Best of LinkedIn: Artificial Intelligence CW 23/ 24

00:00:00: This episode is provided by Thomas Allgeier and Franis, based on the most relevant LinkedIn posts about artificial intelligence from calendar weeks twenty-three in twenty four.

00:00:09: Franus supports ICT and tech providers with AI ecosystem strategy by delivering independent vendor assessment, bilver's buy analysis an ecosystem intelligence that prevents expensive mistakes and positions to providers competitively.

00:00:23: you can find more info.

00:00:24: description.

00:00:26: So what if I told you that despite the literal billions of dollars being poured into enterprise AI right now like nine out of ten executives are getting absolutely zero Measurable productivity gains out of it.

00:00:39: Wow, yeah zero

00:00:40: Right zero.

00:00:41: so today we're going to look at The Hidden Traps causing That massive failure and We're pulling this directly from the sharpest insights shared across the ICT in tech industry on LinkedIn.

00:00:50: Yeah And If You Are an Enterprise Leader tuning In?

00:00:55: Pretty critical.

00:00:56: We're going to break down how the illusion of AI governance is leaving companies completely exposed and why the models you rely on could just vanish overnight due to geopolitical shocks, plus the highly misunderstood security flaws inside autonomous agents.

00:01:12: And finally Why those promise productivity gains?

00:01:15: You mentioned are basically just flatlining.

00:01:17: it really Really is a sequence of overlapping crises, honestly.

00:01:21: Because as these tools move out the sandbox and into live corporate environments there's just fundamental misunderstanding what it actually means to manage them.

00:01:30: Yeah that makes sense.

00:01:31: And I think most dangerous misunderstandings starts with how companies buy this technology.

00:01:37: Okay yes!

00:01:38: I want look at point raised by Dr.

00:01:40: Andre Bates on this because she highlighted trap guarantee happening in board rooms everywhere right now.

00:01:46: Oh for sure.

00:01:48: She pointed out that companies are confusing procurement with governance.

00:01:52: Like you'll hear a CTO say, oh we're secure because we only use Microsoft Copilot right?

00:01:57: They just treat the vendors brand name as their entire risk strategy

00:02:00: which is such a fundamental misunderstanding of liability.

00:02:03: I mean Microsoft might provide the infrastructure sure but they absolutely do not assume The operational risk of your specific deployment exactly like if your team uses co-pilot to summarize highly sensitive client financial data or say, to draft a legally binding contract and it hallucinates at critical term.

00:02:23: Microsoft isn't the one getting sued!

00:02:25: No definitely not.

00:02:26: The Deployer always bears their responsibility...always.

00:02:29: But how does an enterprise actually wrap its arms around that?

00:02:33: Because if buying is secure tool isn't enough.

00:02:36: what Does actual governance look like on a day-to-day basis?

00:02:39: Well,

00:02:39: Gabriel Millian and Bisha Kubica provided a really clear architectural view of this on LinkedIn.

00:02:45: They brought up what they call the six layer AI stack And they argue that most enterprises approach this entirely backward Like leadership usually starts at the very top layer which is compliance in audit.

00:02:55: Oh right!

00:02:56: They draft massive PDF policy document about ethical A.I use Yeah Just you know consider the job done

00:03:02: Exactly, but I mean...I imagine a static PDF doesn't do much good if you don't even know what software your employees are actually running on their laptops.

00:03:10: Yeah that is the core issue

00:03:12: right?

00:03:12: Yeah yeah Million and Kubica point out that You cannot audit What you haven't inventoried.

00:03:17: so The foundational layer of That stack has to be an AI inventory.

00:03:21: So mapping everything out.

00:03:22: Right, you have to actively scan for shadow AI employees expensing random AI meeting summarizers or plugging proprietary code into some free web browser prompt

00:03:35: which happens all the time.

00:03:37: and if You haven't mapped those systems tag them with an owner And assigned a risk tier.

00:03:42: your compliance policy is basically just corporate fiction.

00:03:45: Okay Let's unpack this for second because I want to push back on that a bit sure in inventory Sounds like a massive constantly outdated spreadsheet.

00:03:54: Like is the goal really just to catalog every single tool?

00:03:59: because that seems like administrative bloat.

00:04:01: I mean, it's like trying to buy roof insurance for our house before you've even poured The foundation or counted how many rooms?

00:04:05: You have.

00:04:06: That's a great analogy.

00:04:07: But it's not actually about the tool itself It's but the application of the tool.

00:04:11: okay

00:04:11: also.

00:04:12: well Abhijir Rudra's analysis of the EU AI act makes this distinction incredibly urgent.

00:04:18: August two, twenty-twenty six is the enforcement deadline for high risk systems under the act.

00:04:23: Right

00:04:23: and the penalties are severe.

00:04:26: we're talking up to thirty five million euros or seven percent of global turnover.

00:04:31: wait really?

00:04:32: Seven percent of Global Turnover that is.

00:04:34: I mean That Is The Kind Of Fine that alters a company's trajectory For A Decade.

00:04:39: Absolutely

00:04:40: It'S Massive.

00:04:41: And Rudra Emphasizes A nuance That Answers Your Question About Administrative Bloat.

00:04:45: The EU AI Act doesn't classify risk based on the technology itself, you know?

00:04:50: It classifies risks based on use case.

00:04:52: Oh!

00:04:52: The use-case got it

00:04:53: Right.

00:04:53: So if your marketing team uses a large language model to draft the social media posts, that's minimal risk.

00:04:58: no big deal sure.

00:04:59: but If you're HR department uses that exact same underlying model To screen resumes and like filter out job candidates That instantly becomes A high-risk system subject to intense regulatory scrutiny.

00:05:10: Ah

00:05:10: I see.

00:05:10: so The inventory isn't just a list of vendor names.

00:05:13: It is map Of how technologies intersecting with human lives And actual business outcomes

00:05:18: Precisely Which leads How executives actually need monitor this stuff.

00:05:22: Tiffany Massen shared a brilliant framework for this.

00:05:26: What does she suggest?

00:05:27: She argues that real governance doesn't live in a binder on a shelf, it lives in a dashboard.

00:05:33: A true board packet for AI shouldn't just show cost and usage metrics.

00:05:37: It needs to show incident reports And most importantly, it needs the override rate.

00:05:42: The override rate meaning how often a human employee actually looks at the AI's output and says no thats wrong.

00:05:48: I'm changing.

00:05:49: Yes, exactly.

00:05:51: Think about the mechanics of a human in-the-loop policy.

00:05:54: if A company claims a human reviews every single AI decision but The override rate on the dashboard is literally zero percent What does that tell you?

00:06:03: It tells you the human isn't actually reviewing anything at all.

00:06:06: They are just blindly clicking approve because of automation bias.

00:06:09: Exactly You nail it.

00:06:11: or conversely If the override rate Is like ninety percent your ai is essentially useless and actively slowing your team

00:06:19: down.

00:06:20: That's a really good point, right?

00:06:21: So that one metric the override rate tells The board if their governance is actually functioning in reality or If it's just theater.

00:06:29: Okay so let's say you build that perfect six-layer stack.

00:06:32: You have your inventory you monitor use cases Your board is actively watching the override rates.

00:06:38: You feel completely in control.

00:06:39: Yeah, he

00:06:39: feels safe.

00:06:40: But

00:06:40: then the infrastructure itself.

00:06:42: this gets ripped away because here's where it gets really interesting.

00:06:45: You can't govern a system if the rug gets pulled out from under you.

00:06:48: Oh, yeah This is where we hit the geopolitical reality of AI.

00:06:53: We saw a massive shockwave in this space recently Detailed in post by Joseph Nogle and Christoph Skarten.

00:06:59: The

00:06:59: anthropic situation right?

00:07:00: Yeah Anthropic launched their fable five and mythos Five models.

00:07:04: these were highly anticipated incredibly capable systems And within three days literally three days the US Commerce Department ordered to block on all foreign nationals using those models.

00:07:14: Three days, that is just staggering.

00:07:17: They cite a national security concerns around potential safety jail breaks so the models had to be taken offline?

00:07:23: Let's just pause and think about enterprise implications of that.

00:07:26: for second You don't even have take political stance to see massive business risk here.

00:07:31: Not at all.

00:07:31: Like if you are company operating internationally And spent six months integrating that specific model into your core customer service workflow... ...you're suddenly broken overnight

00:07:44: And not because you missed a server payment or had a technical outage, You are broken because of government decree issued thousands miles away.

00:07:52: Exactly

00:07:53: Mark Schmidt framed this perfectly on LinkedIn.

00:07:56: He said AI is no longer just software tool It's strategic infrastructure

00:08:01: Which means treating an AI vendor like a standard SAWS vendor Is incredibly dangerous.

00:08:05: Like if your CRM goes down it's a headache.

00:08:08: But the cognitive engine processing entire supply chain.

00:08:11: logistics gets geo blocked overnight.

00:08:13: Your business stops

00:08:14: completely stops and Schmidt argues this requires a total shift in architecture.

00:08:19: Companies can no longer afford these single vendor dependencies, right?

00:08:23: You have to design model portfolios.

00:08:25: you need an architecture that allows you to hot swap from an anthropic model To a meta-model or maybe an open source alternative without rewriting your entire application.

00:08:37: But building that kind of portability and routing all those API calls to different massive language models, That brings up a terrifying financial reality doesn't it?

00:08:47: Oh absolutely.

00:08:48: The cost side is wild!

00:08:51: Damien Kopp did some analysis on the unit economics of this industry right now, and then numbers are completely unsustainable.

00:08:57: He cited margins for AI providers on inference at around negative ninety-four percent.

00:09:02: Negative ninety four?

00:09:03: Yeah we need to define inference here because it's the hidden tax of the entire AI boom.

00:09:07: go forward.

00:09:08: so training a model is like sending it to medical school.

00:09:11: It costs hundreds of millions of dollars up front.

00:09:13: but inference the model actually practicing medicine.

00:09:17: Every time you type a prompt, servers have to crunch the math to give an answer

00:09:21: Right?

00:09:21: Right now!

00:09:22: The cost of the compute required to generate that answer is vastly higher than what providers are charging for it.

00:09:28: So

00:09:28: its totally artificial ecosystem.

00:09:30: Providers take massive losses on every API call just lock in market share and starve out competitors.

00:09:37: It's classic capital funded land grab.

00:09:40: But those venture capital subsidies won't last forever.

00:09:43: No We are basically taking a five dollar Uber to the airport right now, like back in twenty fourteen.

00:09:49: Yes

00:09:49: great comparison!

00:09:50: It feels magic at that time but meter is about break and real bill coming.

00:09:56: Once subsidies dry up The true cost of token based API calls will hit enterprise balance sheets

00:10:03: which brings us to Borigur's analysis.

00:10:06: He argues that because of these looming token costs, local models and sovereign infrastructure are basically becoming mandatory for enterprises.

00:10:14: To control their own unit economics.

00:10:15: you mean?

00:10:16: Exactly!

00:10:17: If your AI usage scales from say ten employees testing it out to ten thousand employees running automated workflows all day long... Your inference bill goes exponential.

00:10:27: You eventually have to stop renting intelligence by the token And start owning the inference on your own hardware.

00:10:32: Okay so To escape these exponential API costs and to handle these complex multi-step workflows, companies are trying to automate the human out of loop entirely.

00:10:43: Right moving to agents?

00:10:44: Yeah they're moving away from chatbots that just answer questions And they are deploying autonomous AI agents That actually take actions on their behalf.

00:10:53: But that introduces some terrifying new attack surfaces.

00:10:55: It

00:10:55: is a massive paradigm shift.

00:10:57: You are moving from an advisor to an actor and it completely rewrites The enterprise security model.

00:11:02: Alexander Torlo had such a brilliant chilling way of explaining the mechanics of this shift.

00:11:07: Oh, The Key's analogy?

00:11:08: Yes He used the analogy of giving someone the keys to a building.

00:11:12: If you give a human employee secure access to a corporate system They walk to their desk they log in these one key they navigate To there specific dashboard they do Their job and they log out.

00:11:22: right.

00:11:22: but an agent doesn't behave like A human at all

00:11:25: not even close.

00:11:26: hmm

00:11:26: If you give an AI agent access, it instantly uses all of it.

00:11:30: It probes.

00:11:31: every single directory file and API endpoint that can reach for a human permission is the ceiling.

00:11:41: It will utilize every ounce of access it is granted instantly without any Of the natural hesitation that usually acts as a buffer in human workflows.

00:11:51: and not lack of hesitation Is exactly what turns theoretical risks into active exploits.

00:11:56: rock Lamber has discussed The newly released oas Jenny I security.

00:11:59: twenty-twenty six report, And the findings are deeply concerning.

00:12:02: What did the reports say?

00:12:03: Well last year the Security community was talking about agentic threats mostly as thought experiments but today Those theoreticals are documented.

00:12:11: CVE's.

00:12:13: Yeah, common vulnerabilities and exposures.

00:12:16: And the ultimate irony here The traditional security controls we trusted like allow lists Are actually becoming the attack surface.

00:12:24: Wait how does a Security Control Facility an Attack?

00:12:27: That seems entirely counterintuitive.

00:12:29: I know it is crazy.

00:12:30: Let us look at an Allow List.

00:12:31: Say you deploy in AI coding agent To keep it safe You configure an allow list so it can only execute commands using approved internal tools.

00:12:40: Okay,

00:12:40: make sense!

00:12:41: Now imagine an attacker hides a malicious instruction –a prompt injection inside of database that the agent is supposed to summarize– The agent reads hidden instructions believes its legitimate command from user and executes

00:12:53: them.

00:12:54: But wouldn't they allow us block it?

00:12:56: No because the agent uses the approved internal tool.

00:13:01: The security perimeter just auto-approves the exploit because agent itself is asking for permission.

00:13:06: Oh wow, so it's like a Trojan horse?

00:13:08: Exactly!

00:13:09: This means that at application layer safety which makes sure model doesn't do bad things and security keeping bad actors out are now exactly same job.

00:13:18: That is terrifying.

00:13:20: So if agents are running wild like this and they're essentially weaponizing their own access, do we just need to give them more context?

00:13:26: To make the smarter?

00:13:27: yeah If you can process a million tokens of data at once Shouldn't it be smart enough to recognize a hidden prompt injection?

00:13:36: You would assume that right but more data equals better reasoning.

00:13:40: But Sarah Soleimani provided some fascinating insight into why that completely fails.

00:13:46: She warns against a phenomenon she calls context pollution.

00:13:50: Stuffing a one million token window with data actually makes the agent worse, not better.

00:13:55: Wait really?

00:13:56: What does context pollution look like under the hood?

00:13:58: Think of it this way.

00:13:59: imagine asking human to solve complex math problem but you make them do while sitting at desk covered in ten thousand post-it notes containing irrelevant emails, debug logs and half finished

00:14:11: thoughts.

00:14:12: Adding another fifty thousand post-it notes doesn't make them smarter it paralyzes them.

00:14:17: So when you give an AI agent a massive million token context window It isn't gaining clarity at all

00:14:23: No!

00:14:23: It gets distracted by its own garbage logs.

00:14:26: Every time the agent uses a tool reads file or makes logical step... ...It dumps that text back into its own prompt Very quickly, it just gets lost in its own noise.

00:14:35: A crucial security rule buried fifty thousand tokens deep simply gets ignored.

00:14:41: The model effectively develops amnesia.

00:14:43: So how do you fix that?

00:14:44: If you need the agent to perform a complex multi-step task what's the solution?

00:14:50: Soleimani points out that the fix is architectural.

00:14:53: You don't build one massive agent with a giant context window... ...you build smaller focused subagents.

00:14:59: Oh I see.

00:14:59: And then you have a lean orchestrator that delegates.

00:15:02: You give a sub-agent, a clean empty context window.

00:15:05: It performs one specific task writes its findings to a database and only reports two sentence summary back the main agent.

00:15:12: A Clean Context will outperform a massive context every single

00:15:15: time.

00:15:16: That makes alot of sense.

00:15:17: So let's callback connect the dots here.

00:15:19: Look at whole board.

00:15:20: We have governance gaps and structures built upside down.

00:15:23: We have model dependency risks, an inference cost waiting to explode And we have wild autonomous agents drowning in context pollution... ...and turning security allow lists into attack vectors.

00:15:34: It's a lot!

00:15:36: So with all this tech deployed why aren't businesses actually moving faster?

00:15:40: Which brings us full circle of the question you asked at very beginning

00:15:43: Why are nine and ten executives telling the National Bureau of Economic Research that they have seen zero productivity gains from AI over the past three years?

00:15:53: Simon Taylor shared that report, And it is a harsh reality check.

00:15:57: So what's the real issue?

00:15:58: Olga Patapsiva perfectly diagnosed What's actually happening inside these companies.

00:16:03: She noted that organizations frequently have an eight-week approval queue for an eight second AI task.

00:16:08: Oh my god, yes!

00:16:09: Wait time isn't a technology problem it's an operating model problem.

00:16:12: Exactly let's visualize that An employee uses AI to instantly draft a complex compliance report.

00:16:19: The technology reduced the three day writing process down to eight seconds.

00:16:23: Amazing right?

00:16:24: Incredible.

00:16:25: But then That report has be reviewed by manager Sent to legal Checked by compliance And signed off by director.

00:16:32: All of those people are still operating at human speed, bottlenecked by legacy meeting schedules and inbox fatigue.

00:16:39: The AI eliminated the processing time but the wait-time remained exactly the same.

00:16:44: we're just automating broken processes.

00:16:46: you're just rushing to hurry up in weight

00:16:48: precisely if your operating model is fundamentally broken.

00:16:52: dropping an incredibly fast engine into it doesn't make the car go faster.

00:16:56: It just tears a transmission apart

00:16:59: And that dysfunction is having a profound psychological impact on how we work, too.

00:17:03: I want to bring in Diego Seroa and Steve Suarez's perspective on this.

00:17:07: Yeah their post was fascinating.

00:17:08: They argue that AI Is essentially making us incredibly fast at avoiding the spaces where actual thinking occurs Like We are terrified of The blank page

00:17:17: You really are?

00:17:18: The moment we face A difficult strategic question... ...we reach for a prompt window To generate a framework or draft.

00:17:25: We use the tool to fill the silence before we've even formulated an original thought.

00:17:29: But Soroa and Suarez say that their real competitive advantage for a company isn't figuring out what to automate next,

00:17:35: right?

00:17:36: The Real Competitive Advantage lies in identifying your golden links...

00:17:40: The Golden Links!

00:17:40: I love that term those rare non-linear moments into business process where human judgments simply must override the algorithm.

00:17:47: Exactly

00:17:48: it's the pause It is looking at perfectly generated data set having intuition.

00:17:54: The math is right, but the market context feels wrong.

00:17:57: We are going to pivot.

00:17:58: Yeah If you use AI to instantly bridge every single gap in your thinking You lose the friction that creates real insight.

00:18:06: And if we fail to protect those golden links?

00:18:17: You have to remember the fundamental law of large language models.

00:18:20: They are statistical prediction engines trained on massive data sets, so by definition they average everything out and produce what statistically looks like an acceptable median response.

00:18:31: Right!

00:18:31: So if you're a entry-level employee who struggles with writing... The AI lifts your work up into professional baseline

00:18:38: Sure But What If You Are A Senior Expert?

00:18:42: What if your value to the company is your unique idiosyncratic way of analyzing a problem?

00:18:47: Then it hurts you.

00:18:48: Exactly!

00:18:49: If you rely on AI to draft your memos or design your strategies, The model will actually pull your work down to the median.

00:18:56: It smooths out your edges...it forces your unique expertise into the exact same statistical middle as everyone else

00:19:03: Because AI averages everything out Everything's starting look and sound the same.

00:19:09: You have two rival companies both using the same foundational models, prompting them with the same industry data generating the exact same perfectly-average pitch decks.

00:19:17: Yep!

00:19:18: If we aren't careful... We're just using a golden shovel to dig up perfectly average

00:19:22: grave.".

00:19:22: And Alasenko highlights hidden consequence of this perfectly averaged grave — A massive epidemic of imposter syndrome.

00:19:28: Oh that makes so much sense

00:19:30: Right?

00:19:30: imagine an employee who sounds incredibly articulate and data driven on Slack or over email entirely armed by AI generated insights.

00:19:38: But then they are asked to negotiate a live deal, or have to whiteboard the problem off-the-cuff in meeting room and AI safety net is just gone.

00:19:46: The gap between their synthetic perfectly average online persona and their actual unaugmented capability creates massive anxiety.

00:19:54: We're building corporate culture where everyone sounds brilliant on paper but no one knows how actually think through when wifi goes down.

00:20:03: It's an arms race of noise.

00:20:05: We are generating more synthetic content just to summarize the synthetic content we receive from someone else, and the only way out of it is deliberately build friction back into our days.

00:20:14: To protect the override rates secure golden links.

00:20:17: stop confusing speed with actual progress.

00:20:20: Exactly Well, if you enjoy this episode new episodes drop every two weeks.

00:20:25: Also check out our other editions on ICT and Tech digital products & services cloud sustainability in green ICT defense tech and health tech.

00:20:32: Thank You so much for diving deep with us today And don't forget to subscribe!

00:20:44: what if your company's most valuable asset isn't your tech stack at all?

00:20:48: What if it is actually leaderships willingness to sit in a quiet room, unplug from the prompt window and think.