Best of LinkedIn: Artificial Intelligence CW 03/ 04

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 O three and O four.

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.

00:00:24: It's going to be back.

00:00:25: And, you know, looking at the insights from the last couple of weeks in January, The whole tone just feels different.

00:00:32: It really does.

00:00:32: I mean, we're not in the playground anymore, are we?

00:00:34: Not at all.

00:00:35: If twenty-twenty-four and twenty-twenty-five were the wow years, you know, everyone gasping because a computer wrote a poem or something.

00:00:41: Yeah, the magic show.

00:00:42: Exactly.

00:00:43: Then early twenty-twenty-six feels like, I don't know, the year of the hangover, or maybe sobering up is a better way to put it.

00:00:49: I think sobering up is fair.

00:00:51: We're definitely seeing a narrative arc.

00:00:52: The hype cycle is settling and we're entering a phase of what I call operating discipline.

00:00:57: The conversation is just pivoted so hard.

00:00:59: It's not, look what this model can do anymore.

00:01:02: Now it's... How do we run this safely, legally, and profitably at scale?

00:01:07: And that profitable part seems to be the one everyone's really struggling with.

00:01:11: It's the shift from magic tricks to mechanics.

00:01:13: For sure.

00:01:14: So for this deep dive, we're going to navigate that shift through a few key lenses.

00:01:18: We've got to talk about governance, which I know sounds dry, but wait until you hear the numbers.

00:01:23: Those

00:01:23: numbers are big.

00:01:24: Then there's the ticking clock in Europe with the AI Act, and then the tech itself.

00:01:29: Because while the lawyers are panicking, the engineers are shifting from chatbots to full-blown agents.

00:01:36: To the AI operating system.

00:01:37: Yeah, it's a dense couple of weeks, but I think you're right.

00:01:39: We have to start with a topic that's probably ruining sleep for every CEO right now.

00:01:43: Governance, yeah.

00:01:44: Let's unpack this, because usually you say governance and people's eyes glaze over.

00:01:48: They

00:01:48: tune out, yeah.

00:01:49: But looking at a post from Julian Noble, that attitude is basically a death wish in twenty-twenty-six.

00:01:55: It is.

00:01:56: Noble really laid it on the line.

00:01:57: He's saying air compliance has mutated.

00:02:00: It's no longer just an IT ticket.

00:02:02: Right.

00:02:02: It's not, hey, IT, can you check the software anymore?

00:02:04: No.

00:02:05: It's now a critical enterprise risk.

00:02:07: And he's not just talking about the tech breaking.

00:02:09: He's talking about exposure.

00:02:11: hidden exposure.

00:02:12: That's the key part.

00:02:13: This is that shadow AI problem.

00:02:15: The recruitment tool HR bought that's biased the marketing sauce that hallucinates claims about your product.

00:02:21: The board isn't even tracking half of this stuff.

00:02:24: And Noble points to the big financial stick here.

00:02:26: Yeah.

00:02:26: With the EU AI Act in full force.

00:02:28: We're not talking about a slap on the wrist.

00:02:30: No.

00:02:30: We're talking penalties up to thirty five million euros or seven percent of global turnover.

00:02:36: Seven percent.

00:02:36: That is a resume generating event for a CEO.

00:02:39: That is pack your box and leave money.

00:02:41: It absolutely is.

00:02:43: But here's the nuance, and Joanna Miley had a great insight on this.

00:02:46: Just reacting to that fear by trying to comply might also be the wrong

00:02:50: move.

00:02:50: What do you mean?

00:02:51: Well, she argues that viewing this as just AI Act plus done is a massive strategic error.

00:02:58: She talks about a layered control stack.

00:02:59: A control stack.

00:03:01: That sounds like a cybersecurity principle, like defense in depth, but for AI.

00:03:05: That's exactly it.

00:03:06: Think about a bank today.

00:03:08: You're not just juggling the AI Act.

00:03:10: You've got GDPR, NIS-II, DeWara's, all at the same time.

00:03:15: It's a regulatory minefield.

00:03:17: It is.

00:03:17: And Mylar's point is, you have to pick a control posture.

00:03:22: Are you doing the bare minimum, or are you building control first?

00:03:26: designs.

00:03:26: Which sounds great in theory, but that creates friction.

00:03:29: I mean, I saw Shea Brown's post about procurement and it really hit home.

00:03:33: He's seeing deals just stall out.

00:03:34: The computer says no moment.

00:03:36: Exactly.

00:03:37: And it's not because the startups have bad tech.

00:03:39: It's because they walk into a meeting and procurement pulls out a fifty page questionnaire.

00:03:43: Yep.

00:03:44: And Brown says they're failing because they can't answer the boring questions.

00:03:46: Yeah.

00:03:47: Are you ISO forty two thousand one certified?

00:03:49: Have you tested for bias?

00:03:51: Do you train on customer data?

00:03:52: And

00:03:52: the founder is just staring blankly wanting to show off their cool demo.

00:03:56: And the deal just dies right there.

00:03:58: Brown's insight is that if you can't get past procurement in twenty twenty six, you don't have a business.

00:04:03: It doesn't matter how smart your model is.

00:04:05: This

00:04:05: just feels like a reckoning for the move fast and break things crowd.

00:04:09: And Paula Falcon gave it a name I love.

00:04:11: Architectural bankruptcy.

00:04:13: It's a stinging phrase, isn't it?

00:04:14: It

00:04:15: hurts because it's true.

00:04:16: It is.

00:04:17: Falcon says if you spent a ton of money on AI in twenty twenty five without a governance framework, you didn't buy innovation.

00:04:25: You've funded architectural bankruptcy.

00:04:26: Wow.

00:04:27: The failures we're seeing aren't because people chose the wrong LLM.

00:04:31: They failed because of bad enterprise architecture and unstructured data.

00:04:35: You can't build a skyscraper on a swamp.

00:04:37: You just get a faster way to make mistakes.

00:04:39: And a faster way to get fined, which brings us, I think, to the geography of those fines.

00:04:44: We have to talk about Europe.

00:04:45: All

00:04:45: right.

00:04:45: Theme two, the clock is ticking.

00:04:48: Matt Elsie flagged a specific deadline that everyone needs to have circled in red.

00:04:52: August, that is the deadline for full obligations on high-risk systems.

00:04:56: Which

00:04:56: is only what, six months away?

00:04:58: And we should clarify, high-risk catches a lot more than just, you know, medical robots.

00:05:02: Oh, much more.

00:05:03: If your AI agent influences hiring, sets credit limits, determines insurance premiums, or denies refunds.

00:05:09: So basically half the back office of any modern company.

00:05:13: Pretty much.

00:05:14: And Elsie points out you need eight core pillars of compliance ready.

00:05:18: But the hardest one, I think, is real human oversight.

00:05:22: And that doesn't just mean a person watching a screen.

00:05:24: No, not at all.

00:05:26: It means a human with the authority and the competence to hit the stop button.

00:05:30: If an algorithm denies a loan and a human just rubber stamps it without knowing why, that's not oversight.

00:05:36: That's non-compliance.

00:05:37: And that connects to this idea of sovereignty, right?

00:05:40: For years, it was just about data location.

00:05:42: Is it in Frankfurt or Virginia?

00:05:44: Yeah, that was the old definition.

00:05:45: Yeah.

00:05:45: Rob Thomas posted something that suggests that's totally outdated.

00:05:49: He argues sovereignty has evolved from geography to operation.

00:05:52: He calls it decision authority.

00:05:55: So it's about who's the boss?

00:05:56: Basically, who owns the outcome?

00:05:58: Sovereignty in twenty twenty six means proving where the inference happened and who was in control.

00:06:03: If you use a black box model and can't audit the logic, you don't have sovereignty.

00:06:07: You just have a European server.

00:06:08: That is a critical distinction.

00:06:11: It's accountability, not residency.

00:06:13: This must be a nightmare for regulated sectors.

00:06:16: I saw Eric Van Dyke talking about MedTech.

00:06:18: MedTech is the deep end of the pool here.

00:06:20: Van Dyke says Brussels is trying to align the AI Act with the existing medical device regulations.

00:06:26: But right now, it's a double burden.

00:06:28: So for them, AI is less... innovation lab.

00:06:31: And more compliance permit.

00:06:32: Exactly.

00:06:33: But here's the paradox.

00:06:35: While the legal teams are demanding human oversight, the tech itself is evolving to operate without it.

00:06:41: We're moving from chatbots to agents.

00:06:43: This is the exciting part.

00:06:45: And we're seeing a lot of buzz from people like Nandan Malikara about the AI operating system.

00:06:49: This is a huge conceptual shift.

00:06:51: Think about the old swivel chair workflow.

00:06:53: Oh, I know it will.

00:06:54: Open chat GPT, ask it to draft an email, copy.

00:06:57: swivel to Outlook, paste, swivel to SAP, download a CSV.

00:07:01: It was exhausting and just so inefficient.

00:07:04: Malikar argues the AIOS gets rid of that.

00:07:06: The human just says, get the QFOR report and email a summary to the board.

00:07:11: And the

00:07:11: AI traverses the ERP, pulls the CRM data, does the whole thing.

00:07:15: The AI is the orchestrator.

00:07:17: The enterprise apps become the back end.

00:07:19: Malikar appointed to Claude Coat, hitting a billion dollars in ARR faster than chat GBT as a signal.

00:07:27: The market is paying for execution, not just conversation.

00:07:30: Simon Taylor had a great analogy for this.

00:07:32: He called the modern agent a co-worker, not just a tool.

00:07:36: Right.

00:07:36: He used an example with Figma and Slack.

00:07:39: One prompt.

00:07:40: draft the Slack update, visualize the new timeline in Figma, and create the tasks in Asana.

00:07:46: One prompt, three different applications.

00:07:48: It understands the workflow context.

00:07:50: Exactly.

00:07:51: And that shift from generating text to executing actions changes how you build these things.

00:07:56: BriefCashorePandy had a strong post on this.

00:07:58: He says, twenty twenty six is all about building reliable systems.

00:08:02: So my library of perfect prompts isn't enough anymore.

00:08:04: It's not useless, but it's not enough.

00:08:07: Bandy says prompt engineering is being replaced by system engineering.

00:08:11: The stack has an orchestration layer, a memory layer, and crucially an evaluation layer.

00:08:16: Because you can't just eyeball it anymore, you need automated testing if an agent is sending emails on its own.

00:08:21: If you're just tweaking prompts in a playground in twenty twenty six.

00:08:24: You're already behind.

00:08:25: Okay, but where's the money?

00:08:26: Is this just cool tech or does it pay off?

00:08:29: I saw John Godel's breakdown for chief revenue officers.

00:08:31: Yes,

00:08:32: Godel frames this as an operating leverage tool.

00:08:35: Operating levers?

00:08:36: That's music to a CFO's ears.

00:08:38: It is.

00:08:39: Think about a deal desk.

00:08:40: Usually a huge bottleneck.

00:08:42: An agent can auto build that approval packet in seconds.

00:08:46: It reduces the cycle time of deals.

00:08:48: And if you close a deal three days faster, that's direct revenue impact.

00:08:51: That's cash flow.

00:08:52: Exactly.

00:08:53: But... And there's always a but.

00:08:55: There always is.

00:08:56: Is it actually working?

00:08:57: This brings us to the reality check part of this.

00:09:00: Let's play Devil's Advocate.

00:09:01: Hina Purohit posted about the productivity paradox.

00:09:04: The numbers she cited were

00:09:06: weird.

00:09:07: They're stark.

00:09:08: She cites data showing seventy-four percent of teams report productivity gains from AI.

00:09:13: They feel faster, but only eleven percent report measurable value.

00:09:18: Wait, what?

00:09:18: So people are churning out more stuff, more code, more slides, but it's not moving the needle on the business.

00:09:24: We might just be generating noise faster.

00:09:27: And Werner Heistek pointed out something even crazier about software development.

00:09:30: The

00:09:30: coding assistants?

00:09:31: I love those.

00:09:32: People

00:09:32: do, but Heistek points to cases where AI coding caused a nineteen percent slowdown.

00:09:39: A slowdown?

00:09:40: How is that even possible?

00:09:42: It writes the code instantly, but is it the right code?

00:09:45: The slowdown is the verify tax.

00:09:47: The human developer has to spend hours debugging and fixing the messy code the AI generated.

00:09:53: I feel that.

00:09:54: It's like being an editor instead of a writer.

00:09:55: Sometimes editing takes way longer.

00:09:57: So how do we fix

00:09:58: this?

00:09:59: Aman Ruiz suggests it's an organizational design problem.

00:10:02: He talks about the AI native squad.

00:10:04: Squad, not just one AI wizard on the team.

00:10:06: Exactly.

00:10:07: Ruiz argues that having one wizard doesn't help if everyone else is stuck in twenty twenty two.

00:10:11: The friction kills the speed.

00:10:13: The whole squad needs to level up together.

00:10:15: That makes total sense.

00:10:16: Culture eats AI for lunch, apparently.

00:10:18: Does.

00:10:19: And one last thing before we wrap.

00:10:21: The huge looming risk with these agents.

00:10:24: Security.

00:10:25: Specifically, prompt injection.

00:10:27: Bernard Maher flagged this.

00:10:28: And for a long time, I just thought that was a prank.

00:10:30: Tricking a chatbot into saying something

00:10:32: similar.

00:10:32: When it's just a chatbot, it is a prank.

00:10:35: But now connect it to an agent that can access your ERP, your bank account.

00:10:39: Right.

00:10:40: If I can trick the agent, I'm not getting a poem.

00:10:43: I'm transferring money or deleting the customer database.

00:10:46: Prompt injection becomes a critical vulnerability.

00:10:49: It's not a toy anymore.

00:10:50: It's a security breach waiting to happen.

00:10:52: So if we pull all this together... the wow phase is dead.

00:10:57: We need strict governance.

00:10:58: We need solid architecture.

00:11:00: And we're shifting to agents that can do real work, but that comes with a whole new set of risks.

00:11:06: And we have to do it all while navigating a productivity paradox where faster doesn't always mean better.

00:11:11: It sounds like twenty twenty six is the year of the grown-up conversation about AI.

00:11:16: I think that is the perfect way to put it.

00:11:17: Okay,

00:11:17: let's wrap this up.

00:11:19: If you enjoyed this episode, new episodes drop every two weeks.

00:11:22: Also check out our other editions on ICT and tech, digital products and services, cloud, sustainability and green ICT, defense tech and health

00:11:30: tech.

00:11:31: And if I can leave you with one final thought, we are moving from the wow phase to the how phase.

00:11:37: The winners in twenty twenty six won't be the ones with the best models.

00:11:41: The winners will be the ones with the best architecture and governance to run them safely.

00:11:45: Thanks for joining us on the deep dive.

00:11:46: Don't forget to subscribe.

00:11:48: See you next time.