Latest / Best of LinkedIn: Digital Transformation & Tech / Best of LinkedIn: AI & Agentic Systems CW 29/ 30

Transcript

Best of LinkedIn: AI & Agentic Systems CW 29/ 30 · Best of LinkedIn: Digital Transformation & Tech

• 0:00
  Speaker 0: This episode is provided by Thomas Allgaier and Frenneth, based on the most relevant LinkedIn posts about AI and agentic systems from CW-Twenty-Nine and Thirty.
• 0:08
  Speaker 0: Frenness supports high CT 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 provider's competitively.
• 0:22
  Speaker 0: you can find more info in description.
• 0:24
  Speaker 0: So, um imagine firing your absolute most reliable employee and replacing them with a brand new piece of software.
• 0:30
  Speaker 0: And then realizing like three days later that the software just hallucinated to policy exception an autonomously signed illegally binding contract that costs you company millions.
• 0:39
  Speaker 1: I mean That's it right?
• 0:40
  Speaker 1: That is the ultimate modern enterprise nightmare Because for decades we've uh We've treated software like a power drill.
• 0:48
  Speaker 1: Right yeah You pick it up you aim at Pull the trigger.
• 0:50
  Speaker 1: The tool does that.
• 0:51
  Speaker 1: heavy lifting, sure but human hands are providing that continuous supervision.
• 0:55
  Speaker 1: But now we're basically handing that drill a blueprint and just leaving the room...the tool is getting up unplugging itself walking over to the project and you know deciding which walls need holes
• 1:06
  Speaker 0: Which is terrifying.
• 1:07
  Speaker 0: And that brings us to the core of today's deep dive.
• 1:11
  Speaker 0: We are unpacking the top AI and agentic systems trends That are driving the most critical conversations across LinkedIn right now.
• 1:19
  Speaker 0: Our mission here for you listening is to cut through all that theoretical hype and look at the brutal reality of how autonomous agents are reshaping business economics, actively breaking our security frameworks.
• 1:31
  Speaker 0: And sparking a massive global sovereignty race?
• 1:34
  Speaker 0: Yeah!
• 1:34
  Speaker 1: To even have this conversation we have to isolate specific technology.
• 1:38
  Speaker 1: that's breaking things because we aren't talking about chatbots who write your marketing emails anymore.
• 1:43
  Speaker 1: We're talking about agentic systems
• 1:44
  Speaker 0: Totally different ballgame
• 1:45
  Speaker 1: Exactly!
• 1:46
  Speaker 1: And Jothi Murphy shared really elegant way to visualize this on LinkedIn using the concept of Russian nesting dolls.
• 1:54
  Speaker 0: Oh, I like that
• 1:55
  Speaker 1: right?
• 1:55
  Speaker 1: So AI terminology nested inside itself generative AI is stuff that responds to prompts and creates content.
• 2:00
  Speaker 1: That's an inner doll.
• 2:01
  Speaker 1: Ai agents are the newest outermost doll
• 2:04
  Speaker 0: Right?
• 2:04
  Speaker 0: so gen ai gives you The text but the agent takes a high-level goal independently selects unnecessary tools And then executes some multi step process with almost zero supervision.
• 2:14
  Speaker 0: i mean they Are completely different beasts But here's Enterprise leadership teams are treating them like they're the exact same thing.
• 2:21
  Speaker 1: They really are and
• 2:22
  Speaker 0: that misunderstanding is directly causing this massive wave of pilot
• 2:27
  Speaker 1: failures.
• 2:27
  Speaker 1: Oh, the failure rates are staggering right now.
• 2:30
  Speaker 1: Yeah Eric Pilkington actually brought some highly uncomfortable math to the table regarding these enterprise pilots.
• 2:37
  Speaker 1: He pointed out that ninety five percent of generative AI enterprise pilots Are currently showing absolutely zero measurable PNL impact.
• 2:45
  Speaker 0: wait zero zero
• 2:46
  Speaker 1: pnl impact.
• 2:47
  Speaker 1: After billions of dollars in enterprise spend across the board.
• 2:50
  Speaker 0: That is a brutal reality check.
• 2:51
  Speaker 0: So what is actually driving that disconnect?
• 2:53
  Speaker 0: I mean, why is it feeling so hard?
• 2:55
  Speaker 1: It really comes down to the ruthless mechanics of compounding reliability In an agentic workflow.
• 3:01
  Speaker 1: So people look at an AI agent that is say ninety five percent reliable on a single task and think great Ninety-five percent as an A plus
• 3:08
  Speaker 0: which i mean for a single generative ai prompt.
• 3:11
  Speaker 0: it Is
• 3:12
  Speaker 1: right but agents don't execute single tasks they execute workflows.
• 3:17
  Speaker 0: Ah, okay.
• 3:18
  Speaker 0: so if you tell an agent to process a vendor invoice it has to read the email extract the PDF data cross-reference the vendor ID in the database check the approved budget draft the approval and hit send in the accounting software.
• 3:33
  Speaker 0: that isn't one step...that's what.
• 3:34
  Speaker 0: maybe twelve consecutive steps?
• 3:36
  Speaker 1: Exactly!
• 3:37
  Speaker 1: So that agent is ninety five percent reliable at each individual.
• 3:41
  Speaker 1: You have to multiply that ninety five percent probability against itself twelve times and when you do That math your end-to-end success rate drops to fifty four percent.
• 3:49
  Speaker 0: Wow, which makes it basically a coin toss.
• 3:52
  Speaker 0: exactly I mean if a human employee flipped a coin To decide whether an invoice actually gets paid or get lost in the void.
• 3:57
  Speaker 0: You'd fire them immediately.
• 3:59
  Speaker 1: And this is precisely why these pilots aren't scaling.
• 4:01
  Speaker 1: because what a multi step autonomous process fails.
• 4:04
  Speaker 1: at Step eight A human has to step in, figure out what the agent did wrong, reverse bad actions and finish job manually.
• 4:11
  Speaker 1: You haven't saved time.
• 4:12
  Speaker 1: you've just created a highly complex troubleshooting task.
• 4:15
  Speaker 0: Yeah, that makes total sense.
• 4:17
  Speaker 0: And this maps perfectly onto a framework Nandan Mulakara shared recently called the AI Automation Funnel.
• 4:23
  Speaker 0: He noted only thirteen percent of companies successfully scale AI automation because they fundamentally misunderstand the sequence of transformation.
• 4:32
  Speaker 1: Okay so what's his funnel look like?
• 4:34
  Speaker 0: Well he has five stages Eliminate, Standardize, Optimize, Automate and Innovate.
• 4:40
  Speaker 1: Let me guess Companies are just jumping straight to stage four Automate
• 4:45
  Speaker 0: completely bypassing the unsexy, difficult work of stages one through three.
• 4:50
  Speaker 0: They don't eliminate redundant tasks or standardize their messy data.
• 4:55
  Speaker 0: they just buy an AI agent and drop it onto a Birken workflow.
• 4:59
  Speaker 0: but wait if we're just deploying these agents on to broken inefficient legacy processes aren't we basically like strapping a Ferrari engine into a horse-drawn carriage?
• 5:07
  Speaker 0: That's
• 5:07
  Speaker 1: a perfect way to put it.
• 5:08
  Speaker 1: you're just magnifying your inefficiency at machine speed.
• 5:11
  Speaker 1: Mark Byershutter actually highlighted this exact flaw in how companies approach AI transformation workshops.
• 5:18
  Speaker 1: What's
• 5:18
  Speaker 0: happening?
• 5:18
  Speaker 0: the workshops?
• 5:19
  Speaker 1: Well, he noticed that leadership teams sit down and they ask where can we use AI?
• 5:24
  Speaker 1: but That question inherently assumes that your current processes are sacred And must be preserved.
• 5:29
  Speaker 0: you're just trying to cram new technology into old boxes.
• 5:32
  Speaker 1: precisely mark argues that the organizations Actually seeing P&L impact or asking a much more aggressive question They're asking If we built this business from scratch today, would the process even exist?
• 5:46
  Speaker 1: They aren't trying to automate yesterday's workflows.
• 6:08
  Speaker 0: Sandeep S. made a fascinating prediction about this, he believes agentic AI will make traditional seat based software pricing completely obsolete by twenty-twenty eight.
• 6:17
  Speaker 1: I mean that makes perfect sense when you think about the mechanics of saws over the last two decades.
• 6:22
  Speaker 1: Yeah You buy software by the seed.
• 6:24
  Speaker 1: One license equals one human user logging into the dashboard.
• 6:29
  Speaker 0: Right but agents don't need dashboards They don't have thumbs they don't take lunch breaks and they definitely Don't need a UI.
• 6:36
  Speaker 0: so If one autonomous agent is executing the workload of fifty human analysts via API calls in the background, you only need one seat.
• 6:45
  Speaker 1: Yeah...the math breaks down
• 6:46
  Speaker 0: Exactly!
• 6:48
  Speaker 0: So Sandeep argues that because this, The fundamental unit value and enterprise are shifting from access to tools To accuracy decisions.
• 6:56
  Speaker 0: That's
• 6:57
  Speaker 1: a profound economic pivot.
• 6:58
  Speaker 1: Software vendors will be forced to charge for outcomes the system orchestrates Rather than logins.
• 7:03
  Speaker 1: it supports
• 7:04
  Speaker 0: Definitely
• 7:05
  Speaker 1: But this brings us to a massive looming crisis.
• 7:07
  Speaker 1: If software pricing shifts because agents are acting autonomously, what happens to governance when those autonomous decisions are wrong?
• 7:15
  Speaker 1: Oof!
• 7:16
  Speaker 0: Yeah that is the prillion dollar question right there Because when the tool gets up and walks around, your old compliance manual is essentially useless.
• 7:24
  Speaker 0: Alan Parris highlighted a striking prediction from Gartner that really puts this into perspective.
• 7:29
  Speaker 1: What did Gardner say?
• 7:30
  Speaker 0: They expect by twenty-twenty seven forty percent of enterprises will actually demote or entirely decommission their autonomous AI agents.
• 7:39
  Speaker 0: Forty
• 7:39
  Speaker 1: percent!
• 7:39
  Speaker 1: That's a massive retreat
• 7:41
  Speaker 0: It is.
• 7:41
  Speaker 0: And the kicker is they won't decommision them because the agents perform poorly and scaps that only surface when things break in production.
• 7:52
  Speaker 0: Alan framed the distinction beautifully, he said.
• 7:54
  Speaker 0: traditional model governance asks if a model is fit for purpose.
• 7:58
  Speaker 0: meaning does
• 8:05
  Speaker 1: Can we trust it to delete a record?
• 8:07
  Speaker 1: can We Trust It To Move Money.
• 8:08
  Speaker 1: By the way, if you want to keep up with these massive architectural shifts before they break your own production environment make sure to subscribe so you catch our future deep dives!
• 8:17
  Speaker 0: Yeah definitely hit subscribe.
• 8:18
  Speaker 1: but going back to Alan's point about being fit-to-act this governance gap isn't just a corporate IT headache anymore...it is becoming a systemic risk.
• 8:28
  Speaker 0: Right because these agents aren't just operating in a vacuum They are interacting.
• 8:34
  Speaker 0: Jonas Jacoby actually looked at a recent speech by Sarah Breeden, who's the deputy governor of The Bank Of England.
• 8:39
  Speaker 1: Okay what was her take?
• 8:41
  Speaker 0: She essentially admitted that current financial regulations have a massive glaring blind spot when it comes to agents
• 8:47
  Speaker 1: because the entire global Financial Regulatory Framework assumes there is a human hitting the approve button
• 8:55
  Speaker 0: right.
• 8:55
  Speaker 1: The whole system relies on the concept of a human in-the-loop.
• 8:58
  Speaker 1: Exactly!
• 8:59
  Speaker 1: But at the speed, volume and scale that autonomous agents operate relying on a human to verify every single microtransaction trade or loan adjustment is just completely unrealistic.
• 9:10
  Speaker 0: Yeah we're talking about agents autonomously interacting with payment rails – the underlying networks that move money between banks.
• 9:16
  Speaker 0: If an agent hallucinates a market signal and executes ten thousand microtrades in a minute….
• 9:21
  Speaker 0: …The Human In-the Loop Is Basically Just A Spectator Watching The Crash
• 9:25
  Speaker 1: Which means the security architecture has to fundamentally change.
• 9:29
  Speaker 1: And Steve Neury brought up a critical point regarding this, he noted that agents aren't inherently malicious but To be useful they require deep unfettered access right?
• 9:39
  Speaker 1: They need to query your databases read Your customer records.
• 9:42
  Speaker 1: cross-reference internal policies trigger workflows.
• 9:46
  Speaker 1: Because of that access requirement.
• 9:48
  Speaker 1: steve argues That the data layer itself is becoming The new security boundary.
• 9:53
  Speaker 1: It's no longer enough to just put a firewall around the application layer.
• 9:56
  Speaker 0: Okay, but hold on.
• 9:57
  Speaker 0: if the data layer is The new boundary aren't we essentially handing?
• 10:00
  Speaker 0: A highly capable autonomous intern the master key To the corporate filing cabinet and just like hoping We can track what they photocopy were giving these systems the keys to the kingdom we are
• 10:12
  Speaker 1: And that Is why securing at the source isn't No longer optional.
• 10:15
  Speaker 1: you have to build resilience directly into the database.
• 10:18
  Speaker 1: You need to know exactly what the agent pulled, why its logic dictated that pull and where it routed that information.
• 10:23
  Speaker 1: It requires patching faster than attackers And maintaining a perfect forensic trail
• 10:28
  Speaker 0: Which sounds like an absolute compliance nightmare Especially for smaller companies That don't have infinite IT budgets.
• 10:34
  Speaker 1: Yeah Sounds daunting
• 10:35
  Speaker 0: But there is really interesting kind of counterintuitive take here.
• 10:39
  Speaker 0: Adam Helbig pointed out that for European small and medium businesses, the EU-AI Act shouldn't necessarily be viewed as a crushing compliance burden.
• 10:49
  Speaker 0: Wait!
• 10:50
  Speaker 1: How is massive regulatory framework not a burden on an SMB?
• 10:54
  Speaker 1: Because...
• 10:55
  Speaker 0: That exact governance discipline the act demands is precisely the blueprint required to make agentic AI reliable enough to deploy.
• 11:03
  Speaker 1: Oh I see Yeah.
• 11:04
  Speaker 0: If you are forced to build the safety rails, implement data layer security and establish clear audit trails for compliance.
• 11:12
  Speaker 0: You're inadvertently building the exact architecture needed to trust your agents to run autonomously.
• 11:17
  Speaker 1: That is a brilliant reframe.
• 11:18
  Speaker 1: The Compliance mandate forces you to build infrastructure of trust.
• 11:22
  Speaker 1: You literally earn right to scale automation by proving governance.
• 11:27
  Speaker 1: We need to zoom out from corporate IT for a minute because this governance challenge is rapidly spilling onto the geopolitical stage.
• 11:35
  Speaker 1: As agentic AI proves its sheer economic power, governments are realizing that simply regulating AI from outside isn't sufficient anymore.
• 11:44
  Speaker 0: They need to own the underlying infrastructure.
• 11:47
  Speaker 0: We are watching a real-time shift from governments acting as referees to government's trying to buy the stadium.
• 11:53
  Speaker 1: Yeah, that is good way of putting
• 11:54
  Speaker 0: it.
• 11:54
  Speaker 0: And Agnisey Janoshan highlighted major geopolitical development in her recent post.
• 11:59
  Speaker 0: She noted reports that Sam Altman offered US Government A five percent equity stake In open AI.
• 12:05
  Speaker 1: Just put some numbers on that Five percent stake would be worth roughly forty two billion dollars.
• 12:09
  Speaker 1: Yeah,
• 12:09
  Speaker 0: forty two billions.
• 12:11
  Speaker 0: The proposal supposedly suggests other major AI labs could follow this exact template.
• 12:17
  Speaker 0: Just to be clear, we are strictly looking at the structural implications of this based on the sources.
• 12:21
  Speaker 0: We aren't taking any political stance on whether this is right or wrong
• 12:24
  Speaker 1: Right force
• 12:25
  Speaker 0: but as Agnese points out if a government takes an equity stake It becomes a financial co-owner of the foundational models that global businesses run on.
• 12:33
  Speaker 1: But doesn't that create and immediate glaring conflict?
• 12:36
  Speaker 1: Of interest I mean.
• 12:38
  Speaker 1: If a government holds a massive equity stake in a leading AI lab Doesn't that fundamentally blur the line between a sovereign regulator and a corporate shareholder?
• 12:47
  Speaker 1: How can any government objectively police, restrict or penalize a company That it directly profits from.
• 12:54
  Speaker 0: That is the exact tension The sources are flagging.
• 12:57
  Speaker 0: Symmetra data analyzed this dynamic And argued that Washington's apparent shift for merely regulating AI to actively buying into It should be a massive wake-up call For Europe.
• 13:07
  Speaker 1: interesting what's his solution?
• 13:09
  Speaker 0: Sumitra argues that AI sovereignty requires a patient cross-border strategy.
• 13:13
  Speaker 0: He points to how European nations historically pooled their capital and political will to build Airbus.
• 13:19
  Speaker 1: Ah, that's a fascinating historical parallel because airbus wasn't built by single startup right?
• 13:24
  Speaker 1: Yeah It was collective sovereign effort to ensure Europe isn't entirely dependent on American aerospace engineering.
• 13:30
  Speaker 1: So Sumitra is saying Europe needs that exact same collective infrastructure for AI, something robust enough to outlast shifting political wins.
• 13:39
  Speaker 0: And we're actually starting to see that Airbus model take shape specifically at the language level.
• 13:43
  Speaker 0: Tina Austin posted about an emerging concept she calls Ai nationalism.
• 13:48
  Speaker 0: She highlighted a recent launch of Amalia in Portugal.
• 13:52
  Speaker 0: It is the first open-source model built specifically for European Portuguese.
• 13:56
  Speaker 1: Wait, why does a country need its own specific language model when major global models already translate Portuguese perfectly well?
• 14:03
  Speaker 0: Because translation isn't comprehension.
• 14:06
  Speaker 0: Countries are waking up to that AI infrastructure is fundamentally public sector infrastructure.
• 14:12
  Speaker 0: If an AI system doesn't deeply understand the local cultural context, this specific regional legal nuances or historical dialect it simply cannot be trusted to operate public services.
• 14:22
  Speaker 1: That makes total sense!
• 14:24
  Speaker 1: You can't rely on a model trained predominantly in American Internet data with American Cultural Guardrails... ...to autonomously govern European Municipality or Process Regional Tax Law
• 14:33
  Speaker 0: Exactly.
• 14:34
  Speaker 0: and This push for sovereign localized models is entirely rewriting the map of global compute power.
• 14:41
  Speaker 0: Allie Kay Miller shared some eye-opening data about this.
• 14:45
  Speaker 0: We have this in French assumption that American labs are the undisputed permanent center of gravity for AI, but Allie noted on open router platform Chinese models specifically from Xiaomi recently crossed.
• 15:00
  Speaker 1: Let's quickly clarify a token volume for context.
• 15:02
  Speaker 1: That metric isn't just about how many models exist, right?
• 15:05
  Speaker 1: It measures the actual raw amount of data being processed and generated by users.
• 15:09
  Speaker 1: it is a metric of actual real-world adoption in workload execution
• 15:13
  Speaker 0: Right!
• 15:14
  Speaker 0: And to put that forty five percent perspective OpenAI was sitting at seven point four percent on this same platform.
• 15:19
  Speaker 0: Yeah, and Ali pointed out The brutal economic reality driving this shift.
• 15:24
  Speaker 0: These open source models are proving to be eighty percent as capable as Frontier models For roughly twenty percent.
• 15:29
  Speaker 1: Well, if you're a corporate procurement officer looking to deploy agents across thousands of internal workflows that math is undeniable.
• 15:36
  Speaker 1: You don't need a genius-level frontier model to process vendor invoices... ...you just need a model that's good enough and incredibly cheap
• 15:45
  Speaker 0: which completely decentralizes the power dynamics in industry.
• 15:48
  Speaker 0: If we look at insights from Ali Miller alongside a geographic map shared by Tim Cardin A very clear picture emerges.
• 15:56
  Speaker 0: Cardin mapped out the thirty-five most important AI companies on earth right now.
• 16:00
  Speaker 1: And where are they?
• 16:01
  Speaker 0: Well, The frontier of AI capability is no longer isolated to Silicon Valley.
• 16:05
  Speaker 0: it's Mistral in Paris Deep Mine and Eleven Labs In London and massive open course hubs across Asia & Middle East.
• 16:13
  Speaker 1: So the monopoly on intelligence Is breaking apart.
• 16:16
  Speaker 1: We're entering an era of sovereign AI Where individual nations and regional blocks are ensuring they have their own localized compute, their own culturally aligned models and their own secure data layers.
• 16:26
  Speaker 1: Because think about the brisk profile if your entire national economy you're banking sector in your public services run on an agentic system.
• 16:34
  Speaker 1: You simply cannot afford a scenario where A foreign entity whether that's attack corporation or a foreign government can just flip a switch And turn your economy off.
• 16:43
  Speaker 0: so true?
• 16:44
  Speaker 0: So If we pull all of these threads together We've tracked a massive evolution today.
• 16:49
  Speaker 0: We started with the sheer failure rate of generative AI pilots because companies didn't understand a compounding math of agentic workflows.
• 16:57
  Speaker 0: we saw how true adoption requires redesigning operating model which destroys traditional software pricing, right?
• 17:04
  Speaker 0: We looked at critical need to secure data layer as agents gain autonomy.
• 17:07
  Speaker 0: and finally this entire shift is forcing governments build their own sovereign infrastructure.
• 17:13
  Speaker 1: It's massive amount of systemic change happening simultaneously.
• 17:17
  Speaker 1: You know, if we connect the corporate struggles to the geopolitical race there's one clear defining through line.
• 17:23
  Speaker 1: As we transition away from humans using AI as a static tool like that power drill we talked about earlier and move toward organizations orchestrating complex human agent collaboration We are gonna hit a wall And The ultimate bottleneck won't be the compute power and it won't Be the technology itself?
• 17:41
  Speaker 1: The bottleneck is going To be authority.
• 17:44
  Speaker 0: Wait break That down.
• 17:46
  Speaker 0: why Authority?
• 17:47
  Speaker 1: because authority is inextricably linked to liability.
• 17:51
  Speaker 1: When a system just acts as an advisor, like when it drafts a report or suggests a marketing strategy the human hits send right?
• 17:58
  Speaker 1: The Human holds the Authority and the human holds the Liability if goes wrong.
• 18:01
  Speaker 1: but when agent operates autonomously the paradigm breaks.
• 18:05
  Speaker 1: When an Agent independently decides to approve high risk loan Or deny critical health care claim or execute multi-million dollar algorithmic trade that crashes portfolio who actually owns the liabilities?
• 18:16
  Speaker 0: huh Is it the enterprise that deployed the agent?
• 18:19
  Speaker 0: is at developer who trained model?
• 18:21
  Speaker 0: or in this new geopolitical reality, government shareholder owns an equity stake.
• 18:26
  Speaker 1: We don't know because framework simply do not exist.
• 18:29
  Speaker 1: yet we are sprinting into a era where have to figure out how scale trust governance and legal liability exactly as fast scaling machine intelligence.
• 18:40
  Speaker 1: because an autonomous drill is an absolute miracle of modern engineering, right up until the exact moment it starts drilling into a load-bearing wall without asking for
• 19:07
  Speaker 0: permission.