Best of LinkedIn: Digital Construction CW 42/ 43

00:00:00: Provided by Thomas Allgaier and Franus based on the most relevant LinkedIn posts about digital construction in calendar weeks, forty one and forty two.

00:00:08: Franis is a B to be market research company that supports industrial automation and ICT companies with market intelligence across the construction industry to prioritize segments identify high value accounts and validate use cases.

00:00:21: Yeah just to set expectations right at the top we're diving into the absolute top digital construction trans across linkedin Right now.

00:00:30: So if you're a smart build or manufacturing professional, consider this deep dive your shortcut to staying totally informed without having to sift through all the fluff.

00:00:40: Okay let's unpack this because looking across All The Intelligence we've gathered there is this massive just fascinating central tension tearing Through the construction industry right now.

00:00:50: absolutely We

00:00:51: are essentially trying to bolt A Ferrari engine onto like a nineteen nineties tractor.

00:00:56: that's a great way To put it.

00:00:58: On one hand, we're building this blazing fast highly capable AI tool layer.

00:01:04: But on the other hand We're forcing it to run on top of fragmented siloed and often incredibly messy project data.

00:01:11: And that tension.

00:01:12: well That's exactly where their friction is and The opportunity really.

00:01:16: were watching the collision of the future and past in real time.

00:01:20: Yeah It's wild

00:01:22: because what seeing these highly advanced AI models being layered directly over legacy construction software And it is fundamentally altering how engineers and project managers interact with their models.

00:01:32: Exactly

00:01:34: It's moving from a click-based interaction to a purely conversational one

00:01:38: Which, I mean that gives these legacy applications entirely new superpowers?

00:01:43: That

00:01:43: really does

00:01:44: Like if you are a BIM manager listening to this Think about the sheer amount of muscle memory it takes to navigate a thousand nested drop down menus.

00:01:53: Now imagine just bypassing that entirely by telling an AI what you want.

00:01:58: Right, and we're actually already seeing people build this out.

00:02:02: Julian Pius Martinez documented a really prime example of his workflow.

00:02:07: Oh

00:02:07: yeah the Revit plugin.

00:02:08: Exactly in its single working session he used Claude to build custom Revit plug-in

00:02:13: Which is crazy fast

00:02:15: Insanely fast The mechanism he uses.

00:02:17: critical.

00:02:17: here though He connected language model to a live Revit model using the Model Context Protocol or MCP.

00:02:24: MCP, right?

00:02:25: Yeah and by doing that he essentially built a bridge that translates proprietary BIM geometry into text tokens that AI can actually comprehend.

00:02:34: so used it.

00:02:35: create productivity tab linking three DBIM data directly with lean construction planning.

00:02:41: Doing in single session is just wild.

00:02:42: I mean traditionally building custom plug-in would take development team what Months?

00:02:49: Easily, months at work.

00:02:50: But the real aha moment he shared wasn't just a code generation it was actual output.

00:02:55: Right

00:02:56: By connecting geometric model data to site level logistics you know like historical productivity rates and crew capacities The AI immediately flagged massive missed commitment

00:03:07: And this is before we even started.

00:03:09: right

00:03:09: Exactly This spatial data indicated that specific tasks required four hundred eighty-four man hours But the logistic schedule showed that the crew only had three hundred thirty six available hours.

00:03:19: Wow!

00:03:20: That is the holy grail of lean construction right there?

00:03:23: It really is...

00:03:24: it caught a fundamental mathematical impossibility before a single worker even stepped onto the job site, but I do think we need to bring some gravity into this.

00:03:32: Oh for sure It's not all magic.

00:03:34: No, because deploying this is not as simple as waving a magic wand.

00:03:39: Treating these models like omniscient mind readers gets incredibly expensive incredibly fast.

00:03:44: Yes the token burn.

00:03:45: Yeah.

00:03:46: Natalie Ortega ran into this exact wall when she was experimenting with that same model context protocol workflow.

00:03:52: She burned through four hundred sixty thousand tokens just trying to figure out how to prompt the system correctly

00:03:58: Which really highlights a brutally steep learning curve.

00:04:00: The technology is capable, but the token burn rate can be catastrophic if you don't understand how the model processes spatial data.

00:04:07: Exactly!

00:04:08: Because when you hook up a three-D model through MCP You aren't just sending a flat picture—you're serializing a massive geometric database

00:04:16: Right.

00:04:16: And her takeaway was that treating the AI like it inherently understands architectural intent is a massive mistake.

00:04:23: I mean, It's exactly like asking a human contractor to build a house without handing them a blueprint.

00:04:27: Yeah

00:04:27: you can't just say build-it right

00:04:28: and then getting furious when The kitchen ends up on the second floor.

00:04:32: She noted That Just sending an image and saying Build.

00:04:35: this led To endless super expensive followup queries.

00:04:38: Oh i bet yeah.

00:04:39: Levels shifted unpredictably Dimensions warped...I think she ended Up having to rebuild Structural columns three different times.

00:04:45: She was essentially paying a massive token tax for ambiguity.

00:04:50: I mean, the AI has a finite context window.

00:04:53: if you flood it with vague requests and entire building's worth of data It is absolutely going to hallucinate geometries.

00:05:01: Which is why her lessons here are incredibly pragmatic.

00:05:04: You have provide full unassailable brief upfront Exact dimensions Precise locations Specific materials

00:05:11: Give it boundaries

00:05:12: Exactly And most importantly, you have to stop putting the entire building into one single chat session.

00:05:18: The model does not need the specs for every first floor window.

00:05:20: when you're asking it design a roof dome.

00:05:23: Right?

00:05:23: You've got to isolate context separate sessions or separate elements.

00:05:27: that makes perfect sense from back end processing standpoint and keeps token usage lean

00:05:32: totally.

00:05:33: But if we pull back from the technical weeds for a second, what's really interesting is how this AI integration is disrupting the software market itself.

00:05:42: Oh!

00:05:42: The Revit Killer debate?

00:05:44: Exactly For years, the Holy Grail in contact has been finding a brand new platform to replace legacy BIM software.

00:05:51: but Nicholas Cotellier recently pointed out that BIM managers are actually pausing that hunt entirely

00:05:57: And I can completely understand why.

00:05:59: If you manage a design firm, You probably have decades worth of custom templates families and standardized workflows locked into that legacy software.

00:06:07: It's

00:06:08: all entrenched.

00:06:08: Yeah ripping it out is a logistical nightmare.

00:06:11: Precisely Because AI now allows users to completely bypass the clumsy outdated interface The urgency to replace it just vanishes...

00:06:20: You don't need new application!

00:06:21: You just need a new interface.

00:06:23: Spot

00:06:23: on The old software just becomes a stable historical database in the background.

00:06:28: The AI acts as your hyper-efficient conversational front end, you're talking directly to the data layer.

00:06:34: that flips the entire software model upside down but it also runs completely countered.

00:06:39: how the major vendors are trying to steer the market right now.

00:06:42: yes Autodesk being the prime example.

00:06:45: Right.

00:06:45: take Andrew Anagnos at Autodesc.

00:06:47: he recently stated Point blank that their new cloud and AI platform Forma is designed to entirely replace Revit over time.

00:06:55: a total replacement strategy.

00:06:56: Yeah, they're corporate vision isn't about bolting AI onto a legacy tool.

00:07:01: It's about building a continuous thread.

00:07:03: That connects early-stage design Directly to realtime execution.

00:07:07: because the painful reality Is that rich BIM data almost always dies?

00:07:11: The moment a design is handed over to the contractors

00:07:14: To the people who actually have to build it in the dirt

00:07:16: exactly

00:07:17: which brings us directly into the reality check currently hitting the industry.

00:07:21: The AI tools we just analyzed are undeniably fast, but their output is entirely reliant on quality and continuity of underlying data.

00:07:30: And AI cannot magically fix bad data?

00:07:33: No

00:07:34: it can't.

00:07:35: Altaf Ghaniha recently noted that the industry is currently sliding in to a trough of disillusionment where the Silicon Valley height meets the job site mud.

00:07:45: That's exactly what this is.

00:07:47: In a controlled environment, the demo looks flawless.

00:07:49: The AI pulls up a pristine model flags an issue and everyone in the boardroom cheers

00:07:54: Right high-fives all around.

00:07:55: but when you try to deploy that exact same AI agent on An actual moving project it completely falls apart because the underlying data feeding It is just a chaotic mess of disconnected PDFs and outdated schedules

00:08:09: And we are seeing the hard data to back up that hesitation.

00:08:12: A few leaders in this space, Renata Pinheiro and Jeff Sample.

00:08:16: they pointed to the Bluebeam AEC technology outlook for twenty-twenty seven.

00:08:20: Oh!

00:08:20: That report is

00:08:21: eye opening It really is.

00:08:23: it found only seven percent of surveyed firms have actually embedded AI across multiple core workflows

00:08:29: Just seven percent.

00:08:31: Seven percent.

00:08:32: So.

00:08:32: while LinkedIn might make it look like everyone is running fully autonomous projects, the truth is almost no one has fully operationalized this technology yet.

00:08:40: Because operationalizing AI at scale requires you to do the grueling unglamorous work of structuring your data first?

00:08:48: You can't just buy an AI license and expect it to organize your physical filing cabinets.

00:08:52: Definitely not.

00:08:54: And by the way, if you want to make sure your keeping up with these exact kinds of reality checks and shifts.

00:08:59: Make Sure You Subscribe To This Deep Dive.

00:09:01: We Sift Through The Noise So You Don't Have To Read A Hundred White Papers Every Week.

00:09:05: It Is Definitely The Easiest Way To Stay Calibrated.

00:09:09: Speaking Of How Things Actually Work.

00:09:11: Let's Look At The Mechanics Of These AI Deployments.

00:09:13: Let's Do it.

00:09:14: Matthew Climan Recently Broke Down The Six Fundamentally Different Types Of Construction AI Agents catch all term, but the capabilities are vastly different depending on this specific data pool they're allowed to swim in.

00:09:30: Let's map those out because that distinction really matters.

00:09:33: it completely dictates the output.

00:09:35: so you have agents only read documents

00:09:37: like

00:09:37: drawing specs and contracts.

00:09:39: okay

00:09:39: then you have Agents that plug into project systems reading ERPs and schedules.

00:09:44: You have visual agents analyzing the job site via camera feeds agents that read three D models agents that pull telemetry data from connected tools And finally, agents parsing through chat logs to track human decisions.

00:09:59: So the intelligence of AI is entirely bottlenecked by this specific bucket of data.

00:10:03: it has access too?

00:10:04: I mean...it doesn't know what you can see!

00:10:06: What's fascinating here Is that a highly fluent confident answer built on weak or disconnected evidence is still completely useless answers.

00:10:14: Garbage in garbage out!

00:10:16: Exactly!

00:10:17: Climb and use as brilliant example.

00:10:20: if you ask document reading agent whether the engineering spec requires a specific connection to be torqued to twenty-eight Newton meters, it will find that exact clause in three seconds.

00:10:30: Because its essentially just running a highly advanced reading comprehension test on a PDF?

00:10:35: Right

00:10:36: but if you ask that exact same agent If Connection Number Twenty Four was actually torqu'd to twenty eight newton meters On The Job site this morning well It has absolutely no idea

00:10:47: right!

00:10:47: Its totally blinding reality

00:10:49: because nothing In the documents contains That answer.

00:10:52: It cannot see the physical work, and it's not connected to digital torque wrench used by tradesmen.

00:10:58: So if you have an incredibly smart AI tool but feed it dumb or incomplete data... ...you are just engineering failure at a faster speed!

00:11:04: You get confident hallucinations?

00:11:06: Or this system breaks down entirely.

00:11:08: when asked scale Which perfectly aligns with conclusions Florian Humeur and Matthew Krebs shared.

00:11:14: They pointed out that simply flattening AEC data destroys inherent engineering dependencies.

00:11:19: Oh completely.

00:11:20: If you just take a bunch of raw PDF specs and IFC model in the spreadsheet, dump them all into an AI without building a structured layer?

00:11:28: You strip away all context.

00:11:30: They noted that ninety percent teams are going straight from RAW project files to an AI interface

00:11:35: Which is literal definition for what we said.

00:11:38: Garbage In, garbage out!

00:11:40: You have build unified knowledge layers.

00:11:42: Yeah

00:11:42: If structural wall has an acoustic requirement buried in pdf and a soil bearing capacity requirement in the geotechnical report, both of those data points need to be hard-linked.

00:11:53: To

00:12:00: put it another way flattening project data into an AI is like taking a fully assembled car engine throwing It Into A wood chipper putting The metal shavings In a box And handing it to a mechanic saying start the car.

00:12:11: that Is a terrifying image but very accurate

00:12:15: right?

00:12:15: All the materials are technically in the box, but because you destroyed the dependencies it's completely useless.

00:12:22: And because contractors are realizing that trying to rip and replace massive legacy systems is failing we are seeing a violent pivot in venture capital strategy.

00:12:32: The funding is definitely shifting.

00:12:34: Yeah contact investors are finally accepting that software fragmentation Is just a permanent reality.

00:12:40: They really are.

00:12:41: Maytosh Shoron observed at two companies adaptive and scaffold, just raised forty five million dollars combined.

00:12:48: Both of those startups are operating on the exact same commercial thesis

00:12:51: which is

00:12:52: that construction companies absolutely refuse to do rip-and-replace software implementations anymore.

00:12:57: I mean if you're managing a half billion dollar infrastructure project You cannot afford to halt operations for six months Just to train five hundred subcontractors on any dashboard.

00:13:08: The risk profile isn't sane

00:13:09: exactly.

00:13:11: So instead of trying to be the one platform, rule them all.

00:13:15: these new tools are designed as layers.

00:13:17: Okay integration layers

00:13:19: yeah they sit on top and integrate directly with The existing ERPs.

00:13:23: in localized workflows.

00:13:24: They become the connective tissue between disparate systems rather than forcing a heavy-handed mandate To migrate everything.

00:13:31: And Carter Wittenregg shared an anecdote that perfectly captures why this integration layer is so desperately needed.

00:13:37: He spoke with a construction company doing over one hundred million dollars per year in revenue, and they are still using Excel spreadsheets to build their final

00:13:46: bids.".

00:13:54: And it isn't because they're trying to save money on licenses.

00:13:57: They spend hundreds of thousands a year on top-tier software, but because those walled gardens refuse to talk to each other natively the employees are forced to act as human API.

00:14:07: A Human API?

00:14:08: I love that phrase.

00:14:09: That's true!

00:14:10: A senior estimator has look at one screen export a CSV manually format in Excel and then key same data into project management tool.

00:14:19: another screen

00:14:20: Just burning hours.

00:14:23: So the real value for AI isn't in creating another standalone application.

00:14:28: It's being invisible integration layer that automatically translates data between those systems.

00:14:34: And if you follow where venture capital is moving, You see exactly where market believes massive returns will be.

00:14:40: Alexei Gutev analyzed a hundred and eighty contact companies and their funding rounds across twenty-twenty five and twenty-two six.

00:14:47: What did he find?

00:14:48: Well out of thirty startups successfully raised capitol Thirteen of them operate directly on the physical job site.

00:14:55: So they are abandoning The Office entirely?

00:14:58: They aren't.

00:14:59: and to contrast that if you look at startups focused On design code and regulatory compliance purely office-based software Nine out of fourteen Of those companies Are either totally unfunded or permanently stuck in the seed stage.

00:15:14: Wow, yeah the venture capital isn't flowing To the desk anymore.

00:15:17: it is flowing straight to the dirt.

00:15:20: And Fabio Bronzino breaks down the math on exactly why investors are shifting their focus there.

00:15:25: On any given construction project, The total budget spent on software and IT is roughly one to three percent.

00:15:32: A tiny fraction.

00:15:33: Yeah.

00:15:34: If you're building a new office-based contact app You are engaging in a bloodbath To capture a fraction of a three per cent sliver.

00:15:43: It's an incredibly crowded market fighting over essentially the margin of error under Project Budget

00:15:48: Exactly The massive prize, the other ninety-seven percent of the capital is tied up in physical materials, physical labor and heavy equipment.

00:15:59: That's where actual money changes hands.

00:16:01: And where real risks are?

00:16:03: Right!

00:16:03: Where the schedule bleeds most capital through rework and delays

00:16:08: Which transitions us perfectly into our final look at this all heading.

00:16:13: If a real financial price is tied to labour or material then the technology layer has to physically touch the dirt.

00:16:21: Contact is rapidly getting physical.

00:16:24: We're moving from digital software integration, to physical task automation.

00:16:29: But let's be clear we aren't talking about humanoid sci-fi robots walking around framing walls.

00:16:34: The pragmatic reality is much more grounded.

00:16:36: It has Ivo van Bruckelen delivered a very stark reality check regarding robotics.

00:16:42: There is an enormous amount of hype around deploying thirty thousand dollar general-purpose humanoid robots on sites.

00:16:48: Sure, they look great on video

00:16:50: They do.

00:16:51: but alive construction site as muddy chaotic unstructured and constantly shifting.

00:16:56: Trying to make a bipedal robot navigate that environment requires such a dense array of sensors in computing power That it pushes the ROI out past seven years

00:17:04: which makes it a non starter.

00:17:05: yeah No general contractor is gonna swallow a seven-year payback period on a piece of hardware that might fall into a trench on day three, right?

00:17:14: So the massive margin explosion isn't in general purpose humanoids.

00:17:17: It's happening in highly specialized task specific automation.

00:17:21: Okay like what

00:17:22: we're talking about rugged machines designed to execute one specific high cost bottleneck tasks flawlessly.

00:17:28: autonomous drywall finishing bots.

00:17:31: Autonomous rebar tying machines.

00:17:33: Yeah, Aaron Love highlighted OSNAP layout which is a perfect example of this.

00:17:37: They use an autonomous robotic layout rover that takes the digital BIM plan and prints out lines directly onto physical concrete slab.

00:17:46: It drives around translating the model into reality?

00:17:48: Exactly!

00:17:49: And that eliminates the compounding human error with a tape measure making ¼ inch mistake on first floor.

00:17:55: that magnifies to massive structural problem by third-floor.

00:17:59: And we're seeing that task-specific automation scale up to heavy civil machinery, too.

00:18:04: Haley Swords pointed out a few major moves in this space.

00:18:08: You have companies like Bedrock Robotics Retrofitting existing excavators with sensor arrays.

00:18:13: make them fully autonomous.

00:18:14: That is massive for infrastructure projects

00:18:16: Huge!

00:18:18: On the structural side Icon's Titan system Is bringing commercial scale three D printed buildings closer every day.

00:18:23: reality.

00:18:25: These are no longer just slick tech demos.

00:18:27: They're moving toward actual commercial deployment.

00:18:30: But I want to pull this all the way down to the absolute ground floor because you don't need an autonomous excavator, to radically change your data capture.

00:18:39: Christian Reed pointed out a friction point that is so basic we often completely ignore

00:18:43: it.

00:18:43: What's up?

00:18:44: He noted even adopting something as simple as a Bluetooth connected tape measure Is a transformative step for field data.

00:18:50: Oh!

00:18:51: Because It completely removes The human from the data entry loop.

00:18:54: Exactly Think about the manual process of measuring a cut.

00:18:58: You read the tape, you write that number on a piece of scrap wood and later someone types it into your system.

00:19:04: The human API again?

00:19:05: Right!

00:19:05: A Bluetooth tape measure eliminates the Read, Write & Type steps instantly.

00:19:10: by automating data entry at an exact moment in measurement.

00:19:13: It prepares the field for AI by passively collecting clean data right from source.

00:19:18: That makes total sense.

00:19:19: If you eliminate that micro-friction two hundred times a day across an entire framing crew, You've just unlocked massive workflow efficiency without requiring anyone to learn in your software platform.

00:19:31: And capturing clean data from the field connects perfectly with a critical point made by Stefan Tronberger regarding on site monitoring.

00:19:38: He noted engineering calculations and digital models only give you theoretical plans

00:19:43: Paper vs reality Exactly

00:19:45: because reality is highly dynamic.

00:19:48: Take the pressure of fresh concrete during a pour, for instance.

00:19:50: Okay

00:19:51: The actual physical placement rate unexpected ambient temperature changes this specific behavior of the chemical mix.

00:19:58: They all dynamically alter the actual pressure.

00:20:00: acting against that physical form work in real time.

00:20:03: Yeah A static structural model and office cannot predict a sudden ten degree temperature drop on the site

00:20:08: Precisely Which is why you have to live onsite monitoring To tell system what's actually happening.

00:20:14: The physical measurement is what anchors the digital twin to reality.

00:20:19: You cannot rely on the pristine digital model alone.

00:20:22: It needs a continuous physical feedback loop to be accurate.

00:20:25: Yes, it's all about closing the loop between the theoretical geometry On-the-screen and the physical reality of the concrete in steel.

00:20:31: And as we close out this deep dive that concept of a closed loop brings up A deeply provocative question for you to mull over this week.

00:20:38: I love good hypothetical

00:20:40: well.

00:20:40: As these physical sensors and digital models become seamlessly connected The boundary between the physical job site and digital engineering file starts to disappear entirely.

00:20:50: Okay

00:20:51: Take Hexagon's recent, five hundred thirty-five million dollar acquisition of rock science.

00:20:55: Their stated goal is feed live radar and physical sensor data from the site directly back into their geotechnical models.

00:21:02: So the Geotechnical Model is actively updating and recalculating itself based on the live data streaming in for dirt?

00:21:09: Exactly

00:21:10: Okay.

00:21:10: We are rapidly approaching an era of fully self-calibrating engineering models, but consider the massive legal and professional implications for that shift.

00:21:20: Oh wow yeah.

00:21:21: When AI agent autonomously updates a load bearing structural model based on real time sensor data who legally signs off on this recommendation?

00:21:30: That completely shatters traditional chain liability.

00:21:33: It does

00:21:33: If the model autonomously adjusts required depth to foundation pile because a sensor detected an anomaly.

00:21:41: Does the human structural engineer remain the author of that design?

00:21:45: Or do they simply become a reviewer of AI's autonomous

00:21:48: work?".

00:21:49: That is...a terrifying thought for an engineer!

00:21:53: And heaven forbid, if that foundation eventually fails who holds the ultimate liability?

00:21:58: Is it the software vendor who wrote the

00:22:00: A.I.?

00:22:01: Is it manufacturer or physical sensor that fed data Or is it the engineer of record who trusted this self-calibrating loop?

00:22:08: That's the multi million dollar question that industry has to answer as we transition from static digital files, to autonomous physical workflows.

00:22:17: If you enjoyed this episode new episodes drop every two weeks.

00:22:20: also check out our other additions on smart manufacturing and industrial AI And connected tools in equipment.

00:22:26: Thanks for listening everyone.

00:22:27: Thank you so much for joining us on this deep dive.

00:22:29: And don't forget to subscribe,