Best of LinkedIn: Smart Manufacturing CW 17/ 18

00:00:00: This episode is provided by Thomas Allgaier and Frennus, based on the most relevant LinkedIn posts about smart manufacturing in calendar weeks, seventeen and eighteen.

00:00:09: Frenness is a B to D market research company that supports Smart Manufacturing providers with building feature-by-feature competitive intelligence that shows exactly how their product stacks up against the competition.

00:00:21: you can find more info in the description.

00:00:23: awesome.

00:00:23: so let's get right into

00:00:24: it.

00:00:24: yeah Let's do it.

00:00:25: So imagine finding out line-stopping manufacturing defect like seventy two hours before it actually happened.

00:00:34: Oh, that's the dream right?

00:00:35: Right

00:00:36: but no crystal ball.

00:00:37: No guessing.

00:00:38: just a system That's continuously monitoring you know over a thousand process parameters and yielding something crazy Like a quarter of billion dollars in incremental profit per factory

00:00:48: which is Just a staggering reality.

00:00:50: yeah And today that's our mission for you We're doing a deep dive into the top smart manufacturing trends we've seen across the LinkedIn community over calendar weeks, seventeen and eighteen.

00:01:00: We are looking at how physical AI is exactly what I just mentioned.

00:01:03: Yeah!

00:01:03: And if you look at major themes dominating space over these last couple of weeks especially coming out events like handover mess twenty-twenty six there's very distinct vibe shift.

00:01:15: A Vibe Shift?

00:01:16: How so

00:01:17: Well were entirely passed the hype phase of industry.

00:01:20: four point oh like.

00:01:21: we are no longer talking about these distant you know futuristic concepts.

00:01:25: We're in a period of ruthless industrialized execution.

00:01:30: exactly I mean if your navigating realities of construction and manufacturing industries right now.

00:01:35: separating actual operational shifts from marketing noise is basically full time job.

00:01:39: It

00:01:40: really is.

00:01:40: So we're going to unpack how leaders are actually scaling AI, why digital twins or suddenly these living operational backbones?

00:01:47: How they're fixing the broken data architectures and Physical AI is finally stepping onto the shop floor.

00:01:54: Yeah, and I think we should start with that AI piece because The way it's actually rolling out is completely counterintuitive to what most people thought.

00:02:00: Oh totally Because we all kind of expect a i'd be this really loud flashy Factory brain That just takes over day one right

00:02:07: but the reality on the ground What?

00:02:09: We might call the quiet ai revolution Is the exact opposite.

00:02:13: yeah, I was reading John Healy take on This recently And he made this fantastic observation.

00:02:18: He noted that Ai in the industrial space It isn't a switch.

00:02:22: you suddenly flip.

00:02:23: No, not at all.

00:02:24: it's quietly embedding itself into how systems run day-to-day.

00:02:28: like You don't necessarily notice the AI you just start noticing Fewer issues smoother operations

00:02:35: and a lot less daily firefighting which I know everyone listening can appreciate exactly And Maggie Slowick echoed that perfectly with her takeaways from handover mess.

00:02:43: Generative AI is not showing up as this omnipotent autonomous manager.

00:02:47: its showing up has highly focused task-specific

00:02:52: agents.

00:02:53: Yeah, we're talking about agents embedded in the background to say optimize a project schedule or update of purchase order based on real time supply fluctuation?

00:03:01: Okay

00:03:01: let's unpack this for second.

00:03:02: because wait when I hear agentic AI my mind immediately jumps to.

00:03:07: you know a chatbot confidently hallucinating completely fabricated answer.

00:03:11: Sure yeah that's the fear!

00:03:13: So why on earth would a massive manufacturer hand over procure to pay, To an LLM.

00:03:21: Like isn't that An enormous risk?

00:03:23: Well it's a great question and It really comes down The environment you put the agent in.

00:03:27: Robert Gurvin actually shared some Really prescriptive data on this.

00:03:31: Oh what did he find?

00:03:32: So back office processing costs manufacturers upto five percent of their revenue.

00:03:37: For A ten billion dollar manufacturer That is half a Billion dollars a year.

00:03:42: Wow

00:03:43: just On the Back Office Right.

00:03:45: And Gervin is seeing Agentec AI deliver five to ten times productivity gains in that exact space.

00:03:51: Five-to-ten times, just in the back office?

00:03:54: Yes specifically in procure-to pay and record-to report because those areas are highly standardized.

00:03:59: they carry lower risk than you know tweaking a machine on the shop floor but they offer the absolute highest return on investment.

00:04:05: right

00:04:05: cuz it's invoice transcription its three way matching

00:04:08: exactly It's the places where up to sixty-five percent of manual repetitive work actually lives.

00:04:14: But

00:04:15: okay, I gotta push back a little.

00:04:17: How is an AI agent handling and invoice fundamentally different from the automated scripts or OCR macros we've used in accounting for?

00:04:26: That's true, but the difference is contextual reasoning versus rigid rules.

00:04:31: An automated script is entirely rule-based.

00:04:34: if X happens do Y right?

00:04:36: But If a supplier changes the layout of their invoice your traditional script just breaks because The text isn't in the exact pixel coordinates it expects.

00:04:45: Oh yeah anyone who's managing ERP integration knows that pain

00:04:48: exactly.

00:04:49: But an LLM-backed agent understands the context.

00:04:52: It knows what a total do is, regardless of where it's sitting on the page.

00:04:55: Interesting!

00:04:56: And if a supply chain disruption happens... ...an agent can actually reason through the available alternatives validate material availability across different suppliers and recommend a rerouting action before the disruption becomes production stoppage that actually adapts to messiness in real world.

00:05:13: That adaptability completely changes how we define technology.

00:05:18: And, you know, Alexei Samyshalov brought up a really sharp definition from an industry leader's dialogue on LinkedIn.

00:05:24: What

00:05:24: did he say?

00:05:25: He argued that we have to stop blurring the terms.

00:05:28: like industrial AI isn't just legacy pattern recognition or standard computer vision That we've had for a decade

00:05:35: right.

00:05:35: it's a completely different paradigm.

00:05:37: Yeah

00:05:37: The real shift is taking large language models and applying them directly to production KPIs and engineering data Like, imagine a foundation model trained on your specific technical reports maintenance logs and sensor outputs.

00:05:51: Not just public internet text.

00:05:53: Exactly.

00:05:53: And when you apply that to engineering workflows the reduction in human effort is just immediate.

00:05:58: Gwennel Huey posted an update of the new industrial co-pilot from Schneider Electric & Microsoft...and it's wild!

00:06:05: Oh I saw that That's the agentic system right?

00:06:07: Yeah It actually designs validates and deploys automation logic In single connected workflow.

00:06:13: She noted it's cutting automation engineering time by up to fifty percent.

00:06:16: Half the time, that is huge!

00:06:19: Because isn't just a dashboard making recommendation?

00:06:21: It's system that reasons writes logic and

00:06:25: acts... Okay there's glaring problem here.

00:06:28: if these AI agents are capable of autonomous action like writing logic rerouting supplies, you absolutely cannot let them test those actions on a live factory floor where a mistake costs millions of dollars.

00:06:43: No!

00:06:43: Absolutely not.

00:06:44: that would be a disaster

00:06:45: right?

00:06:45: so they need a sandbox to fail in.

00:06:47: precisely if You're gonna scale autonomous reasoning?

00:06:50: You need an environment That perfectly mirrors reality and that bridges us directly into the second major theme from The last two weeks which is the evolution of digital twins.

00:06:59: Yeah, and Mark Dietrich offered a really sharp critique of how the industry used to handle this.

00:07:04: He was talking about traditional model-based systems engineering or you know MBSC right.

00:07:08: for a long time mbsc Was this massive buzzword that honestly in reality often just resulted in fancy digital vizio drawings?

00:07:15: yeah Just static models

00:07:16: exactly.

00:07:17: organizations treated The three d model as the final deliverable And sure it solved the paper document problem but It just created a new expensive static silo.

00:07:27: Right, and the necessary leap we are seeing now is turning those static models into living executable digital assets.

00:07:35: A true Digital Twin today links directly to physics-based simulations... ...and pulls live PLC data from the

00:07:43: floor.".

00:07:43: And the real world impact of that is staggering!

00:07:46: There's this incredible example shared by Olmal Cascanero Rios and Hank Vildgones about what happens when a global CPG leader actually makes The Digital Twin foundational.

00:07:55: Oh, the PepsiCo example.

00:07:56: Yes!

00:07:57: PepsiCo used the Siemens Digital Twin Composer combined with NVIDIA Omniverse to modernize their facilities and they basically converted entire factories in warehouses into high-fidelity three D digital twins

00:08:08: every single machine

00:08:09: every machine conveyor operator pass all of it mild.

00:08:12: And what's crucial is how they use that unified scene, like I didn't just look at on the screen.

00:08:16: AI ran realistic high speed experiments at scale to validate the line design and material flow before any physical changes occurred.

00:08:24: in their results.

00:08:25: The early results were re-harkable.

00:08:27: PepsiCo saw a twenty percent increase in throughput and ten to fifteen percent reduction capital expenditure.

00:08:34: That's massive!

00:08:35: Why?

00:08:36: Just from simulating?

00:08:37: because

00:08:38: they achieved nearly one hundred percent design validation before the built anything.

00:08:43: They identified up to ninety-percent of potential bottlenecks.

00:08:46: completely virtually it's

00:08:48: like a digital twin is no longer just a blueprint, its high stakes rehearsal space where you can make all your expensive mistakes for free!

00:08:56: Exactly

00:08:57: For you listening.

00:08:58: Just imagine being able test an entirely new assembly line.

00:09:01: virtually You push it to its absolute breaking point, find all the flowers before you even buy a single physical bolt.

00:09:08: It completely de-risks that capital

00:09:10: deployment.".

00:09:10: It really does and is leading two entirely new categories of facility engineering.

00:09:15: Michelle Mitch Peterson brought up this fascinating concept on LinkedIn about the

00:09:20: AI factory.

00:09:20: Yeah

00:09:21: she asks what do we call an industrial plant part power utility or super computer?

00:09:28: Because building infrastructure physically manufacture intelligence like these massive data centers running these models, it is incredibly complex because of the massive power constraints and cooling requirements.

00:09:41: You literally cannot design this in a traditional way?

00:09:44: No you can't!

00:09:44: They're being evaluated on completely new metrics.

00:09:47: Mitch Peterson mentioned time to token which is speed from concept to actual operation tokens per watt which measures extreme energy efficiency.

00:09:59: That's a totally

00:09:59: different language for a plant manager,

00:10:01: Totally you have to simulate the thermal power and system interactions in a digital twin first.

00:10:08: your engineering complex hardware ecosystems that don't even exist yet.

00:10:11: You have to know The cooling system will prevent the servers from melting before you pour the concrete

00:10:16: Which is mind-blowing.

00:10:17: but okay here Is the critical roadblock?

00:10:21: But if the script it's running on is gibberish, The play is still going to be a disaster.

00:10:28: Oh one hundred percent.

00:10:30: You cannot run a living digital twin and you certainly can not train an AI agent If the underlying data Is a chaotic mess?

00:10:37: If your AI is feeding on garbage It is just going To hallucinate efficiently.

00:10:42: And this brings us to theme three.

00:10:44: This is exactly where so many digital transformations just stall out.

00:10:48: It requires a fundamental shift in data architecture.

00:10:51: Zanak Machalla made this brilliant distinction.

00:10:54: regarding the unified namespace or UNS You know, A lot of manufacturers treat The UNS as the ultimate holy grail for all their data problems.

00:11:02: But Machala points out that a UNS only solves the connectivity problem.

00:11:06: Right right

00:11:07: clean.

00:11:07: MQTT topic hierarchy gives you one central hub were all your data lands which is great But it doesn't give you meaning.

00:11:15: Yes, I love the example used for this.

00:11:17: if The system spits out a raw value of let's say seventy two point four That means absolutely nothing to an algorithm Nothing

00:11:23: at all.

00:11:23: is it Celsius?

00:11:25: Fahrenheit ambient room temperature coolant?

00:11:27: You have no idea

00:11:28: exactly.

00:11:29: think of the UNS like a massive factory mailroom.

00:11:33: All the letters from every department arrive in one big sorting bin.

00:11:36: Great.

00:11:37: You have connectivity, but if half the letters are in French a quarter and Mandarin And the rest is some weird technical shorthand your mailroom clerk still can't read them.

00:11:46: That's

00:11:46: a great analogy.

00:11:48: semantics Are what tell you?

00:11:49: The value Is the spindle bearing temperature in Celsius on machine cnc.

00:11:54: Oh four and it is twelve degrees above normal.

00:11:57: and then to really make It work you need an ontology right.

00:12:01: The ontology is the Rosetta Stone.

00:12:02: It's a shared vocabulary across every site, so that when Site A and Site B both output a variable for spindle temperature they mean the exact same mathematical thing.

00:12:11: And finally you need a knowledge graph that stores relationships between these things Turning an alert into actual insight

00:12:18: Exactly Without those layers.

00:12:20: Machalla warns your manufacturing AI agent Is basically just chatbot hallucinates KPIs.

00:12:26: Okay here where it gets really interesting though.

00:12:29: We've been talking about breaking down data silos since, like the dawn of the internet.

00:12:33: That is not a new phrase!

00:12:35: Just

00:12:35: very true...

00:12:36: What makes this Data Fabric Ontology approach actually different from the massive data lakes we all built five years ago?

00:12:43: Didn't already try to solve it by just dumping everything into the

00:12:46: cloud?!

00:12:47: We did but fundamental shift here moving merely storing data and establishing semantic relationships between datapoints.

00:12:56: Shankar Raman pointed out that when companies try to solve this by just dumping everything into a data lake, they often create bigger and more expensive silo.

00:13:05: Because it's still disconnected in meaning?

00:13:06: Exactly!

00:13:07: A Data Lake is holding the raw information.

00:13:10: Someone has to manually pull data from the product life cycle management system check the manufacturing execution systems logs cross-reference supplier scorecards all to figure why a park failed.

00:13:21: So human is still an integrator?

00:13:22: Yes

00:13:23: But an ontology maps those relationships in advance.

00:13:26: It understands that a design tolerance in the PLM is directly related to specific machine vibration and the MES, so the AI can instantly trace equality issue on the floor back to exact engineering

00:13:40: spec.".

00:13:40: That's incredible!

00:13:41: You really cant answer these cross-domain questions without it?

00:13:44: And...the cost of not having this is literally eating into operating margins.

00:13:49: David Rogers posted about Something he calls the builders tax.

00:13:53: Oh, yeah The builder's tax Yeah

00:13:55: for you listening if your spending massive chunks of your it budget just moving data around this is why?

00:14:01: The Builders Tax Is the Massive Cost Of Complex Reverse ETL.

00:14:06: You Know Extract Transform Load Right.

00:14:09: You build Your Applications and Transactual Databases You Build Your Intelligence in a Lakehouse And Then You Spend Millions essentially shipping the data back and forth just to keep them in sync.

00:14:19: Which

00:14:19: requires massive computing power, and manual mapping Exactly!

00:14:23: It's a huge drain

00:14:24: which is why the infrastructure layer so critical right now.

00:14:27: Kudzai Mendetreza has been heavily advocating for common industrial API standard specifically citing i-three x from CESME.

00:14:36: Okay what does that do?

00:14:37: Well, think about it.

00:14:37: The number of specialized software products in the manufacturing stack has just exploded.

00:14:42: Every new tool brings another interface and other proprietary data format.

00:14:46: It's a nightmare for IT.

00:14:48: A common standard like I-IIIx gives manufacturers a lightweight universal way to ask any system what information it has and how is structured.

00:14:57: It's the missing layer for scaling intelligence without paying that builder's

00:15:01: tax.".

00:15:02: So, How do manufacturers navigate this practically?

00:15:04: Because if you're plant manager... You can't just rip or replace your entire IT architecture tomorrow

00:15:09: morning!

00:15:10: No, of course not.

00:15:11: And Muserat Hussein offered a very grounded tip for that.

00:15:14: He said manufacturers need to stop buying buzzwords and start partnering with pragmatic integrators

00:15:19: Pragmatic integrators, I like that.

00:15:21: Yeah these are partners who natively speak both IT and OT.

00:15:26: They're as comfortable discussing cloud API limits As they talk about PLCs And overall equipment effectiveness on the shop floor.

00:15:33: So bridge gap

00:15:34: Exactly!

00:15:35: They value open interoperability standards Rather than locking you into proprietary ecosystems.

00:15:41: And crucially start small.

00:15:42: They prove ROI On a single line before trying to boil ocean.

00:15:46: That makes total sense.

00:15:47: Start small, prove the value.

00:15:49: Yeah Okay.

00:15:50: so we fixed a data pipeline.

00:15:53: We've rehearsed in the digital twin and got agents handling back-office purchasing Yep.

00:15:58: Which brings us to theme four.

00:16:00: How does this actually hit physical shop floor?

00:16:03: Because once you have meaningful data You can finally unleash physical AI.

00:16:08: Yes And Demetrio Spiliopoulos shared some incredible observations from HandoverMess regarding his exact transition.

00:16:16: The arrival of physical AI is no longer theoretical.

00:16:19: What really stood out was the integration of vision language models, or VLMs into robotics.

00:16:24: Okay

00:16:24: impact that for us!

00:16:26: How's a VLM changing a robot?

00:16:28: So historically if you wanted a robot arm to pick up a box You had to hard code its spatial path move x degrees left y degrees down

00:16:34: right very rigid

00:16:35: exactly.

00:16:36: If the box moved two inches the robot just grabbed empty air.

00:16:39: But with a vision-language model workers can give natural language instructions

00:16:42: like talking to chat gpt.

00:16:43: but four robot

00:16:44: Exactly.

00:16:45: You can tell a system, pick up the red box and inspect the barcode.

00:16:49: The robot's camera uses the VLM to visually identify the Red Box understand its geometry in real time And execute the task.

00:16:57: That is wild!

00:16:58: The systems can learn faster and workflows like inspection or picking Can be adapted instantly without heavy hard-coded retraining.

00:17:06: Sean say he did a brilliant analysis on the broader impact of this.

00:17:10: He outlined five forces colliding inside the factory and he made a point that really reframes the whole robotics conversation.

00:17:17: What was his take?

00:17:18: Well we saw human always deployed at The Siemens Electronics Factory in Erlangen, and BMW using them in Leipzig And say he says these aren't just cool novelties or like isolated science experiments.

00:17:30: Right?

00:17:31: They're a necessary response to projected three point eight million workers shortage in U?

00:17:36: S manufacturing over the next decade,

00:17:37: and that is the critical context That so many people miss at that scale of a labor gap.

00:17:42: deploying humanoid robots Is not a labor replacement story.

00:17:45: It isn't operating model shift

00:17:47: because the people just aren't there to hire.

00:17:49: Exactly

00:17:50: hiring alone literally cannot solve A three-point eight million person shortfall.

00:17:55: you have to augment The floor

00:17:56: And the financial impact of deploying physical AI effectively is just massive.

00:18:01: Think back to the hook we started with, Pascal Bergenu shared an analysis from The Semiconductor Industry detailing how Physical AI is monitoring over twelve hundred process parameters continuously

00:18:12: Right.

00:18:12: predicting failures seventy two hours early

00:18:14: Yes It's detecting anomalies seventy-two hours before defects actually manifest on a line.

00:18:20: But How does that translate to profit though?

00:18:22: I mean two hundred and fifty million is huge number

00:18:25: Because In advanced manufacturing, like semiconductors yield is everything.

00:18:30: If you know a wafer's going to fail three days before it does You can adjust the chemical vapor deposition or the lithography alignment in real time.

00:18:38: that yield optimization alone delivers A two hundred and fifty million dollar incremental gross profit per fab.

00:18:44: Wow

00:18:45: Those numbers are incredible for massive operations, but you know we also have to ground this for smaller facilities.

00:18:50: Christurgio offered some very pragmatic advice for small and medium enterprises the SMEs.

00:18:54: Yeah because they don't have two hundred fifty million laying around

00:18:57: Right.

00:18:58: He pointed out that SMEs do not need to start with multi-million dollar humanoid robots.

00:19:04: The immediate low hanging fruit is right sized low-cost semi automated fixtures that are likely already in house.

00:19:11: Oh, like what he calls labor amplifiers?

00:19:14: Yes Labor amplifiers they reduce the physical friction and daily workflows And they multiply the human effort without requiring you to turn the whole shop upside down.

00:19:24: Which brings us to the ultimate question if we're talking about humanoids filling a three point eight million worker gap and Software agents doing the back office purchasing What does this actually mean for the human workforce?

00:19:36: Are we just replacing people or are we fundamentally redesigning The nature of factory work.

00:19:41: We're absolutely redesigning it.

00:19:43: Dean Bartels noted in his insights that leading manufacturers are treating upskilling as part Of the new operating model itself.

00:19:49: Upskilling is mandatory now.

00:19:51: Yeah, they're embedding training directly into technology rollouts because As factories operate faster and systems become more autonomous It moves from manual execution to system orchestration and exception handling.

00:20:06: If you do not upskill the workforce to manage the AI, The structural skills gap will completely undermine the productivity gains.

00:20:16: That is a profound shift.

00:20:18: We've covered a massive amount of ground today from quiet LLM agents reasoning through supply chain disruptions to digital twins acting as high-states rehearsal spaces, the absolute necessity of data ontologies and finally physical AI dynamically acting on the floor.

00:20:33: Yeah!

00:20:34: And if we pull all these threads together Victor M delivered final lingering thought that really captures stakes at this moment.

00:20:41: Oh I saw it was great post.

00:20:43: He noted that industrial AI is rapidly moving out of the pilot era and into the scaling era, but he warned at The Winning Factory in next decade won't necessarily be the most automated one.

00:20:53: Really?

00:20:54: Then what will it be?

00:20:55: It'll be the one with greatest capacity to continually learn from its own data, adapt processes and improve measurably every single shift.

00:21:03: That's a perfect take away.

00:21:05: So as you look your own operations Your own tech stack And team tomorrow morning I want you ask yourself Are you just building a factory with more technology?

00:21:14: Or are you actually building a Factory With greater capacity to learn.

00:21:18: That's the real question!

00:21:19: If you enjoyed this episode, new episodes drop every two weeks.

00:21:23: Also check out our other editions on digital construction and digital power tools.

00:21:28: Thank You so much for taking time to join us For This Deep Dive.

00:21:31: Don't forget To Hit Subscribe So You Never Miss An Insight And We Will Catch You On The Next One.