Best of LinkedIn: Artificial Intelligence CW 25/ 26

00:00:00: This episode is provided by Thomas Allgeier and Frennus, based on the most relevant LinkedIn posts about artificial intelligence from calendar weeks twenty-five and twenty six.

00:00:09: Frenness supports ICT and tech providers with AI ecosystem strategy by delivering independent vendor assessment build versus buy analysis in Ecosystem Intelligence that prevents expensive mistakes.

00:00:23: You can find more info in the description.

00:00:24: Top artificial intelligence trends seen across LinkedIn.

00:00:27: Yeah, exactly that is our focus for today.

00:00:30: But before we really get into it I want you to just imagine a scenario For a second.

00:00:34: let's say your are banking executive

00:00:36: Okay and picturing

00:00:37: And customer applies for loan online.

00:00:40: An AI agent gets the application, cross-references all of the internal databases, flags the applicant for potential fraud denies the loan and then actually updates their permanent credit record.

00:00:51: And it does this in exactly eleven seconds.

00:00:55: Right!

00:00:56: Zero humans reviewed the file.

00:00:58: Zero humans signed off.

00:01:00: Now imagine you get sued tomorrow for discriminatory lending practices.

00:01:04: Who is legally responsible?

00:01:06: Is that YOU?

00:01:07: The developer who built a model?

00:01:09: The vendor who sold you the software.

00:01:11: I mean that is a terrifying question and for a lot of organizations right now, the honest answer Is just nobody knows exactly?

00:01:17: And That is exactly why we are doing this deep dive today.

00:01:20: We are pulling directly from curated LinkedIn insights shared by ICT and tech professionals to map out what is actually happening on the ground Right Now.

00:01:29: yeah in my mission For You listening To This Deep Dive is to help you realize one fundamental truth the era Of you know AI experimentation it's officially over

00:01:38: there really is.

00:01:39: We are firmly in the reality of twenty-twenty six.

00:01:41: The playground has been paved over that whole honeymoon for those have, oh let's see what this chatbot can write?

00:01:46: That's gone.

00:01:47: we've transitioned into a strict era of enterprise governance geopolitical sovereignty and just hard measurable business value.

00:01:53: You could

00:01:54: really feel the shift in tone across the industry.

00:01:57: it's palpable.

00:01:59: I mean we're moving from asking What Can This Technology Theoretically Do To Asking How do we control it?

00:02:06: Who is legally liable for it?

00:02:08: and Is that actually driving a concrete return on investment?

00:02:11: right those are significantly harder questions to answer.

00:02:13: they

00:02:13: Are, looking at the discussions from The past couple of weeks you can kind Of see the panic setting in around compliance.

00:02:19: Yeah because AI governance has morphed From this Conceptual PowerPoint deck into A rigorous unforgiving Operating discipline.

00:02:28: absolutely.

00:02:29: And the catalyst here is obviously the EU AI Act enforcement deadline.

00:02:34: yes

00:02:34: staring us down on August second, twenty-twenty six right.

00:02:37: Tom S. made a brilliant point on LinkedIn recently about this exact thing, he specifically warned financial institutions that are using AI for things like anti-money laundering transaction monitoring or fraud detection.

00:02:48: He basically said you have to build real governance before the August twenty-twenty six date because The gap between a company just saying hey we use AI and being able To actually prove to a regulator That We Have AI Governance that Gap is About to Become Incredibly Expensive.

00:03:04: Yeah, when you look under the hood of what Tom is describing.

00:03:07: The operational reality is just staggering.

00:03:10: It's no longer enough to let an AI work through a queue of compliance alerts.

00:03:15: You have to document its logic and mathematically prove that human oversight Is real Not

00:03:20: just a rubber stamp

00:03:21: Right not someone clicking.

00:03:22: approve And defend that moderns logic To external auditor

00:03:27: Which was huge heavy lift.

00:03:28: And to translate that overwhelming requirement into action, Kuba Sarmak highlighted a highly practical guide recently by Anna Tujikowska and her co-author.

00:03:38: I saw the way!

00:03:38: Yeah

00:03:39: it's great because it strips away all of legal fluff... ...and turns AI Act compliance in concrete operating steps.

00:03:45: So what did those steps actually look like for say IT & Compliance teams who were doing actual work?

00:03:51: Well It means strictly classifying your AI systems at risk level For starters thoroughly mapping out exactly who the end users are and continuously monitoring for model drift.

00:04:03: And by us over time,

00:04:04: right?

00:04:05: It's no longer about a legal team just asking Are we technically a provider or deployer of this software?

00:04:11: it's about executing a granular technical inventory an assigning hard individual ownership to every single algorithm running in your stack

00:04:20: Which naturally brings us to the financial stakes because getting that ownership wrong is going to cost you massively.

00:04:25: Oh yeah, yeah.

00:04:25: Io and Cyprian Popa pointed out everyone talks about a massive thirty-five million euro maximum fine under EU AI Act but very few people understand how it's actually applied.

00:04:35: Right.

00:04:36: its not arbitrary

00:04:37: Exactly Popa noted.

00:04:38: there is an actual deterministic formula the regulator uses to calculate these fines.

00:04:43: It's not just a random number a judge throws at you, it takes into account the severity of infringement and volume of users affected.

00:04:50: your failure to implement mitigation steps

00:04:53: Which means if your CFO doesn't understand inputs on that legal calculus they can't properly provision risk on balance sheet.

00:04:59: Exactly!

00:05:00: You are essentially flying entire AI stack blind And

00:05:03: look... If YOU are executive listening outside Europe Do not think you are exempt from this.

00:05:09: Zuria Nagadia and Sadio Jonas both explicitly warned that these fines, which can reach up to seven percent of your global revenue.

00:05:17: by the way they applied.

00:05:18: any company anywhere in that uses AI to serve European customers.

00:05:23: Seven percent of global revenue is just, it's staggering!

00:05:27: It is if you were a fintech startup in Lagos or US based retail executive deploying dynamic pricing tool.

00:05:34: If your AI model processes European data where its outputs are used the EU You're fully subject.

00:05:39: this law Extra-territorial reach is the defining feature of modern tech regulation.

00:05:44: I have to push back a little here though, i hear all of this and part of me just wonders isn't this really just GDPR?

00:05:50: two point oh?

00:05:51: How do you mean?

00:05:52: It feels like we're gearing up for massive check the box exercise A bunch of consent banners updated privacy policies some legal paperwork that ultimately slows everybody down without fundamentally changing the tech itself.

00:06:04: Yeah!

00:06:04: I get it's very common reaction but its fundamentally incorrect.

00:06:09: AI regulation operates on an entirely different mechanical level than data privacy.

00:06:14: Okay, unpack that for me.

00:06:16: Kevin L meticulously broke this down and linked in recently.

00:06:19: He points specifically to article ten of the AI Act.

00:06:23: Under this Article Active Data Governance Continuous Bias Checks And Rigorous Data Quality Are Now Legal Standards.

00:06:30: Let That Sink In.

00:06:31: Under GDPR Your Job Was Basically Simply To Protect The Data From Being Stolen.

00:06:36: Under the AI Act, you have to mathematically prove that data you fed your model is representative relevant and statistically appropriate for this specific context.

00:06:45: You're using it in

00:06:46: okay?

00:06:46: Wow so garbage in garbage out Is no longer just an IT headache or like a technical glitch.

00:06:51: It is a strict compliance failure.

00:06:53: right exactly if an AI model makes a biased decision feeds that output back into its own training data and compounds that bias over time, breaking the technical cycle is now a legal obligation.

00:07:06: So if IT department can't just slap a privacy banner on this or call it day...and fines are globally enforceable…it means liability leaps right over middle management and lands directly in boardroom.

00:07:18: Yep!

00:07:18: Completely bypasses them.

00:07:19: Which exactly where this concept of the Agentec Liability Fog comes to play.

00:07:25: Shobha Shah posted these five hard-bored governance questions that corporate directors absolutely need to ask before they approve any major AI project.

00:07:34: Those

00:07:34: are so good!

00:07:34: They really were, the two One, who owns the business outcome?

00:07:39: And two what happens when the model is wrong.

00:07:42: And those questions are so critical right now because of how the underlying technology has evolved from generative to agentic and this ties perfectly back to that eleven second loan scenario we opened with.

00:07:52: Salman Lagesse shared that chilling real-world example from Canada a fully autonomous AI agent declined alone application in seconds updated a permanent record.

00:08:05: Wow.

00:08:06: Jess calls this the agentic liability fog, it's what happens when AI acts autonomously making real world decisions and literally no one in the corporate hierarchy can explicitly name who authorize that specific action.

00:08:21: And if you're listening and think autonomous agents are just a fringe use case for tech companies Think again, Brad Wolf shared a jaw-dropping statistic.

00:08:29: Agentech AI adoption is already at eighty eight percent in legal departments and eighty nine percent in recruiting.

00:08:34: that's huge it Is.

00:08:35: these are pure judgment functions?

00:08:36: We aren't talking about Generating a marketing email.

00:08:39: we're talking about an agent the triages of human beings career prospects or Flags A critical liability clause In a vendor contract.

00:08:47: yeah when an agent executes a multi step workflow like That who signs off on It honestly feels Like a corporate game Of musical chairs.

00:08:53: now so

00:08:54: Well, the music is playing at a hundred miles an hour.

00:08:56: Right?

00:08:57: These AI agents are making thousands of micro-decisions in the background but when the music stops say... When there's class action lawsuit or massive financial hallucination no human sitting in accountability chair.

00:09:10: That analogy perfectly captures systemic risk boards we're facing right now And Alexander C. has a really important technical layer to this on LinkedIn, she pointed out that agentic failures compound.

00:09:21: What does it mean exactly?

00:09:23: Well if the traditional generative AI model gives you wrong answer A human spots it fixes and moves on.

00:09:28: The air is isolated.

00:09:30: But when an agent makes autonomous decision It acts on it.

00:09:33: Might send email update database Then use new state To make its next decisions.

00:09:38: Oh man!

00:09:39: Failure doesn't sit still.

00:09:40: It cascades through your systems.

00:09:42: This makes traditional point-in time assurance, like auditing a model once before launch completely useless.

00:09:47: You have to govern the decisions across the entire life

00:09:50: cycle.".

00:09:50: So if boards are suddenly waking up to the terrifying fact that they don't legally own the decisions being made inside their own companies... They're simultaneously waking up an even more severe reality!

00:10:04: geopolitics enters the check.

00:10:05: Exactly!

00:10:06: Think about what happened on June twelfth, Greg Malpass and Ali Yüksel brought up this massive incident where the US government ordered Anthropic to shut down its top models worldwide due to cybersecurity concerns.

00:10:18: I remember that day pure chaos

00:10:20: with zero warning.

00:10:21: thousands of European organizations lost access to their models but more importantly they lost there AI's constitutional memory.

00:10:30: right wait help me visualize for a second though.

00:10:32: are we saying API goes down, a European company doesn't just lose it.

00:10:36: A software tool?

00:10:38: they lose the accumulated business logic of their entire operation.

00:10:41: that is exactly what happens and That concept of constitutional memory Is so vital to understand because when you deeply integrate a frontier model into your enterprise You aren't Just using an out-of-the-box product anymore?

00:10:53: Your fine tuning It you are feeding it.

00:10:56: your corporate rules your specific operational parameters your audit trails All of that custom reasoning only existed inside an infrastructure that Europe didn't own and couldn't protect.

00:11:08: Daniel Yusitalo noted, losing access to a frontier model like Fable V overnight is the ultimate undeniable case for Europe building sovereign AI.

00:11:18: And Deng Hong pointed out a really dangerous geopolitical mismatch here.

00:11:23: Europe is incredibly successful at exporting its rulebook, we always talk about the Brussels effect right?

00:11:28: Sure

00:11:28: where countries all over the world just copy European tech regulations

00:11:31: Exactly.

00:11:32: but Europe isn't exporting it's models or compute infrastructure.

00:11:35: so We are basically renting our society cognitive infrastructure from a geo-political landlord who can change the locks without notice.

00:11:43: That's

00:11:43: great way to put It.

00:11:44: But Can Europe actually catch up?

00:11:46: F. Garrick Verhees suggested building massive data centers.

00:11:50: Is pouring money into server farms enough to solve a sovereignty issue?

00:11:54: Building data centers is a necessary prerequisite, I mean you have to have them but it's fundamentally not the whole solution.

00:12:01: You need the models The specialized silicon and the talent.

00:12:05: Maximilian Hanning highlighted a really critical dynamic here.

00:12:08: Washington has actively tightening control over its AI firms And selectively squeezing Europe out of access cutting-edge frontier models, limiting them only to quote unquote trusted partners under strict conditions.

00:12:24: It's all about controlling the threshold of computing power or FLOPs.

00:12:29: This proves that digital independence is no longer just an economic ambition for Europe, it's a baseline survival metric.

00:12:36: You simply cannot enforce the sovereign rulebook on a cognitive engine.

00:12:40: you do not

00:12:40: control.".

00:12:58: Yet despite all of this, businesses are still rapidly adopting AI.

00:13:03: Why?

00:13:03: Because the financial stakes are simply too high to sit it out.

00:13:07: but The nature of that adoption has completely changed.

00:13:09: business value has officially replaced adoption theater.

00:13:12: Thank

00:13:12: goodness

00:13:13: right.

00:13:14: Michelle Rao offered a really sharp critique of this shift.

00:13:17: She says executives need to stop celebrating adoption metrics, Stop counting the number of software licenses you bought or how many prompts your employees wrote This week.

00:13:26: start asking what KPIs actually moved?

00:13:29: Did Your cycle time improve?

00:13:31: did your expense ratios Actually drop?

00:13:33: this is exactly where The maturity of the market Is showing and it Really just comes down To basic unit economics.

00:13:38: yeah, theft key corner listen made A brilliant observation about this.

00:13:41: she said AI models now run on a meter.

00:13:44: What does she mean by that?

00:13:45: Well, during the early experimental phase AI was often just a flat subscription fee.

00:13:49: Twenty month-a-month whatever.

00:13:50: Now if you are building enterprise applications via APIs every capable model charges per token per task.

00:13:57: You were pushed for the compute power required for every single action.

00:14:01: This usage meter acts as a mirror for the organization.

00:14:04: It exposes which automated work actually creates real value and which works looks busy.

00:14:10: You are literally paying for every decision you hand to the machine.

00:14:13: And how you build the architecture behind that machine determines if you go bankrupt running it.

00:14:18: David Lenthicum shared a striking example of this.

00:14:20: I read that

00:14:21: one, yeah

00:14:21: He looked at three different enterprises That solved the exact same AI problem.

00:14:26: They achieved the exact Same business value But One company's architecture cost one million dollars The second cost three million and the third costs ten million dollars.

00:14:37: Just wild!

00:14:38: It's insane, architecture discipline literally dictates your ROI

00:14:42: And we should probably unpack how that actually happens.

00:14:45: A ten-million dollar architecture is usually what happens when a company blindly sends every single internal data query to massive premium frontier model API.

00:14:54: They pay top tier token costs for simple tasks.

00:14:57: Just

00:14:57: burning money?

00:14:58: Exactly, a one million dollar architecture on the other hand utilizes smart routing.

00:15:04: They might use a cheap, locally hosted open source model for ninety percent of basic sorting tasks and only call the premium API for highly complex reasoning.

00:15:13: Oh that makes sense!

00:15:14: Furthermore they optimize their Retrieval Augmented Generation Pipelines or RG so they aren't unnecessarily feeding massive documents into the context window over-and-over again.

00:15:24: You cannot just throw premium cloud compute at problem.

00:15:27: That makes perfect sense.

00:15:29: And part of that architectural discipline is also how you map the AI to your human organization.

00:15:35: Katia Casano issued a really strong warning against rebuilding your organizational silos inside your AI, like if you are a CT listening right now.

00:15:44: look at your Ork shirt.

00:15:45: If you're deploying dedicated sales bot and a separate finance bot, an isolated support bot you are literally just recreating your company's departmental walls in code.

00:15:54: Exactly!

00:15:55: The leading edge of enterprise architecture is running one shared AI operating system that reconfigures itself per task drawing from a shared data foundation.

00:16:04: Because if we deploy siloed bots aren't we building faster horses instead inventing the car?

00:16:10: We're taking our own dysfunction or communication bottlenecks between departments and automating them

00:16:15: Precisely.

00:16:17: You're just speeding up the inefficiencies that already exist.

00:16:20: And to make things even more complex, The underlying technology is shifting again.

00:16:25: Of course it does!

00:16:26: Georgina Tilly points out that basic generative AI.

00:16:30: you know...the models that predict next word and write an email or summarize a document.

00:16:35: That's now just the floor.

00:16:37: It's a commodity The real competitive gap being formed by organizations building neuro-symbolic reasoning systems.

00:16:44: Harro symbolic reasoning, we hear that term thrown around a lot but let's demystify it for the audience.

00:16:48: What does that actually mean in practice?

00:16:51: It's actually a fascinating combination of two different types of AI.

00:16:55: generative neural networks are incredible at guessing patterns But they are terrible and strict logic which is why they hallucinate facts.

00:17:02: Symbolic AI on the other hand Is The Old School Approach.

00:17:06: It relies On Strict Hard Coded Logic Rules Like If X Then Y. Its Highly Accurate But Very Brittle

00:17:13: Okay, so it can't handle nuance.

00:17:15: Right!

00:17:15: So Neurosymbolic reasoning combines them.

00:17:18: The neural network acts as the creative pattern matching engine but its outputs are constantly checked against a symbolic logic engine based on your company's actual hard facts.

00:17:28: Oh wow Yeah, it allows the AI to literally show its math and prove the causal logic behind his decisions which as we discussed earlier is exactly what the EU-AI Act requires.

00:17:39: Which naturally leads us to humans who have to operate these advanced systems because you cannot restructure the foundational technology of a company without radically restructuring the human's doing their work?

00:17:49: Definitely not!

00:17:49: Trisha Srivastava pointed out that talent market isn't splitting into simple binary categories for winners using AI or losers don't.

00:17:57: Instead, every single role is being entirely redefined from the ground up.

00:18:01: Yes and Andreas Horn offered some incredibly practical advice for professionals navigating this transition.

00:18:07: he says do not fear job losses.

00:18:10: instead focus on mastering the AI tools inside your own specific domain

00:18:14: so leaning into your niche

00:18:16: right because there's a misconception that most valuable person in future is generic AI engineer sitting at lab.

00:18:23: In reality, the person who successfully maps what an AI agent can do in a highly specialized niche whether that's forensic accounting or specialized supply chain logistics is The Domain Expert.

00:18:34: It makes a lot of sense!

00:18:35: The domain expert understands the messy unwritten edge cases of the real world That a generic model misses...that combination Of deep narrow-domain knowledge and broad AI fluency Is THE NEW PREMIUM SKILL.

00:18:48: But the transition to that new reality is causing significant friction and not everyone is thrilled about it.

00:18:54: Sol Rashidi brought up a fascinating cultural data point from recent surveys, thirty one percent of Gen Z feel real anger toward AI in work.

00:19:02: place an education.

00:19:03: That's high number!

00:19:04: It would be really easy for executive to dismiss stubborn resistance but Rashidi argues leaders should see this as honest feedback.

00:19:14: This is a generation actively trying to protect the value of actual learning.

00:19:18: They don't want it just be handed an answer by an agent, they wanna understand how that was formed.

00:19:24: That is really interesting.

00:19:26: It makes me think of navigation!

00:19:27: Yeah, it's exactly like handing a teenager a GPS system before they ever learn how to read a physical map or understand the concept of street grids.

00:19:36: Yes The technology gets them to destination faster But They know deep down that relying purely on machine Is robbing them of foundational understanding Of their own environment.

00:19:46: I Think there was very profound way To look at And it ties directly into a brilliant insight from Dr.

00:19:52: Annette Domes.

00:19:53: She notes that for centuries human intelligence and the ability to produce a correct answer was scarce, and therefore It was highly valued.

00:20:01: now AI is making intelligence abundant and incredibly cheap.

00:20:05: if you need a marketing strategy A block of code or illegal summary You can get it instantly for fractions of ascent right but when answers become cheap commodities The ability to ask the right questions becomes priceless.

00:20:17: human judgment, the wisdom to challenge a model's assumptions and critical thinking required to distinguish signal from noise.

00:20:25: Those are skills that will define value in an AI-saturated workforce.

00:20:29: As we wrap up this deep dive into the reality of AI in twenty-twenty six, I want to leave you with a final provocative thought from Ben Torben Nielsen.

00:20:39: He expands on this exact idea that generative AI made knowledge cheap and he says it forces really uncomfortable question for every professional.

00:20:48: You have to ask yourself does your specific field of work operate like TV repair?

00:20:53: Or does it operate like quantum physics?

00:20:55: Ooh, I liked that.

00:20:56: Right?

00:20:57: Think about it!

00:20:57: A TV repair person doesn't necessarily need to understand the fundamental electronics of every single microchip on the board.

00:21:03: they just need know which component to replace and get screen working.

00:21:06: again Generative AI is perfect for kind of modular replacement level work But in Quantum Physics you cannot look up a breakthrough.

00:21:14: The deep internal understanding of mechanics is actual work.

00:21:17: It raises critical defining question.

00:21:20: everyone listening In the next decade, as autonomous AI agents increasingly take over the execution of tasks knowing which camp your career falls into.

00:21:30: The replaceable component camp or the deep understanding camp will entirely dictate your professional trajectory.

00:21:37: It all circles back to the muddy diagnostic landscape we are navigating right now.

00:21:47: It's completely dead.

00:21:48: We have to look at the rigorous mechanisms of governance, the compounding risks of liability, the efficiency of our enterprise architecture and ultimately the irreplaceable wisdom of humans in.