Best of LinkedIn: Artificial Intelligence CW 19/ 20

00:00:00: This episode is provided by Thomas Algeier and Frennis, based on the most relevant LinkedIn posts about artificial intelligence from calendar weeks nineteen-and-twenty.

00:00:08: Frennes supports ICT and tech providers with AI ecosystem strategy by delivering independent vendor assessment build versus buy analysis an Ecosystem Intelligence that prevents expensive mistakes And positions the providers competitively.

00:00:22: you can find more info in the description.

00:00:24: And welcome back to the Deep Dive, everyone.

00:00:27: We are really excited to get into this one today!

00:00:29: Yeah totally because we're unpacking the absolute top artificial intelligence trends and you know...the real ground truth insights that we've seen across LinkedIn over calendar weeks nineteen and twenty

00:00:42: Right.

00:00:42: And we're specifically filtering this for you, the professionals in the ICT and tech industry because...I mean let's face it there is so much noise out there

00:00:49: right now.

00:00:50: Oh an incredible amount of noise.

00:00:51: Yeah.

00:00:51: So were organizing todays deep dive into three main clusters basically.

00:00:56: First how AI governance.

00:00:58: well its officially becoming a board level execution priority

00:01:01: No longer just a hypothetical debate

00:01:03: Exactly.

00:01:04: Then Cluster two, we're looking at how agentic AI is moving from just being a cool concept into actual production infrastructure.

00:01:13: And then finally cluster three- How enterprise AI adoption is shifting away from vanity metrics and toward you know measurable business impact?

00:01:21: Right.

00:01:22: so let's jump right in to that first cluster... AI governance.

00:01:27: I mean, it feels like he has officially moved out of the philosophy department and straight into

00:01:56: transparent.

00:01:57: You

00:01:57: have to prove they're governed which I feel like.

00:01:59: i see companies all the time trying to find uh loopholes or they're just crossing their fingers hoping that timeline gets pushed back again.

00:02:05: yeah, Which is a super dangerous game.

00:02:08: Oliver Patel actually added a really critical layer of nuance to this exact point on LinkedIn.

00:02:14: What'd he say?

00:02:15: Well, He pointed out that yes there is a recent political agreement To delay enforcement for very small subset of high-risk AI systems too I think December twenty twenty seven but and this is the key part.

00:02:27: The core structure of the AI Act remains completely

00:02:30: intact.".

00:02:31: Right, so you can't just use that tiny delay as an excuse to put off your entire governance overhaul?

00:02:35: Exactly!

00:02:36: The underlying legal framework hasn't changed at all.

00:02:39: Organizations really cannot afford to deprioritize this.

00:02:42: Yeah...and make it super tangible for ya if you're sitting in a enterprise right now.

00:02:46: Aminiciator provided this brilliant framework Of five critical questions That honestly Every tech leader needs to be able to answer today.

00:02:55: Oh, yeah the five questions run through those because they are tough

00:02:57: They are okay.

00:02:58: question one how many AI agents?

00:03:00: Are currently running across your organization

00:03:02: sounds simple but most companies have no idea right.

00:03:05: Question two Which of those systems make decisions affecting EU residents?

00:03:10: a massive compliance trigger rate there yep.

00:03:13: three Can you produce a ninety day audit trail for any AI decision?

00:03:18: four What is the exact threshold where an agent is forced to escalate into a human?

00:03:23: The

00:03:23: human in-the-loop

00:03:24: cutoff.

00:03:25: Exactly, and five who was the named owner of your highest risk systems?

00:03:30: Yeah And Cieta warns you know.

00:03:31: if you can't answer these cleanly You are basically leaving yourself exposed to a thirty five million euro gap under the new

00:03:37: rules.

00:03:38: Wow Thirty five million.

00:03:39: I want zero on that third question though than ninety day audit trail because i think it exposes A really fundamental technical misunderstanding out there.

00:03:48: how so like Comparing it to traditional software.

00:03:50: Exactly, producing an audit trail for traditional software is well It's pretty straightforward right?

00:03:54: It's a rigid flowchart.

00:03:55: if A happens do B?

00:03:57: You just

00:03:57: check the logs you see exactly what branch of logic The code took.

00:04:00: Right

00:04:01: but producing and audit trail For a probabilistic model totally different.

00:04:04: beast large language models don't follow hard-coded logic paths.

00:04:08: They generate outputs dynamically based on you know statistical weights in these massive context windows.

00:04:14: So you can't just print out a log file say hey Here is why the AI denied this person's loan.

00:04:19: No, you absolutely can't!

00:04:20: You would have to capture the exact state of a model's weights—the specific prompt and system instructions….

00:04:26: …the contextual data all present at that exact millisecond it was generated

00:04:30: in.

00:04:31: Which means if you didn't architect your system... ...to capture that whole multi-dimensional snapshot from day one?

00:04:36: You're

00:04:36: out of luck —you cannot reverse engineer The Y ninety days later.

00:04:41: Man so….

00:04:43: Treating governance like it's just some legal checklist is actually a massive technical failure.

00:04:48: A hundred

00:04:48: percent.

00:04:48: Philip Braun tackled this beautifully.

00:04:51: He argued that for a lot of companies their compliance strategy isn't to shield It's a sedative.

00:04:57: Oh Sedative, that is great analogy right?

00:04:59: Like they are passing a building inspection by showing the inspector a really nicely formatted PDF Of The Building Codes instead letting them inspect the actual steel beams in the foundation.

00:05:10: That perfectly visualizes it.

00:05:12: Braun actually points out that compliance under the EU AI Act is an architectural requirement, not a spreadsheet.

00:05:18: He outlined this five-layer model for sovereign infrastructure to handle

00:05:23: it.

00:05:23: Okay what kind of layers are we talking about?

00:05:25: Well things like workload isolation at the bare metal level and hardware attestation.

00:05:31: okay let's break that down because Hardware Attestations sounds some serious deep data center jargon.

00:05:36: Huh,

00:05:36: fair enough!

00:05:37: Basically instead of just trusting a cloud vendors contract that says you know your data is safe hardware attestation means the physical microchip itself mathematically proves That no unauthorized code has tampered with the system before your AI workload runs.

00:05:51: Oh wow so the chip itself Is verifying the integrity?

00:05:54: Exactly and brawn also talks about cryptographic sovereignty where You use advanced encryption to lock Your data so firmly that even The Cloud provider physically holding your servers can't peer inside the black box.

00:06:06: So you're literally hard-coding the law into the plumbing of your tech stack?

00:06:11: That's incredible, but there is a completely different layer to governance that does not show up on any architecture diagram.

00:06:17: Oh!

00:06:18: The human element.

00:06:19: Yes

00:06:20: Tiffany Masson brought out the psychological elements of AI adoption in LinkedIn and honestly it blew me away.

00:06:26: Yeah, her insights were fascinating.

00:06:29: She argues that when governance and AI deployment hit a brick wall inside enterprises leadership usually just blames the training gap.

00:06:36: but Madsen says it's not a training gap at all.

00:06:38: its identity protection.

00:06:40: Right!

00:06:40: The psychology of the expert is so rarely factored into the rollout like she used the medical field as an example right?

00:06:47: Picture someone who has been a radiologist for twenty years.

00:06:51: Their entire professional identity, their status they're value to the hospital.

00:06:55: It's all tied to their unique ability to analyze a complex scan.

00:06:59: And then IT installs an AI tool that reads that identical scan in three seconds.

00:07:03: Exactly if I'm not radiologist my brain does not register.

00:07:07: That as an efficiency upgrade My nervous system registers.

00:07:10: that is professional erasure

00:07:11: precisely Because the machine is suddenly performing the exact cognitive function that made The Human valuable.

00:07:17: Right And when that radiologist pushes back or refuses to use the tool, Or maybe subtly undermines it They aren't lacking technical training.

00:07:26: they are defending their identity

00:07:28: Which becomes a massive governance risk in itself.

00:07:31: Huge because if your human experts feel threatened?

00:07:34: They stop trusting Their own judgment.

00:07:37: They might just blindly accept the AI's output to avoid conflict, or they might reject totally valid A.I findings out of spite

00:07:44: which destroys The entire safety net.

00:07:47: if the humans stop supplying critical thought you essentially Just have an ungoverned machine running.

00:07:52: this show right

00:07:53: and mass on solution here is pretty profound.

00:07:56: To govern a I safely You have to redesign the workflow so the human expert Is elevated to the ultimate judgment layer?

00:08:02: So the

00:08:03: AI is like the resident in the human as the attending doctor

00:08:06: Exactly.

00:08:07: You frame the AI as the tireless assistant doing the preliminary grunt work, but that twenty-year radiologist remains the irreplaceable authority making the final call.

00:08:17: Preserving humans' professional identity is a prerequisite for system safety.

00:08:22: I love it And i really want to pull on this thread of human being and ultimate judgment layer because its sets us in direct collision course with next massive trend.

00:08:32: we saw in data Cluster two Agenetic AI.

00:08:36: Oh yeah, we are moving rapidly past the era of chat interfaces

00:08:41: Right where a human types of prompt reads the text and then the human takes an action.

00:08:45: We're entering The Era Of Autonomous Agents Designed To Take The Action Themselves.

00:08:49: If Governance Is The Boundary Line Agenic AI is the force pushing right up against it

00:08:54: And that transition breaks the fundamental security models that enterprises have relied on for decades.

00:08:59: Shoaiba Khan raised a really brilliant warning flag about this.

00:09:02: The shift in zero trust paradigm.

00:09:04: Yeah,

00:09:05: look at traditional Zero Trust Security.

00:09:06: It assumes someone is always trying to break-in.

00:09:08: so it focuses entirely on controlling access.

00:09:10: Right You authenticate user's identity you check their device and if all matches... ...you unlock door of data.

00:09:17: Because historically, once the human was inside of system... The Human made decisions.

00:09:22: The software just executed whatever the human clicked.

00:09:25: But an AI agent operates totally differently

00:09:27: Exactly!

00:09:27: The Agent is given a high level goal.

00:09:29: It reasons across different options, selects a path and executes it.

00:09:34: Your security stack looks at network traffic And sees perfectly authenticated API call from authorized internal agent.

00:09:40: So it lets action through?

00:09:42: Yep

00:09:43: The vulnerability didn't happen in AccessGate.

00:09:46: It happened a step earlier inside the reasoning layer of the model.

00:09:49: Wow,

00:09:50: so this system verifies that the agent is allowed to say delete files but it doesn't verify if the agent made logical decision to delete those files?

00:09:59: That's Khan whole point.

00:10:01: enterprise risk has shifted from access control to Decision Control.

00:10:05: The modern tech stack just does not have the intervening mechanisms To govern an AI real-time reasoning process before becomes executable command.

00:10:14: Wait, okay.

00:10:14: Let me challenge this for a second because I was reading Chris Leone's thoughts on the exact issue.

00:10:19: aren't The big frontier model vendors trying to solve This reasoning gap just by giving the AI agents better memory?

00:10:27: you mean like larger context windows?

00:10:29: Yeah Like if the model has a massive Context window and remembers the company's purchasing rules from a previous prompt doesn't that effectively govern its decisions?

00:10:39: It is so tempting to think, but Leone points out a really dangerous flaw in that logic.

00:10:44: We are conflating model memory with enterprise state.

00:10:48: they're fundamentally different things.

00:10:50: okay walk me through the difference there.

00:10:52: well model memory is probabilistic.

00:10:54: if an agent remembers a chat log from yesterday where manager approved of five thousand dollar purchase order that it's just historical context.

00:11:00: Okay The agent predicts buying item is correct next step based on text.

00:11:06: But Enterprise State is deterministic reality.

00:11:09: Knowing if that five thousand dollars is actually in the bank account today, knowing if the vendor's insurance Is currently valid.

00:11:15: That truth lives In your ERP Your Enterprise Resource Planning System.

00:11:20: Right The ERP is the actual nervous system.

00:11:23: Holding hard math.

00:11:25: Yes Relying on a language model's probabilistic memory to execute deterministic business logic is just incredibly reckless.

00:11:33: Leone argues that the winners in Enterprise AI won't be the vendors who build best memory retrieval tools.

00:11:40: Who will it then?

00:11:40: It'll be platforms that embed the AI agents directly inside of transactional state-of-the-business, like ERPs and supply chain databases.

00:11:49: So agent is constrained by real time mathematical reality not text prediction.

00:11:54: That distinction really changes how you evaluate every piece of AI software on the market.

00:12:00: And we are already seeing this agentic shift happening in The Wild at a scale that is honestly hard to wrap your head around.

00:12:06: Oh, You mean the Allegro example?

00:12:07: Yes Rob Van Gent and Michael Westfield pointed it out.

00:12:11: Agentic commerce has no longer concept It's live infrastructure.

00:12:15: Allegro which Europe's largest e-commerce marketplace Is now operating as native application inside chat GPT.

00:12:20: The implications for that are just staggering.

00:12:23: Twenty-two

00:12:23: million users, twenty two million people who can now bypass the traditional website search bar entirely.

00:12:30: They are researching comparing and buying products through a conversational interface with an agent

00:12:36: And Westerwheel makes the case that Traditional Search Engine Optimization SEO is effectively dead in this environment.

00:12:43: Yeah think about how our retailer normally operates.

00:12:45: you write product descriptions You design beautiful web pages to appeal to a human eye

00:12:52: But in Agent-to-Commerce, your product attributes—your return policies and technical specs —they are being ingested by an AI agent acting on behalf of the buyer.

00:13:01: The agent does not care about your marketing copy… Right!

00:13:04: It cares about structured data...

00:13:05: Exactly!!!

00:13:06: It's looking at the JSON file to find your API latency Your exact delivery service level agreements Your historical direct rates Fulfillment speed and dare hygiene suddenly become the direct ranking factors for whether the agent recommends your product to the human.

00:13:21: The agent literally stands between a buyer in the seller, so if agents are out there executing commerce or making autonomous reasoning decisions it forces is very uncomfortable question from boardroom which brings us our final cluster

00:13:34: right how we actually keeping score

00:13:37: exactly?

00:13:38: How do you measure value of machine that does work fifty employees?

00:13:43: The honeymoon phase of AI adoption is over.

00:13:47: Enterprises are being forced to pivot from measuring mere activity to measuring hard business impact.

00:13:53: And Andreas Horne had a really fantastic take on this in LinkedIn, he stated that most companies are completely confusing AI adoption with AI Impact.

00:14:01: Oh the vanity metrics!

00:14:03: Yes if you look at the dashboards and most IT departments right now they're bragging about things like how many software licenses were deployed or how many thousands of tokens we consumed Or

00:14:12: How Many Employees Log Into A Co-Pilot Tool.

00:14:14: This Month

00:14:15: Right Tracking token consumption is like a factory manager bragging about how much electricity they used without ever checking if any products actually rolled off the assembly line.

00:14:23: That's a brutal comparison, but it's so true!

00:14:26: Horne argues that the only metrics that matter now are revenue generation time-to-value reduction and freed workforce capacity.

00:14:34: If you cannot draw a direct line between an AI tool or a compounded business value You don't have to change management

00:14:42: problems And the financial sector is already brutally enforcing this standard.

00:14:47: Dr.

00:14:47: Efi Pilariniu highlighted that major global banks institutions like BNY, State Street and CIBC are altering how they communicate with Wall Street.

00:14:56: Yeah!

00:14:57: This a massive shift.

00:14:58: They're no longer hiding AI investments in capital expenditure lines of their IT budgets... ...they report AI as core operating metric on quarterly earnings

00:15:08: calls.

00:15:08: When an institution like BNY starts treating AI as operating leverage, the entire market pays attention.

00:15:14: I mean they are reporting hundreds of AI solutions actively in production.

00:15:18: In

00:15:18: pointing to tangible measurable gains and fraud detection efficiency and credit monitoring accuracy They're setting a standard that says we didn't just buy tool.

00:15:26: this tool drove exactly this much bottom line value.

00:15:29: But implementing this at scale reveals a massive hidden vulnerability in how companies are structured, and I know you wanted to talk about that Lumen case study.

00:15:38: Oh yes!

00:15:39: Christopher Rainey recently interviewed the Senior Vice President of People Ops at Lumen And...the case study he shared is just mind-bending.

00:15:47: Lumen isn't running some small pilot program.

00:15:49: They currently have five thousand AI agents running live across their business.

00:15:55: Five thousand autonomous agents, that's a whole new workforce.

00:15:58: A blended work force alongside humans.

00:16:00: and Rainey asked them the million dollar question when you deploy AI at that massive scale what breaks first?

00:16:07: And it wasn't the tech was it?

00:16:08: Nope!

00:16:09: It wasn't API limits or cloud infrastructure...it was leadership.

00:16:13: The actual management structure of this company broke down

00:16:16: which honestly makes perfect sense if we look how traditional middle-management operates.

00:16:20: Right, think about a manager whose entire job for the last decade has been assigning routine tasks tracking activity and hoarding specialized information as a form of job security.

00:16:31: And

00:16:31: suddenly five thousand Asians are handling all that routine work instantly!

00:16:35: The AI has democratized that specialized information to every junior employee in the division...the managers literally had nothing left to track.

00:16:44: This perfectly aligns with what Fenke Cornelisyn termed the transformation paradox.

00:16:48: She highlighted a staggering finding on LinkedIn, ninety-five percent of generative AI rollouts delivers zero measurable return on investment.

00:16:57: Zero?

00:16:58: How do you deploy and get zero return.

00:17:03: Because companies are treating AI like a software patch rather than an organizational redesign, Corneliusen argues that if you bolt a highly efficient AI agent onto a fundamentally broken internal process—you don't get transformation... You just get faster-broken meeting.

00:17:19: Oh!

00:17:19: A Faster Broken Meeting?

00:17:21: That it's painfully close to home for anyone working in tech doesn't.

00:17:24: If your company's key performance indicators still reward the old shape of work like billing by hour or producing a sheer volume of code regardless quality, The Old System will always reject new technology.

00:17:35: So you have to redesign roles themselves?

00:17:37: Exactly!

00:17:38: The role itself has be the unit of redesign.

00:17:41: You need change communication lines and incentives Not just give an employee faster typing tool.

00:17:48: I want zoom out from enterprise level for second To look at how this transformation plays on macro scale Because George Osborne shared an initiative that acts as a brilliant blueprint for rapid literacy.

00:17:59: The Malta Initiative, right?

00:18:01: Yeah

00:18:01: the nation of Malta is rolling out a program to give all five hundred and fifty thousand of its citizens access to chat GPT plus for a full year.

00:18:10: but... And this is the genius part they tied the access to the completion of a mandatory AI Literacy course.

00:18:17: That is incredible.

00:18:18: Treating artificial intelligence access like a national utility,

00:18:21: right?

00:18:22: It cuts through all the endless corporate strategy papers and multi-year consultation processes.

00:18:27: Malta recognizes that you do not build AI literacy by talking about it in seminars.

00:18:32: You built it by radically lowering the cost of experimentation.

00:18:35: Exactly!

00:18:36: You put the tool on the hands of the population pair with practical guardrails And let them figure out how to apply into their daily lives.

00:18:42: It is the ultimate exercise in applied execution over theoretical planning.

00:18:47: And, you know, SIRS is a really powerful mirror for everything we have unpacked today... Oh definitely!

00:18:52: ...whether you are regulating algorithms in the EU deploying agentic commerce infrastructure or redesigning roles at a major telecom The technology is ruthlessly exposing the flaws and our outdated operating models.

00:19:05: Yeah We have to redesign the tracks while the locomotive is already top speed

00:19:09: which leaves us with a final, highly personal question for you to consider.

00:19:13: If you look at that Lumen case study where thousands of agents are deployed to handle the routine transactional friction of the business what happens to your role tomorrow?

00:19:22: if an agent has perfect access to your company's enterprise state and can execute the daily processes flawlessly What exactly is the unique irreplaceable judgment Headcount will achieve short-term savings, sure.

00:19:38: But the professionals who use this moment to offload the routine and elevate their own capacity for high level critical judgment—those are ones that dictate future of.