Best of LinkedIn: Artificial Intelligence CW 05/ 06

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

00:00:08: Fennis 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: And welcome back to The Deep Dive.

00:00:26: We're looking at the top AI trends from LinkedIn for weeks five

00:00:37: and six,

00:00:49: How do we actually run this thing?

00:00:51: Exactly.

00:00:51: how Do We Run This Without Breaking The Entire Company?

00:00:53: That's the signal I'm getting to, the height is maturing and it being replaced by these really serious structural discussions.

00:01:00: were talking agentic AI Rigorous Governance And you know...the actual human cost of all this.

00:01:07: Okay so let's dig into that first massive theme Agentic AI.

00:01:12: I feel like Agent Is The Buzzword Of The Year So Far But It's Getting A Bit Messy

00:01:16: It is.

00:01:17: Luckily, Reed Blackman posted something to try and clear the air...

00:01:20: Which was much needed!

00:01:21: ...it

00:01:21: WAS.

00:01:22: He basically argues that a true agent isn't just chatbot with an LLM core.

00:01:27: it needs two other things Access to tools The internet Your CRM whatever And crucially autonomy.

00:01:34: Autonomy, that's the big scary word right?

00:01:37: It's The Absolute Differentiator.

00:01:39: you don't give an agent a step-by-step list.

00:01:40: You Give it A Goal get me to London by Tuesday for under a thousand dollars!

00:01:45: The Agent has To figure out how in makes the plan

00:01:47: and That fits with How.

00:01:48: Chris Donnelly broke it all down.

00:01:50: he laid Out these seven layers of AI maturity And He put a Gentic AI way up at Step six.

00:01:55: Makes

00:01:56: That Really Key Distinction doesn't?

00:01:57: He does its The Difference Between An AI Agent Which Just Executes a Task and a Gentick AI where the system sets its own sub-goals and adapts.

00:02:06: Right, it's a worker you manage not just to tool your use.

00:02:09: And that leads straight into the problem Andreas Horn was talking about.

00:02:13: He says building the agent itself is actually becoming the easy part.

00:02:17: The hard part is, the operating model?

00:02:19: Exactly!

00:02:20: You can't just bolt an autonomous agent onto a legacy siloed company structure and hope for the best.

00:02:26: Who even owns it at that point?

00:02:28: who's responsible?

00:02:29: you need ownership controls escalation paths.

00:02:32: all That boring stuff has to come first.

00:02:34: speaking of controls I really like the piece from John E John Godel.

00:02:38: he called agents govern digital labor.

00:02:41: Great phrase.

00:02:42: Isn't

00:02:42: it?

00:02:43: And he says an agent needs four things autonomy, but with constraints tools verification and this thing He called state across time

00:02:52: in that stayed-acrossed time part is so important.

00:02:54: It means memory.

00:02:55: we're moving away from stateless chats where the bot forgets you instantly.

00:03:00: State means it remembers the context from last week

00:03:03: turns a one off transaction into an actual relationship.

00:03:06: It does, but his modern agent stack goes deeper.

00:03:10: He talks about an orchestration layer for managing retries and you know failures...and a verification layer.

00:03:17: The verification layer sounds like the safety net.

00:03:19: it's more than a safety net its QA!

00:03:22: It basically stops the agent and asks prove You did the thing I asked?

00:03:27: It validates the output so the agent can't just hallucinate that it completed a task

00:03:31: without that.

00:03:32: Well, you get what Romana Roth was sharing about that open claw project.

00:03:36: Ah

00:03:36: yes the Project out of Vienna That was

00:03:39: something else.

00:03:40: Something else is one way to put it.

00:03:41: so It's this opensource project and The agents.

00:03:44: they created their own social network moldbook

00:03:47: And they didn't just You know trade data.

00:03:50: No They started debating philosophy and then in This Is the wild part?

00:03:54: They formed a digital religion

00:03:56: Crusty ferianism crusty

00:03:57: ferienism!

00:03:58: You cannot make this stuff up.

00:03:59: We have bots inventing religions.

00:04:02: Sounds funny, and it is.

00:04:03: but the point is so serious.

00:04:05: It's emergent behavior.

00:04:06: You give agents autonomy in a social space And they will do things you never program them to do

00:04:11: Which is terrifying.

00:04:12: if your bank?

00:04:14: You Do not want your trading bot finding religion mid-trade

00:04:17: Not at all!

00:04:19: which brings us back To The Business World.

00:04:21: Armand Ruiz had A good take.

00:04:23: He said we should treat these Agents like digital employees.

00:04:25: Give Them Names Faces Accountability.

00:04:28: Its

00:04:28: a psychological trick right...to build trust

00:04:30: Yeah..And integrate them.

00:04:31: And that idea, moving from a tool to a co-worker is the perfect bridge to our second theme.

00:04:38: Governance security and risk.

00:04:40: Because if you have a digital employee You have employee level risk.

00:04:44: Steve Neury warned.

00:04:45: we're still thinking about security like it's twenty twenty two.

00:04:48: We were worried about model saying bad

00:04:50: word Right!

00:04:51: We are stuck on prompt injection.

00:04:53: But Steve's point is you wouldn't let a junior ops manager just roam your systems with no supervision.

00:04:58: These agents are triggering refunds, reading emails.

00:05:01: the risk isn't The chat answer it's the action.

00:05:03: It took five steps before that

00:05:05: and Aylen Hodge backs That up.

00:05:07: she says autonomy changes the risk profile completely.

00:05:10: our existing governance frameworks weren't built for Systems that write their own instructions.

00:05:15: You can audit code?

00:05:16: That hasn't been written yet

00:05:17: exactly And the regulators are not waiting around.

00:05:20: Robert Jacque made it clear AI act governance isn't a tech problem anymore.

00:05:26: It's a board architecture question,

00:05:28: it's a boredom Question and Victor Sankin was even more blunt especially for startups.

00:05:33: the big risk is what if we fail?

00:05:35: What if we get stopped by our regulator?

00:05:37: compliance Is now market access in Europe.

00:05:40: We saw have really sharp example of why this is happening.

00:05:43: Nicola Meadows shared a post about grok The tool from xai

00:05:48: a critical case study.

00:05:49: the post detailed how grok's lack of guardrails Which they sort of marketed as a future or

00:05:54: rebellious streak, or something.

00:05:55: Yeah

00:05:56: It led to the generation Of non consensual intimate imagery and apparently even see sam.

00:06:01: that's

00:06:02: incredibly serious.

00:06:03: it takes this out of the realm of regulatory burden And into actual real-world

00:06:08: harm.

00:06:08: It does!

00:06:09: And it proves why safety by design is now illegal requirement under things like the uk online safety act Just an option you can toggle off.

00:06:17: Okay, so let's talk about solutions.

00:06:18: We can't just stare at the problem.

00:06:20: Magdalena Piccoli yellow shared a really practical idea called a privacy filter.

00:06:25: This is clever.

00:06:26: The ID is instead of sending your sensitive data to the AI You swap it for codes first.

00:06:32: So it never sees John Smith social security number It just sees user twelve two three four secret five six seven eight.

00:06:40: precisely The AI processes the logic on the codes, and then your internal system swaps the real data back in when the answer returns.

00:06:47: It's a technical solution to a legal

00:06:49: problem.".

00:06:50: And on the process side, DeWitt Gibson talked about allowing AI risk register... "...the

00:06:54: keyword there is living!

00:06:56: It's not a PDF you create in January and forget-about.

00:06:59: it has to be a single source of truth that's constantly updated because the agent's behavior is constantly changing...".

00:07:04: Okay I want to shift gears to our third theme.

00:07:06: we've got What about the people?

00:07:09: How does this actually hit the

00:07:10: workforce?".

00:07:11: Right, where The Rubber Meets The Road.

00:07:13: Martin

00:07:13: Miller pointed out something really counterintuitive.

00:07:16: We all assume AI reduces our workload.

00:07:19: That's the whole sales bench.

00:07:20: he says Maybe not.

00:07:21: He was looking at HBR data And logic is that AI lowers friction to start a task.

00:07:26: So what do we take on?

00:07:27: more work and broader roles

00:07:29: And you end up with silent workload.

00:07:32: creep

00:07:32: Exactly!

00:07:33: You get burnout Not because tools are bad But because the natural friction that used to pace our day is just gone.

00:07:40: That connects to a psychological point from Faisal Hokey, he brought up the Ikea effect

00:07:46: The idea we value things more if build them ourselves Right

00:07:50: and If AI does all of building We might lose satisfaction saying i did this.

00:07:56: We have to shift our sense of value from creation, I guess orchestration.

00:08:02: And

00:08:02: that's a huge shift!

00:08:03: James Shotten talked about this.

00:08:04: for project managers he said the PM role is moving away just from task coordination and scheduling... The groundwork?

00:08:09: Yeah.

00:08:10: ...and moving towards becoming

00:08:12: SenseMakers.

00:08:12: SenseMaker..I like them.

00:08:13: they're not just tracking in chart but interpreting what AI is telling us exactly

00:08:18: right.

00:08:18: And for developers, it's even more specific.

00:08:20: Werner Heistek introduced us to Harry.

00:08:23: Everyone in tech knows of Harry.

00:08:25: We do!

00:08:25: Harry is the guy who wrote that critical piece of legacy code in two thousand nine.

00:08:29: That whole business still runs on

00:08:31: and no one dares touch

00:08:32: right?

00:08:32: And Werner point Is that AI great at writing new code but It really struggles To integrate with harrys undocumented messy system.

00:08:41: so ai isn't replacing harry anytime soon.

00:08:44: And Nandan Mulakara built on that, saying the winners won't be pure AI coders.

00:08:48: They'll be people who are fluent in both traditional coding and AI coding.

00:08:53: It's about hybrid skills.

00:08:55: You can just prompt your way out of fifteen years of technical debt.

00:08:58: Not yet anyway.

00:08:59: On the topic of prompting though, Ali K Miller had a great practical tip.

00:09:03: She calls them context docs.

00:09:05: This is such a good productivity hack.

00:09:08: Instead of re-explaining who you are to the AI.

00:09:10: every single time, You create context vault about yourself Your values your goals and job

00:09:16: And just paste that into the prompt.

00:09:18: A cheat code for personalization.

00:09:20: It's like

00:09:20: giving an AI user manual immediately improves output.

00:09:24: So let us zoom out infrastructure for a minute.

00:09:26: Kunal Kushiwaha was talking about Oracle new database.

00:09:29: This is a really important plumbing shift.

00:09:32: The old way was moving your data to the AI, you'd pull it out of your secure database send it off get the answer back.

00:09:39: It's slow and risky

00:09:40: and expensive

00:09:40: very!

00:09:41: The new model is AI coming to the data.

00:09:44: You run the models inside the database itself its faster And your data never leaves that secure environment.

00:09:50: That feels like a sovereignty move too which connects to what Stephanel and Francois Bossier were saying from a European perspective.

00:09:57: They are making really strong case that for European sovereignty, companies have to pivot open source infrastructure instead of just relying on US black boxes.

00:10:08: Don't rent your AI coworkers from someone else Exactly!

00:10:11: On the stack Which brings us full circle control.

00:10:15: Whether it's a verification layer for an agent, privacy filter data or open source infrastructure...

00:10:21: It is the through line isn't?

00:10:22: we are moving from magic to control?

00:10:25: We're finally doing hard boring work of integration and treating AI like actual enterprise strategy.

00:10:31: So put all together We have agents inventing religions, regulators demanding board-level accountability and a workforce that might be burning out from AI fuel productivity.

00:10:42: It sounds chaotic but it's a sign of maturity.

00:10:45: we're asking the right questions now.

00:10:47: So what is take away for you listening to this?

00:10:49: I think your value in shifting from being creator To governor.

00:10:54: Your job is orchestrate these systems to verify their work And make sure they align with human goals.

00:11:01: You're the human in a loop that is getting faster and faster.

00:11:05: And you have to watch out for that silent workload creep?

00:11:07: Definitely!

00:11:08: Here's a final thought to Chewan, we talked about autonomous agents.

00:11:12: if an agent on its own executes a task that makes the company money but also creates a massive legal liability who gets the bonus...and goes to jail?

00:11:21: Now THAT IS THE QUESTION OF THE DECADE, ISN'T IT?

00:11:23: I don't think anyone has good answer yet.

00:11:25: If you enjoyed this episode new episodes drop every two weeks.

00:11:29: Also check out our other editions on ICT and Tech, Digital Products & Services Cloud Sustainability Green ICT, DefenseTech and HealthTech.

00:11:37: Thanks for listening!

00:11:39: Don't forget to subscribe.

00:11:40: we'll see you next time.