Best of LinkedIn: Artificial Intelligence CW 49/ 50

00:00:00: This deep dive is provided by Thomas Allgaier and Frennus, based on the most relevant LinkedIn posts about artificial intelligence in calendar weeks, forty nine and fifty.

00:00:09: Frennus supports enterprises with market and competitive intelligence, decoding emerging technologies, customer insights, regulatory shifts and competitor strategies.

00:00:19: So product teams and strategy leaders don't just react, but shape the future of AI.

00:00:25: Welcome to the deep dive.

00:00:28: So this week, we're digging into a huge pile of insights from LinkedIn, really looking at the top artificial intelligence trends that we're defining the end of the year.

00:00:37: Yeah, and if you're expecting another chat about the latest model, writing better poetry, it's not what's happening.

00:00:43: The whole conversation is shifted.

00:00:44: It's no longer about the novelty of it all.

00:00:46: It's now about control.

00:00:47: It's about execution and building actual systems.

00:00:50: We're seeing leaders, engineers, regulators, they're all grappling with how to govern, orchestrate, and you know, critically, how to secure AI when it leaves the lab.

00:01:00: So it's moved from what if to how to.

00:01:02: Exactly.

00:01:03: The focus is one hundred percent on measurable value now.

00:01:06: OK, so let's unpack all that.

00:01:07: Our mission here is to pull out the most important nuggets for you if you're a professional in the ICT and tech industry.

00:01:13: And we're going to cover everything from why these huge pilots are failing to which global regulations you absolutely cannot ignore.

00:01:21: And that failure point, I mean, that's where we have to start.

00:01:23: The data is pretty sobering.

00:01:26: It really is.

00:01:27: Let's get into it.

00:01:28: The hard truth about adoption.

00:01:31: We're calling this first theme AI strategy and the execution gap.

00:01:36: The big consensus we're seeing is that AI underperformance, it isn't really a model quality problem.

00:01:43: It's a failure of governance and just, well, weak execution.

00:01:47: It's even more than that.

00:01:48: It's actually a retreat.

00:01:49: John Nordmark shared this truly surprising data point.

00:01:52: I

00:01:52: saw that one.

00:01:53: AI adoption in big companies is actually declining, not rising, not even plateauing.

00:01:57: They tried it and now they're pulling back.

00:01:59: Declining, so not just slowing down, but actually going backwards.

00:02:02: That completely flies in the face of the industry narrative.

00:02:05: It does.

00:02:05: And Stephanie Gradwell frames the why perfectly.

00:02:09: She says most leadership teams don't really have an AI problem.

00:02:12: OK.

00:02:12: They have an execution problem.

00:02:14: The models, they work beautifully in the lab, you know, and a proof of concept.

00:02:17: In a clean room.

00:02:18: Exactly.

00:02:19: But then you drop it into the messy reality of business workflows and legacy systems.

00:02:25: and the core business KPIs, they don't change.

00:02:28: They just can't seem to bridge that gap from a successful pilot to actual enterprise value.

00:02:34: So the challenge isn't building the shiny new tool.

00:02:38: It's redesigning the entire factory floor to actually use it.

00:02:41: That's the perfect analogy.

00:02:43: And the solution we're seeing mentioned everywhere is a return to basics.

00:02:46: Start small, stay focused, iterate.

00:02:48: Right.

00:02:49: Kaushik Bay suggests this very practical four-step approach.

00:02:52: First, you rigorously map your workflow pain points.

00:02:55: Okay, standard stuff.

00:02:56: Right.

00:02:56: But then second, you pick a high-frequency low-risk task, something that won't, you know, crash the business if it fails.

00:03:01: Third, you automate it with a controlled agent.

00:03:04: And then finally, you measure the impact for a very strict thirty days.

00:03:07: Wait, but how is that different from what we were doing five years ago?

00:03:10: You know, mapping pain points, running thirty-day tests.

00:03:14: That's a great question.

00:03:16: The difference is the type of automation.

00:03:18: That this isn't just scripting a task.

00:03:20: It's about introducing a controlled agent that can operate with some autonomy.

00:03:24: And if you get that quick, quantifiable ROI win, that's your first.

00:03:30: internal success story.

00:03:32: And that's what unlocks the budget.

00:03:33: That's what unlocks the budget and the will for a bigger transformation, which brings us all back to leadership.

00:03:39: Neil Harrison reflected on this.

00:03:41: He said the failure in twenty twenty-five adoption wasn't a lack of ambition.

00:03:45: It was a lack of leadership capability.

00:03:47: Exactly.

00:03:48: Real

00:03:48: success means redesigning the work itself for human AI collaboration, not just trying to bolt a new tool into a broken process and that change management, that's the hard part.

00:03:58: That idea of integrated systems, of systems that collaborate, that's a perfect segue into our next big theme.

00:04:03: Absolutely.

00:04:04: The focus has really moved away from just a single model's performance and toward what everyone's calling the agentic shift and orchestration.

00:04:11: Okay, the word agent is everywhere right now.

00:04:15: What does that actually mean for someone in tech?

00:04:17: Well, Breesh Gashore Pandey points out that the foundational models, the LLMs, they're becoming commodities.

00:04:23: They're interchangeable.

00:04:24: Right.

00:04:24: The real value, and so the real engineering focus, has shifted to what he calls agentic stacks.

00:04:31: These are the system layers that handle things like memory, complex execution, and the orchestration that controls the whole process.

00:04:39: This is where it gets really interesting for me.

00:04:41: Nandan Malakara defines what agents do differently.

00:04:45: They plan, so they break big tasks into smaller steps.

00:04:49: They coordinate systems, they learn from feedback, and this is the key part, they intelligently invoke other tools to act.

00:04:56: They're not just tools, they're intelligent partners.

00:04:58: And this isn't just hype, it's delivering real economic results.

00:05:01: Richard Turen reported on an anthropic survey.

00:05:04: It found eighty percent of technical leaders are already reporting measurable economic returns from agents.

00:05:09: Eighty percent, that's huge.

00:05:11: It is.

00:05:12: The most mature area, no surprise, is software development.

00:05:15: Adoption there is at ninety percent.

00:05:18: We're seeing examples like Neobank N-twenty-six achieving seventy percent automation in certain processes.

00:05:24: Okay, eighty percent measurable return.

00:05:26: Are we talking about small pilot projects here?

00:05:28: or is this scalable, sustained ROI?

00:05:31: That is the big question.

00:05:33: And the industry's answer is to build the infrastructure to make it scalable.

00:05:37: That's where orchestration comes in.

00:05:39: You can't just have a fleet of autonomous agents running wild.

00:05:42: You need a control plane.

00:05:43: You need a control plane.

00:05:45: Vishal Mishra announced Microsoft's agent three sixty five framing it exactly as that the new enterprise control plane

00:05:52: control plane.

00:05:53: so the foundational models are becoming interchangeable.

00:05:55: isn't creating a Microsoft controlled highway just trading one kind of vendor lock-in for another?

00:06:00: that's

00:06:00: the strategic risk for sure but it shows how urgent the need is for a system of systems they call it the highway traffic control and safety system for all your AI agents.

00:06:10: the point is leaders know they need centralized management to scale this safely

00:06:14: because that creates a massive amount of risk.

00:06:16: Massive risk.

00:06:18: Rob Vanderveer flagged cross user prompt injection as the single biggest threat in agentic AI right now.

00:06:24: So what is that exactly?

00:06:26: Imagine

00:06:26: you have a privileged automated agent scanning say a public forum.

00:06:32: An attacker can sneak a malicious instruction, the prompt injection, into a totally normal looking comment.

00:06:38: And because the agent is designed to follow instructions.

00:06:40: It follows the malicious one.

00:06:42: That instruction could tell it to leak sensitive data, delete files, change configurations, anything.

00:06:48: Wow.

00:06:48: So the recommended defense is pretty harsh, but it's necessary.

00:06:52: You have to limit an agent's access to untrusted data dynamically.

00:06:56: You have to treat all agents as potential bad actors and operate with zero trust.

00:07:01: That security concern flows right into the regulatory landscape, which is trying to build a cage around all this

00:07:06: power.

00:07:06: Let's turn to that.

00:07:07: governance, trust, and the regulatory reality.

00:07:11: The EU AI Act is really anchoring this whole global debate.

00:07:14: Yeah, but there's a lot of, let's say, paranoia out there.

00:07:18: Before teams panic and start huge compliance projects, Alex Sureski gives some simple advice.

00:07:23: Just as the first question, have you checked whether the AI Act even applies to your system at all?

00:07:28: So many teams just slap AI enabled on a marketing page and create their own regulatory headache.

00:07:34: Right.

00:07:34: You have to look at the actual definitions.

00:07:36: Kimberly Friesen shared some key indicators for figuring this out.

00:07:40: A regulated system has to be machine-based, have a certain level of autonomy, and that threshold can be pretty low.

00:07:46: Okay.

00:07:47: And most importantly, it has to have a real influence on a physical or virtual environment.

00:07:53: A lot of simple prediction tools just fall outside of that scope.

00:07:56: But if your system does fall into that high-risk category, things like credit scoring, underwriting, hiring tools, what are the immediate requirements?

00:08:04: Hegbert Weatherborne lists them out.

00:08:05: You need rigorous risk management, demonstrable data quality controls, detailed technical documentation, mandated human oversight, and verifiable accuracy.

00:08:14: It's non-negotiable.

00:08:15: And the penalties are serious.

00:08:17: We're talking up to fifteen million or three percent of your global annual revenue.

00:08:21: So what's the blueprint for actually doing all that?

00:08:24: Gurdiep Singh Chopra positions ISO forty two thousand one as that blueprint.

00:08:30: It's the international standard for an AI management system.

00:08:33: It basically operationalizes the law.

00:08:35: It tells you how to systematically meet those legal obligations.

00:08:38: And for all of us who rely on third-party vendors for this stuff, what do we need to be demanding from them?

00:08:44: Patricia Bertini says you have to proactively demand responsible AI documentation.

00:08:50: That means getting training data summaries, a clear copyright compliance policy, and model documentation forms.

00:08:56: If a vendor can't give you that.

00:08:58: It's a huge red flag.

00:08:59: It's a huge red flag.

00:09:00: And it means the legal exposure falls squarely back on you, the client.

00:09:03: What's fascinating is the global fragmentation here.

00:09:06: Lexi Reese discussed the new one rule AI executive order in the U.S.

00:09:10: from the Trump administration.

00:09:11: It favors voluntary standards.

00:09:13: A completely different philosophy from the EU.

00:09:15: Exactly.

00:09:16: And that creates a massive strategic headache for global companies.

00:09:19: The reality is you have to build your systems to the strictest standard, which is the EU AI Act, and then kind of work backwards from there.

00:09:25: So the EU is setting the global bar effectively.

00:09:29: They

00:09:29: are.

00:09:29: And then you have countries like Italy.

00:09:31: According to Bethesimo Novec, going even further, they're passing their own national laws to criminalize harmful deep fakes, which goes beyond even the EU framework.

00:09:41: It's a fractured landscape.

00:09:43: It really is.

00:09:44: And all that complexity hits the engineering teams directly.

00:09:48: Which brings us to our next theme.

00:09:50: Right.

00:09:50: Engineering, IP, and the future of work.

00:09:53: The tools are changing so fast, the roles have to change with them.

00:09:57: Absolutely.

00:09:57: For developers, the shift is huge.

00:10:00: Dr.

00:10:00: Ingir Ganem described it as moving from the busy bee who does hands-on coding to the air traffic controller who oversees, reviews, architects, and orchestrates the whole system.

00:10:10: The value isn't in how many lines of code you type anymore.

00:10:13: And that's all being driven by AI coding assistance.

00:10:16: Hague Lupesco shared some stats.

00:10:17: Ninety percent adoption reported.

00:10:19: Developers shipping sixty percent more pull requests.

00:10:22: Sounds amazing.

00:10:23: It does, but here's the catch.

00:10:24: Quality metrics like change failure rate, they don't consistently improve.

00:10:28: So adoption doesn't equal impact.

00:10:30: Not automatically?

00:10:32: You need a structured rollout and proper enablement to make sure you're getting high quality alongside that high velocity.

00:10:38: Which means the old problems haven't gone away.

00:10:40: Nick Kay pointed out that classic vulnerability, SQL injection, cross-site scripting are still common.

00:10:46: And worse, the LLM assistance can actually amplify those bad patterns if the developer doesn't have the fundamental knowledge to spot insecure code in the output.

00:10:55: Wow.

00:10:56: Okay, and adding another layer of risk to all this is the IP battle brewing in Europe.

00:11:02: Yes, the details from this landmark ruling in Germany from the Munich regional court against open AI They're absolutely vital for anyone in this space.

00:11:11: So what does the German court actually rule?

00:11:14: It was a case brought by Jima, the German music rights organization.

00:11:17: The court found that the AI's memorization and then reproduction of copyrighted song lyrics was infringement.

00:11:23: And they rejected the usual defense.

00:11:25: They rejected the text and data mining defense.

00:11:28: The argument that training on copyrighted data is fair use.

00:11:31: It didn't hold up.

00:11:32: The court basically said if the model reproduces the material, it means it memorized it and that's infringement.

00:11:38: which creates a ton of legal uncertainty because that ruling goes directly against a recent decision in the UK on the same issue.

00:11:45: Exactly.

00:11:46: It fractures the legal landscape for generative AI across Europe.

00:11:50: Let's zoom out for a second to the macroeconomic impact.

00:11:53: Envato Ordax reported a staggering statistic.

00:11:56: AI infrastructure was responsible for ninety-two percent of US GDP growth in the first half of twenty twenty-five.

00:12:03: Ninety-two percent.

00:12:04: That's not just a tech story.

00:12:05: That's the entire macroeconomic growth engine.

00:12:08: It's immense.

00:12:09: But Zemina Ahmad warns that it's leading to what she calls jobless growth.

00:12:13: Productivity is skyrocketing.

00:12:15: But employment in IT and business services is actually declining.

00:12:20: Companies are quietly pursuing a strategy of more output with fewer people, and that's hitting junior staff and freelancers the hardest.

00:12:26: A really sobering perspective.

00:12:28: It connects that initial execution gap we talked about to this ultimate goal of massive, efficient, and potentially job-displacing growth.

00:12:34: We've covered

00:12:35: a lot of ground, so to summarize this whole deem dive, the AI battleground It's not about the models anymore.

00:12:41: It's about the systems, the governance, and the organizational change you need to actually get real value.

00:12:47: Right.

00:12:47: The focus now is all on control, security, and global compliance.

00:12:53: And if we think about the long-term human cost of this, there's an interesting point from Borogur.

00:12:58: What's that?

00:12:59: He raised the risk of Gen Z. of younger workers heavily relying on AI for critical social health and career decisions.

00:13:07: And for us, as the professionals building and deploying these systems we just said need to be treated with zero trust, that raises a really important question.

00:13:14: Which is?

00:13:15: What kind of organizational liability are we creating when we inherit a workforce that's potentially de-skilled in critical thinking and independent judgment?

00:13:24: That vulnerability created by a reliance on the very systems we're trying to lock down, that's something every leader needs to be thinking about urgently.

00:13:31: If you enjoy this deep dive, new episodes drop every two weeks.

00:13:34: Also check out our other editions on ICT and tech, digital products and services, cloud, sustainability and green ICT, defense tech and health tech.

00:13:43: Thank you for joining us for this deep dive.

00:13:45: Don't forget to subscribe.