Latest / Mind Cast / The AI Reckoning | Where Billions Vanish and What Actually Delivers

Transcript

The AI Reckoning | Where Billions Vanish and What Actually Delivers · Mind Cast

• 0:00
  SPEAKER_00: Here is a number I want you to sit with for just a second.
• 0:04
  SPEAKER_00: $684 billion.
• 0:07
  SPEAKER_00: That is how much was invested globally in enterprise AI in 2025 alone.
• 0:13
  SPEAKER_00: It is a figure so large it almost loses meaning.
• 0:18
  SPEAKER_00: So let me make it meaningful.
• 0:20
  SPEAKER_00: Of that$684 billion, over$547 billion delivered zero.
• 0:27
  SPEAKER_00: I mean zero measurable return.
• 0:32
  SPEAKER_00: Not a disappointing return.
• 0:34
  SPEAKER_00: Not a modest return.
• 0:36
  SPEAKER_00: Zero.
• 0:37
  SPEAKER_00: Nothing.
• 0:38
  SPEAKER_00: Gone.
• 0:39
  SPEAKER_00: An MIT study found 95% of generative AI enterprise pilots failed to produce any trackable profit and loss impact whatsoever.
• 0:49
  SPEAKER_00: Meanwhile, 42% of U.S.
• 0:52
  SPEAKER_00: enterprises abandoned their broad AI initiatives in 2025, up from just 17% the year prior.
• 1:01
  SPEAKER_00: So here is my question.
• 1:03
  SPEAKER_00: If AI is the most transformative technology of our generation, why is the corporate world hemorrhaging money into it at a scale that would make any rational investor reach for the exit?
• 1:15
  SPEAKER_00: That is what we are going to find out, and the answer will completely reshape how you think about where AI actually works and where it is simply the most expensive emperor's new clothes in business history.
• 1:30
  SPEAKER_00: Welcome to Mindcast.
• 1:32
  SPEAKER_00: I'm your host Will.
• 1:33
  SPEAKER_00: And if you are new here, welcome.
• 1:35
  SPEAKER_00: This show is for curious minds who want to go deeper than the headlines.
• 1:40
  SPEAKER_00: We don't do hype, we don't do buzzwords, we chase insight and follow the evidence wherever it leads.
• 1:47
  SPEAKER_00: Today's deep dive is based on exclusive research that is not publicly available anywhere else.
• 1:53
  SPEAKER_00: So let's get into it.
• 1:57
  SPEAKER_00: Key insight number one.
• 1:59
  SPEAKER_00: Why enterprise AI is burning billions.
• 2:03
  SPEAKER_00: Enterprise AI projects fail at a rate of 80 to 90%.
• 2:08
  SPEAKER_00: For context, standard non-AIIT projects fail at roughly 40 to 50%.
• 2:14
  SPEAKER_00: So we have a technology sold as the future of everything failing at double the rate of the very infrastructure it is supposed to replace.
• 2:24
  SPEAKER_00: The answer lies in three structural failures the research identifies with remarkable clarity, not random misfortunes, they are repeatable, predictable, and almost entirely avoidable if you know what to look for.
• 2:39
  SPEAKER_00: The first structural failure is what researchers at Rand describe as top-down mandates without clear problem definitions.
• 2:47
  SPEAKER_00: Picture this: a CEO reads a breathless profile of a competitor deploying AI.
• 2:52
  SPEAKER_00: The board gets nervous.
• 2:54
  SPEAKER_00: A mandate comes down from the executive suite.
• 2:56
  SPEAKER_00: We need to be an AI company.
• 2:59
  SPEAKER_00: Hundreds of millions of dollars get allocated.
• 3:01
  SPEAKER_00: Consultants arrive, slide decks proliferate, and somewhere in all of that, nobody asks the one question that actually matters.
• 3:09
  SPEAKER_00: What specific problem are we solving?
• 3:12
  SPEAKER_00: AI is not a strategy, it is a tool.
• 3:15
  SPEAKER_00: And deploying a tool without defining the problem it solves is like hiring a world-class surgeon and saying, just do something.
• 3:23
  SPEAKER_00: The results are predictably catastrophic.
• 3:27
  SPEAKER_00: The second structural failure is catastrophic data unreadiness.
• 3:32
  SPEAKER_00: Gartner shows 85% of AI projects fail because of poor data quality, not poor AI models, poor data, and only 12% of organizations have AI-ready data architecture.
• 3:47
  SPEAKER_00: 88% of enterprises are trying to run a Formula One race car on a dirt road.
• 3:54
  SPEAKER_00: The research includes a case study that is genuinely staggering.
• 3:58
  SPEAKER_00: A Fortune 500 manufacturer invested$12 million in an AI initiative.
• 4:04
  SPEAKER_00: Partway through, they discovered 47 legacy data formats and 23 disconnected databases that could not talk to each other.
• 4:14
  SPEAKER_00: Fixing just the data infrastructure before a single line of AI code could function required an additional$8 million.
• 4:24
  SPEAKER_00: They canceled the entire project.
• 4:27
  SPEAKER_00: $20 million.
• 4:29
  SPEAKER_00: Gone.
• 4:30
  SPEAKER_00: Not because the AI was bad, because nobody checked whether the foundation could support the building before construction began.
• 4:38
  SPEAKER_00: The third structural failure is what the research calls the open loop problem, perhaps the most subtle, but in many ways the most damaging.
• 4:46
  SPEAKER_00: Generative AI produces probabilistic outputs.
• 4:50
  SPEAKER_00: It gives you a statistically plausible answer.
• 4:53
  SPEAKER_00: In an unstructured workflow where humans have to eyeball every recommendation and decide whether to act on it, that creates an almost impossible cognitive burden.
• 5:04
  SPEAKER_00: The result?
• 5:06
  SPEAKER_00: Frontline workers stop trusting it.
• 5:08
  SPEAKER_00: 67% of failed enterprise AI projects cite frontline worker resistance as the primary barrier to adoption.
• 5:16
  SPEAKER_00: In one documented case, warehouse managers overrode AI recommendations 78% of the time.
• 5:23
  SPEAKER_00: At that point, you are just adding an expensive suggestion box that everyone ignores.
• 5:28
  SPEAKER_00: So that is the anatomy of failure.
• 5:31
  SPEAKER_00: No clear problem definition, catastrophically unready data, and an open loop that breeds organizational resistance.
• 5:38
  SPEAKER_00: Three causes, 80 to 90% failure rates, half a trillion dollars evaporated.
• 5:47
  SPEAKER_00: Key insight number two, the hidden side effect.
• 5:51
  SPEAKER_00: AI is breaking software itself.
• 5:54
  SPEAKER_00: Now I want to pivot to something I think is the most counterintuitive finding in this entire body of research.
• 6:00
  SPEAKER_00: The report calls it the developer productivity paradox.
• 6:04
  SPEAKER_00: Here is the surface level story, and it sounds great.
• 6:08
  SPEAKER_00: AI coding tools have reduced pull request lead times by 32.4%.
• 6:14
  SPEAKER_00: Individual developers are writing code faster than ever.
• 6:17
  SPEAKER_00: On a dashboard measuring individual throughput, everything looks fantastic.
• 6:22
  SPEAKER_00: But here is what the dashboard does not show you.
• 6:26
  SPEAKER_00: Longitudinal research across 211 million lines of code tells a very different story at the team level.
• 6:35
  SPEAKER_00: Code churn is up 139%.
• 6:38
  SPEAKER_00: Code duplication has increased by 800%.
• 6:42
  SPEAKER_00: 12.3% of all AI-generated code is a duplicate of something already in the code base.
• 6:49
  SPEAKER_00: Refactoring activity is down 39.9%, meaning developers are generating code but not maintaining the health of the systems it lives inside.
• 6:59
  SPEAKER_00: Technical debt is growing 30 to 41% faster.
• 7:04
  SPEAKER_00: And in a landmark study by METR, developers on complex real-world tasks were 19% less productive when using AI tools than without them.
• 7:17
  SPEAKER_00: Think about what this means.
• 7:19
  SPEAKER_00: AI is making individual developers faster at writing code and simultaneously making the overall systems they build slower, messier, more fragile, and more expensive to maintain.
• 7:32
  SPEAKER_00: The illusion of productivity at the cost of long-term system integrity.
• 7:38
  SPEAKER_00: The root cause?
• 7:40
  SPEAKER_00: Architectural blindness.
• 7:42
  SPEAKER_00: AI completes local patterns within its immediate context, but cannot see the global system, like a builder who lays bricks incredibly fast but has never seen the blueprints.
• 7:54
  SPEAKER_00: Every brick looks perfect, but the wall is crooked, and nobody noticed because they were watching the bricklaying speed.
• 8:04
  SPEAKER_00: Here is where the research makes a crucial distinction.
• 8:07
  SPEAKER_00: There is a domain where AI works brilliantly in software, specialized systems engineering.
• 8:13
  SPEAKER_00: In environments using model-based design, hardware in the loop simulation, or tools like MATLAB and Simulink, AI performs with striking reliability.
• 8:24
  SPEAKER_00: Physics-based equations mathematically constrain what the AI can output.
• 8:29
  SPEAKER_00: Hallucinations are physically impossible when the system is governed by thermodynamics or fluid dynamics.
• 8:35
  SPEAKER_00: The feedback is either correct or incorrect by definition.
• 8:39
  SPEAKER_00: No ambiguity.
• 8:41
  SPEAKER_00: Open loop versus closed loop, language governed versus physics governed, the single most important conceptual dividing line in all of applied AI today.
• 8:54
  SPEAKER_00: Key insight number three.
• 9:08
  SPEAKER_00: Industrial AI, the factory floor.
• 9:11
  SPEAKER_00: The key difference: industrial AI ingests deterministic physical inputs, three-dimensional laser scans, X-ray photon attenuation data, load cell signals, acoustic fast Fourier transform spectra.
• 9:26
  SPEAKER_00: Inputs governed by physical laws that don't bend to human interpretation.
• 9:31
  SPEAKER_00: Because the feedback is physical, outputs are reliable, auditable, and directly traceable to hard business metrics.
• 9:41
  SPEAKER_00: Let me give you some concrete examples because the numbers here are remarkable.
• 9:46
  SPEAKER_00: Start with Ishida's multi-head weyer systems.
• 9:50
  SPEAKER_00: These are AI-driven precision weighing platforms used in food manufacturing.
• 9:55
  SPEAKER_00: Before AI integration, product giveaway, the excess product packed above the target weight, which the manufacturer absorbs as a direct cost, ran between 2 and 5% of total throughput.
• 10:08
  SPEAKER_00: After AI integration, giveaway dropped to within 0.2 grams per pack.
• 10:16
  SPEAKER_00: 0.2 grams.
• 10:18
  SPEAKER_00: Think about what that means at industrial scale.
• 10:21
  SPEAKER_00: Throughput improved by 30 to 40%.
• 10:24
  SPEAKER_00: Energy consumption dropped by 20%.
• 10:27
  SPEAKER_00: This is AI paying for itself in weeks, not years.
• 10:31
  SPEAKER_00: Then there's Merrill's Sensor X in poultry processing, combining X-ray imaging with three-dimensional laser scanning to perform what used to be a manual labor-intensive inspection process.
• 10:43
  SPEAKER_00: 70 breast caps per minute, 20% capacity boost, manual rework substantially eliminated, a quality bottleneck turned into a quality guarantee.
• 10:55
  SPEAKER_00: In salmon processing, a sector with notoriously tight margins, Botter and Marilec AI systems have pushed consistent fillet yields from the traditional range of 78 to 82% up to 85 to 88%, precision within plus or minus two grams across every single fillet.
• 11:16
  SPEAKER_00: The financial impact is captured in multi-million dollar annual margin improvements per facility.
• 11:22
  SPEAKER_00: And the environmental benefit?
• 11:24
  SPEAKER_00: A 35% reduction in global seafood waste.
• 11:28
  SPEAKER_00: That is a measurable consequence of deterministic physics-governed AI making better decisions than humans can at speed.
• 11:37
  SPEAKER_00: And then there is predictive maintenance.
• 11:40
  SPEAKER_00: Augury's industrial AI platform uses vibration and acoustic analysis to detect bearing wear weeks before the bearing actually fails.
• 11:51
  SPEAKER_00: Weeks before failure.
• 11:53
  SPEAKER_00: Maintenance teams schedule downtime on their terms, not scramble to respond to a catastrophic breakdown.
• 12:00
  SPEAKER_00: The result is a 30 to 50% reduction in unplanned downtime.
• 12:05
  SPEAKER_00: The economics are stark.
• 12:07
  SPEAKER_00: Avoiding just two unexpected breakdowns per year generates over$1 million in savings.
• 12:14
  SPEAKER_00: The software costs between$200,000 and$600,000 annually.
• 12:20
  SPEAKER_00: Industrial analytics platforms like Site Machine and Odin Technologies are delivering 12 to 22% improvements in overall equipment effectiveness, 10 to 25% reductions in scrap, and up to 40% cuts in downtime costs, at scale, measurable results.
• 12:41
  SPEAKER_00: On the factory floor, industrial AI is not eliminating workers, it is upgrading them.
• 12:47
  SPEAKER_00: Workers who spent shifts on grueling manual tasks now supervise AI systems through HMI touchscreen interfaces.
• 12:56
  SPEAKER_00: Their tacit knowledge, how humidity affects dough viscosity, how a specific raw material batch behaves under heat gets embedded directly into the AI model.
• 13:07
  SPEAKER_00: These operators become model calibrators, the human intelligence layer that makes the AI smarter over time.
• 13:14
  SPEAKER_00: That is not displacement, that is evolution.
• 13:18
  SPEAKER_00: Better outcomes for the worker and the business simultaneously.
• 13:24
  SPEAKER_00: The synthesis, the reality check framework.
• 13:28
  SPEAKER_00: So, what do we do with all of this?
• 13:30
  SPEAKER_00: Let me pull all of these threads together and give you something concrete and actionable.
• 13:35
  SPEAKER_00: The research distills the conditions for high ROI AI into five core principles.
• 13:41
  SPEAKER_00: Think of these as your field guide for evaluating any AI investment.
• 13:46
  SPEAKER_00: Principle 1, physical telemetry grounding.
• 13:50
  SPEAKER_00: AI delivers consistent, auditable ROI when its inputs are governed by physical laws, not open-ended human language.
• 13:59
  SPEAKER_00: The moment your AI system is processing laser scan data or acoustic spectra, you have closed the door on hallucination.
• 14:07
  SPEAKER_00: The moment it is processing natural language in an unstructured environment, that door swings wide open.
• 14:14
  SPEAKER_00: This is not a hard and fast rule about generative AI being useless, it is a precise observation about where the physics of the problem protects you from the probabilistic unreliability of the model.
• 14:26
  SPEAKER_00: Principle 2.
• 14:27
  SPEAKER_00: Closed loop operational context.
• 14:31
  SPEAKER_00: AI recommendations that sit on a dashboard get ignored.
• 14:36
  SPEAKER_00: Full stop.
• 14:37
  SPEAKER_00: The research is conclusive on this.
• 14:39
  SPEAKER_00: If a human being has to see the recommendation, evaluate it, and then manually implement it, you will get resistance, delay, inconsistency, and eventual abandonment.
• 14:50
  SPEAKER_00: High ROI Industrial AI writes its adjustments directly back to the machines, PLCs, servo drives, way gates.
• 14:59
  SPEAKER_00: The loop is closed, the system acts, the human monitors.
• 15:03
  SPEAKER_00: That is the model that works.
• 15:05
  SPEAKER_00: Principle 3.
• 15:06
  SPEAKER_00: Edge first micro latency architecture.
• 15:10
  SPEAKER_00: Cloud AI introduces latency, and on a production line where physical processes are happening in real time, latency is not just inefficiency, it is a fundamental incompatibility with the task.
• 15:23
  SPEAKER_00: Sub-10 millisecond inference delivered at the edge via neural processing units is what enables AI to participate in physical control loops.
• 15:33
  SPEAKER_00: If your AI requires a round trip to the cloud to make a decision, it cannot keep up with the physics.
• 15:39
  SPEAKER_00: This is not a technical preference, it is a physical requirement.
• 15:43
  SPEAKER_00: Principle 4.
• 15:45
  SPEAKER_00: Direct coupling to hard PL metrics.
• 15:49
  SPEAKER_00: This one is simple and it is ruthless.
• 15:52
  SPEAKER_00: If the AI's output cannot be audited on a monthly profit and loss statement, do not fund it.
• 15:59
  SPEAKER_00: Define the metric before you build anything.
• 16:02
  SPEAKER_00: Reduce product giveaway from 2.5% to 0.2%.
• 16:07
  SPEAKER_00: Reduce unplanned downtime by 30%.
• 16:10
  SPEAKER_00: Increase fillet yield from 80% to 87%.
• 16:14
  SPEAKER_00: Concrete, measurable, auditable.
• 16:17
  SPEAKER_00: If you cannot write the success metric in those terms before you start, you are not ready to invest.
• 16:23
  SPEAKER_00: Principle 5.
• 16:24
  SPEAKER_00: Frontline Augmented Orchestration.
• 16:28
  SPEAKER_00: Involve the frontline workers from day one, not as an afterthought.
• 16:33
  SPEAKER_00: Not in a change management session three weeks before go live.
• 16:36
  SPEAKER_00: From day one.
• 16:38
  SPEAKER_00: Their tacit knowledge is not a nice to have.
• 16:41
  SPEAKER_00: It is mission critical data.
• 16:43
  SPEAKER_00: And when workers understand that the AI is designed to eliminate physical strain, not eliminate them, the resistance dynamic flips entirely.
• 16:52
  SPEAKER_00: 67% of enterprise AI projects fail because of frontline resistance.
• 16:57
  SPEAKER_00: The mirror image is that organizations who genuinely co-design AI with frontline workers report dramatically higher adoption and dramatically better outcomes.
• 17:07
  SPEAKER_00: Five principles use them as a filter for every AI investment conversation you have from this point forward.
• 17:16
  SPEAKER_00: Let me bring this home.
• 17:26
  SPEAKER_00: Top-down mandates without problem definitions, catastrophically unready data, open loop architectures that breed resistance, not the failures of bad technology, the failures of bad deployment.
• 17:40
  SPEAKER_00: But on the factory floor, in the way gates, X-ray scanners, acoustic sensors, and bearing monitors, AI is delivering genuine, measurable miracles, physics-governed inputs, closed loop control, workers more skilled and safer than before.
• 17:56
  SPEAKER_00: Salmon yields up nine percentage points, unplanned downtime cut in half, product giveaway at a fraction, documented outcomes.
• 18:07
  SPEAKER_00: The difference between those two worlds is not the sophistication of the AI.
• 18:12
  SPEAKER_00: It is the presence or absence of physical grounding, closed loop control, and a clear, audifable success metric defined before a single dollar is spent.
• 18:22
  SPEAKER_00: The question is not should we use AI?
• 18:24
  SPEAKER_00: That ship is sailed.
• 18:26
  SPEAKER_00: The question is, are we deploying it where physics protects us from its failures and where the feedback loop is closed by something more reliable than human subjectivity?
• 18:36
  SPEAKER_00: If you can answer yes, you are in the 20% actually getting ROI.
• 18:41
  SPEAKER_00: If you cannot, you now know exactly what to fix.
• 18:46
  SPEAKER_00: A final note: the research in this episode is exclusive to Minecast and not publicly available anywhere else.
• 18:53
  SPEAKER_00: If this episode made you think, subscribe, leave a review, and send it to one person who needs to hear it.
• 18:59
  SPEAKER_00: A colleague about to green light an AI project, a founder deciding where to invest.
• 19:04
  SPEAKER_00: It could save them from a very expensive mistake.
• 19:07
  SPEAKER_00: I'm Will, this is Minecast.
• 19:10
  SPEAKER_00: Stay curious, stay rigorous, and I'll see you in the next one.