06:19: It is but his take away really practical.

00:06:23: ESA's compliance shouldn't mean building entirely new, parallel structures in your company.

00:06:28: So no isolated AI ethics board that meets twice a year?

00:06:31: Right!

00:06:32: You need to perform gap assessments and integrate the AI requirements into quality management systems you already have.

00:06:38: That makes systemic sense.

00:06:40: But there is human side of this governance too which Seagrid Burj van Royen brought up regarding clinical AI in healthcare.

00:06:46: This is a fascinating dynamic.

00:06:48: Yeah she argues The best governance document in the world won't make a clinician actually trust an AI tool.

00:06:55: Trust breaks into real-world when a doctor is exhausted and under time pressure.

00:06:59: And

00:06:59: then, most dangerous part isn't even when they reject the AI?

00:07:02: No!

00:07:03: It's when that AI is right MOST of the time because THEN clinicians stop actively agreeing... ...and just start clicking through.

00:07:10: Active trust degrades to what she calls passive consent

00:07:14: Which is terrifying…in clinical setting

00:07:16: Completely.

00:07:17: Tools have to be designed to explicitly communicate their uncertainty, to force the human to wake up and pay attention.

00:07:24: And that danger of passive consent gets infinitely riskier when we transition from AI just gives advice... ...to a GENTIC.

00:07:31: AI actually takes action on our behalf.

00:07:33: Here's where it really get interesting.

00:07:36: Renee Kay highlighted this recent incident in UK AI Security Institute.

00:07:40: It proves that agentic risks are way beyond just, you know, prompt injection.

00:07:45: We aren't

00:07:45: just tricking chatbots to write bad poems anymore

00:07:48: Exactly!

00:07:49: we now have agents capable of unsanctioned supply chain compromise.

00:07:53: they can spin up servers execute transactions...we absolutely need a defense in depth layered security approach for this.

00:08:00: Thomas Sowering from SAP put this into perspective perfectly.

00:08:03: He said, AI hasn't really changed what we defend against in cybersecurity but it has completely changed.

00:08:09: how fast that happens?

00:08:10: Because the time to impact is shrinking?

00:08:12: Exactly!

00:08:12: It's compressed for weeks down literally minutes.

00:08:16: Having a human respond at human speed no longer viable as first line of defense.

00:08:20: We need machine-speed defenses.

00:08:22: Wait, but if the machine is defending at machine speed and the attacker's attacking at machine-speed where does human even fit in?

00:08:29: Are we literally just bystanders watching algorithms

00:08:32: fight?".

00:08:32: It's a valid concern.

00:08:33: But we are the ones who control boundaries.

00:08:36: Moshkin Lefebvre provided framework for this.

00:08:39: Agent autonomy isn't binary switch... Right!

00:08:44: It should scale based on task risk.

00:08:46: You have to design sandboxes asking two things What consequences of agent being

00:08:51: wrong?

00:08:52: and is the action reversible?

00:08:53: Ah, reversibility.

00:08:55: Drafting an email as reversible wiring fifty grand it's not

00:08:58: Exactly.

00:08:59: But defining that sandboxes getting weird because our legacy security systems can't even recognize good agents.

00:09:06: Nikhil Kassidy shared this highly specific kind of hilarious example from The FinTech Space

00:09:11: About fraud models.

00:09:12: Yeah Right now Fraud detection models are misclassifying legitimate AI purchasing agents at suspicious spots

00:09:19: Because they're too fast.

00:09:20: Right.

00:09:21: The fraud models learned what suspicious means based on human typing rhythms like hesitation, the way a mouse jiggles.

00:09:27: A good AI agent has no hesitation.

00:09:29: It looks exactly Like a malicious script.

00:09:32: So this system is blocking the exact efficiency it was meant to protect

00:09:35: Exactly And Michael Westerwell brought up a wild solution To this authorization problem.

00:09:41: World ID just launched iris verified identity proofs For AI agents.

00:09:46: Wait!

00:09:46: Like scanning your eyeball for the AI.

00:09:48: Literally Yes You scan your eyeball to verify you're human, and hand that cryptographic proof of human...to the AI agent.

00:09:57: Oh wow!

00:09:58: What's fascinating here is that the bank security system stops asking the robot who it IS…and starts asking WHO stands BEHIND IT?

00:10:06: That is incredible.

00:10:08: And it perfectly brings us to this final piece – the workforce itself because at the end of day all comes down to humans standing behind a machine with quality of their judgement

00:10:17: which in pretty fragile state right now.

00:10:19: honestly

00:10:19: Yeah, Joshua Miller shared this statistic.

00:10:21: that is just a nightmare for management.

00:10:24: A KPMG survey of forty-eight thousand people found that fifty seven percent of employees are hiding their AI use at work.

00:10:32: Over half your workforce is secretly using it

00:10:34: and presenting the Ai's work as they're own original thinking.

00:10:38: Imagine doing performance reviews right now.

00:10:40: You can't tell if strategy belongs to you employee or if its just AI polish.

00:10:45: No,

00:10:47: you have to promote on the actual thinking.

00:10:48: You've heard them articulate in their room.

00:10:51: This is exactly what Saul Rashidi calls The Messy Middle.

00:10:54: Everyone expects massive productivity But an MIT study found that ninety-five percent of organizations saw no measurable return On Their Gen AI Pilots.

00:11:03: Ninety five percent.

00:11:04: That was a staggering failure rate for pilots.

00:11:07: It Is

00:11:07: And it gets crazier.

00:11:09: A Metra randomized trial showed that experienced developers using Ai tools were actually nineteen percent slower

00:11:15: Slower Slower.

00:11:17: But when surveyed, they believed there were twenty percent faster.

00:11:20: How do you lose that much time and still feel like your flying?

00:11:24: Because AI commoditizes the execution.

00:11:26: Typing the prompt feels instant but then the developer spends three hours untangling hallucinations in debugging sloppy logic that they didn't write.

00:11:35: so The real work shifts entirely to oversight.

00:11:38: right critical thinking taste discernment these are the new scarce capital.

00:11:44: If you don't have deep domain expertise, You can even tell if the AI's output is good or just confidently wrong.

00:11:50: Okay let's unpack this with a final story from Jimmy Acton because it perfectly captures everything we've talked about.

00:11:57: He watched this high-end robot teleoperation demo.

00:12:00: The motion capture one

00:12:01: Yeah!

00:12:01: A robot was copying every move of human operator zero leg.

00:12:05: But then the human accidentally flinched in pain

00:12:08: And the robot copied him perfectly.

00:12:10: Perfectly mimicked the flinch and inside there profound.

00:12:14: Perfect mimicry is not good judgment.

00:12:17: the hard part of building AI and integrating it into your company isn't getting It to follow instructions, its teaching when not too.

00:12:26: That's the ultimate through line here.

00:12:27: Over last fifteen minutes or so we've seen that technology is incredibly capable.

00:12:32: but our operating models are governance cues security layers.

00:12:36: they're all playing a frantic game of catch up

00:12:39: Absolutely.

00:12:40: Ultimately human judgement is only thing that scales AI safely.

00:12:45: It really is, and I want to leave you with one final kind of provocative thought on that note.

00:12:50: Joseph Noggle and Dr.

00:12:52: Benjamin Blau were discussing this recent claim from OpenAI

00:12:55: about the fluid dynamics problem?

00:12:57: Yes!

00:12:58: The Navier-Stokes math problem.

00:13:00: OpenAI says they solved it in just eighty eight hours.

00:13:04: The AI agents traded two point seven million messages and spat out a one hundred sixty six page paper

00:13:10: which sounds like a monumental breakthrough.

00:13:13: It

00:13:13: does, but top human mathematicians are looking at this paper and saying it's nearly unreadable.

00:13:19: The logic is so dense an alien that humans can barely verify the work.

00:13:23: So we have the answer But no actual comprehension

00:13:26: Exactly!

00:13:29: Here something for you to mull over This week.

00:13:31: As AI radically accelerates what We Can produce Are we investing enough in our ability To actually understand results?

00:13:40: If a machine gives you the right answer but can't trace how it got there, well is that actually progress?

00:13:46: That's the big

00:14:02: question.