<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Henry Costa on Medium]]></title>
        <description><![CDATA[Stories by Henry Costa on Medium]]></description>
        <link>https://medium.com/@henry.costa?source=rss-e2cc0bd3556a------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*uXh4avzdUJGhrKB5W7_gLw@2x.jpeg</url>
            <title>Stories by Henry Costa on Medium</title>
            <link>https://medium.com/@henry.costa?source=rss-e2cc0bd3556a------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 22:36:38 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@henry.costa/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Industrial DataOps. Turning Plant Data Into Trusted Data]]></title>
            <link>https://medium.com/@henry.costa/industrial-dataops-turning-plant-data-into-trusted-data-e552c3a0bcab?source=rss-e2cc0bd3556a------2</link>
            <guid isPermaLink="false">https://medium.com/p/e552c3a0bcab</guid>
            <category><![CDATA[unified-namespace]]></category>
            <category><![CDATA[industrial-dataops]]></category>
            <category><![CDATA[un]]></category>
            <category><![CDATA[dataops]]></category>
            <category><![CDATA[iiot]]></category>
            <dc:creator><![CDATA[Henry Costa]]></dc:creator>
            <pubDate>Tue, 19 May 2026 20:28:18 GMT</pubDate>
            <atom:updated>2026-05-19T20:35:00.338Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xlmE6b6uaoBZdU1i" /></figure><p>Most manufacturing teams I’ve worked with have the same quiet problem.</p><p>They can connect machines. They can collect data. They can even stream it to a data lake. But when someone tries to <em>use</em> it. For dashboards, OEE, anomaly detection, quality investigations, AI. The whole thing turns into a messy mix of “which tag is correct?”, “why did it change?”, and “who owns this data anyway?”</p><p>That’s where <strong>Industrial DataOps</strong> comes in.</p><p>Not as a new tool. Not as a buzzword. As the missing operating layer between your shop floor and your data lake.</p><h3>The pattern I’ve seen over and over</h3><p>A typical story looks like this:</p><p>We connect equipment and systems across a big manufacturing site. PLCs, SCADA, historians, MES, lab systems. We set up a modern pipeline and start streaming data up to cloud storage and analytics platforms.</p><p>At first it feels like progress. Data is flowing. People celebrate.</p><p>Then reality hits.</p><ul><li>A key tag changes after a controls update. Nobody tells the analytics team.</li><li>The same pump exists under three different names in three systems.</li><li>A line stops for 30 minutes. SCADA shows one reason. MES shows another. The historian shows… nothing useful.</li><li>A data scientist asks, “Which temperature is product temperature?” and five people argue for two weeks.</li></ul><p>The data lake is full. But trust is empty.</p><p>So we end up doing what we always do. We build a bunch of one-off fixes.</p><p>A script here. A mapping file there. A custom dashboard logic layer. A manual spreadsheet for “golden tags.” It works. Until it doesn’t.</p><p>Anyway, that’s the gap Industrial DataOps is meant to close.</p><h3>What Industrial DataOps actually is</h3><p>Industrial DataOps is the discipline of making industrial data:</p><ul><li><strong>reliable</strong></li><li><strong>traceable</strong></li><li><strong>consistent</strong></li><li><strong>deployable</strong></li><li><strong>governed</strong></li></ul><p>From the moment it leaves a machine or system. To the moment it shows up in analytics, reporting, or AI.</p><p>If you already know DevOps, it’s a similar idea. But applied to industrial data products.</p><p>Not “move fast and break things.” More like “move carefully and don’t break the batch record.”</p><h3>The 5 things Industrial DataOps controls (in real life)</h3><p>Here’s what I typically focus on when putting this in place.</p><h4>1. Naming and structure. The Unified Namespace problem</h4><p>If your plant data doesn’t have consistent structure, every downstream consumer has to rebuild context from scratch.</p><p>I’ve seen teams spend months building analytics logic, only to realize the real issue was that:</p><ul><li>assets were named differently by each integrator</li><li>tags were inconsistent across lines</li><li>site standards existed, but weren’t enforced</li></ul><p>A <strong>Unified Namespace (UNS)</strong> helps. But only if it’s treated like a product with rules and ownership, not a diagram in a slide deck.</p><p>Industrial DataOps is how you keep the UNS clean over time.</p><h4>2. Data quality checks. Not after the fact</h4><p>Most teams detect bad data when somebody complains. That’s too late.</p><p>DataOps puts checks in the pipeline:</p><ul><li>missing values</li><li>out-of-range values</li><li>flatlined signals</li><li>timestamp issues</li><li>duplicated events</li><li>unit mismatches</li></ul><p>And yes. This sounds boring. It is boring. It’s also the difference between “we have data” and “we can use data.”</p><h4>3. Version control and change management for data structures</h4><p>This is the one that makes some people uncomfortable.</p><p>My honest opinion. If you can change a tag, a UDT, a topic structure, or a calculation in production with no traceability. You don’t have a data platform. You have a science experiment.</p><p>In regulated environments, it’s even more serious. You need a clear story for:</p><ul><li>what changed</li><li>why it changed</li><li>who approved it</li><li>when it was deployed</li><li>what it impacted</li></ul><p>Industrial DataOps borrows from software practices. Git-based versioning, controlled promotion between environments, structured release notes. With the right level of validation and review.</p><h4>4. Context and semantics. The part data lakes can’t guess</h4><p>A data lake stores bytes. It doesn’t understand your process.</p><p>So when you send raw signals upstream without context, you force every consumer to rebuild meaning.</p><p>DataOps builds and maintains the “semantic layer,” things like:</p><ul><li>asset identity and hierarchy</li><li>equipment state models</li><li>product and batch context</li><li>good signal definitions</li><li>event definitions, with consistent logic</li></ul><p>This is where a lot of value sits. It’s also where a lot of projects quietly fail, because everyone assumes “we’ll add context later.”</p><p>Later never comes.</p><h4>5. Ownership. Who is on the hook when it breaks</h4><p>If nobody owns the pipeline end-to-end, the pipeline will break end-to-end.</p><p>Industrial DataOps makes ownership explicit:</p><ul><li>Who owns the naming standard?</li><li>Who owns the data product for Line 3 packaging?</li><li>Who owns the connector to the historian?</li><li>Who is responsible for the “gold” dataset used in reporting?</li></ul><p>Without this, every issue becomes a blame game between OT, IT, automation, and analytics. I’ve watched that movie too many times.</p><h3>A small example that shows why this matters</h3><p>One time we were trying to build a simple KPI dashboard. Nothing fancy.</p><p>Just availability and a couple downtime reasons.</p><p>The data existed in three places:</p><ul><li>SCADA had state changes</li><li>MES had production events</li><li>Historian had tag history</li></ul><p>The dashboard looked wrong. People argued. Operators didn’t trust it.</p><p>After digging in, the issue wasn’t “bad analytics.” It was basic DataOps gaps:</p><ul><li>the line state tag had been re-mapped during a controls update</li><li>the downtime reason codes in MES had changed, and the mapping table wasn’t updated</li><li>timestamps between systems weren’t aligned the same way</li></ul><p>We fixed it, but the real lesson was this.</p><p>The work wasn’t building a dashboard. The work was building the layer that makes dashboards boring to maintain.</p><p>That layer is Industrial DataOps.</p><h3>What to do if you want this. Without overcomplicating the scope</h3><p>If you’re thinking, “This sounds big,” you’re right. So start small.</p><p>Pick one production area and define one “data product,” for example:</p><ul><li>Line status and downtime events</li><li>Critical process parameters for one unit operation</li><li>Batch context plus key quality attributes</li></ul><p>Then build the DataOps basics around it:</p><ul><li>standard names and structure</li><li>automated data quality checks</li><li>versioned definitions</li><li>clear ownership</li><li>controlled deployment path</li></ul><p>Once that works, scale to the next one.</p><h3>Final thought</h3><p>Most teams don’t have a data lake problem. They have a <em>trust</em> problem.</p><p>Industrial DataOps is how you earn trust in industrial data. The slow way. The real way.</p><p>And yes, it takes discipline. But the payoff is huge.</p><p>Because when your data foundation is solid, everything above it gets simpler. Dashboards, investigations, analytics, AI. Even day-to-day operations.</p><p>And honestly. Simple is the whole point.</p><p><em>Originally published at </em><a href="https://iiotblog.com/2026/05/19/industrial-dataops-turning-plant-data-into-trusted-data/"><em>https://iiotblog.com</em></a><em> on May 19, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e552c3a0bcab" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Agentic AI on the Shop Floor: Real vs Hype in 2026]]></title>
            <link>https://medium.com/@henry.costa/agentic-ai-on-the-shop-floor-real-vs-hype-in-2026-1036a4a96e4a?source=rss-e2cc0bd3556a------2</link>
            <guid isPermaLink="false">https://medium.com/p/1036a4a96e4a</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[iiot]]></category>
            <category><![CDATA[industrial-ai]]></category>
            <category><![CDATA[innovation-manufacturing]]></category>
            <dc:creator><![CDATA[Henry Costa]]></dc:creator>
            <pubDate>Mon, 18 May 2026 17:16:13 GMT</pubDate>
            <atom:updated>2026-05-18T17:22:26.432Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DHiHcm8ssrXL2bPO" /></figure><p>There’s a lot of noise around Agentic AI right now. If you’re in manufacturing, you’ve probably seen the demos. An AI agent that “autonomously optimizes your production line.” Another one that “self-heals equipment failures.” Sounds amazing, right?</p><p>So let me share what I’m actually seeing on the shop floor in 2026, after 20+ years connecting machines, data, and people across global manufacturing networks.</p><h3>First, let’s get the basics straight</h3><p>Agentic AI means AI systems that can perceive, decide, and act on their own, without a human clicking “approve” every time. Think of it as the difference between a dashboard that shows you a temperature spike and a system that detects the spike, checks the context, adjusts a setpoint, and logs the action. All by itself.</p><p>That’s the promise. And honestly, some of it is starting to work. But most of it? Not yet.</p><h3>What’s actually real today</h3><p>Here’s what I’ve seen working in real manufacturing environments, not in demos or slide decks.</p><h4>1. Anomaly detection with automated alerts</h4><p>We’ve had real-time monitoring for years. Streaming data from PLCs, SCADA systems, and historians into cloud platforms. Running trend analysis, batch overlays, tracking KPIs. That part is mature. What’s new is adding AI models that learn normal patterns and flag deviations before an operator notices them. I’ve seen this running at scale, handling hundreds of thousands of tags across multiple sites. It works. But here’s the thing, it still sends an alert to a human. The “agentic” part, where it takes action on its own, is mostly limited to low-risk scenarios like adjusting a refresh rate or triggering a maintenance ticket.</p><h4>2. Predictive maintenance, sort of</h4><p>Everyone talks about predictive maintenance like it’s solved. It’s not. What I’ve seen work is pattern recognition on time-series data, things like spotting unusual vibration patterns or pressure readings that historically led to equipment failures. But going from “this looks off” to “automatically schedule a technician and order the part”, that’s still mostly manual. The AI can suggest, but the decision chain involves too many variables, spare parts availability, production schedules, compliance requirements. Especially in regulated industries, you can’t just let an agent decide to shut down a line.</p><h4>3. Data contextualization getting smarter</h4><p>This one is underrated. We’ve spent years building streaming engines that move data from equipment to data lakes, enriching it with metadata like site, building, unit, and batch information. Now AI is helping with the boring but critical part, auto-tagging, standardizing naming conventions across sites, and even suggesting data models based on existing structures. It’s not glamorous, but it saves hundreds of hours. And it’s a real use case where AI acts semi-autonomously without breaking anything.</p><h3>What’s still hype</h3><h4>1. “Fully autonomous production optimization”</h4><p>I keep seeing vendors demo this. An AI agent that watches your entire production line and continuously optimizes it. In reality, manufacturing processes are incredibly complex. A pharmaceutical batch process has hundreds of parameters that interact in non-linear ways. You can’t just let an agent tweak things. GxP compliance alone requires documented justification for every process change. FDA 21 CFR Part 11 doesn’t care how smart your AI is. You still need audit trails, role-based access, and human accountability.</p><h4>2. “Self-healing factories”</h4><p>Great concept. Terrible execution so far. The idea is that when something breaks, the AI diagnoses the root cause, reconfigures the process, and keeps production running. What actually happens? The AI might correctly identify the root cause 60–70% of the time. But reconfiguring a production process autonomously? In a GxP environment? With cybersecurity requirements for OT-cloud integration? No. We’re not there.</p><h4>3. “Drop-in AI agents for any equipment”</h4><p>Every IIoT platform vendor now has an “AI agent” feature. But here’s what they don’t tell you. These agents need clean, contextualized, properly streamed data to function. And most manufacturing sites are still working through basic connectivity challenges. Equipment diversity, legacy protocols, historian migrations, bandwidth limitations, local security policies. I’ve seen sites where just getting a stable data flow from a PLC to the cloud took months of work. You can’t layer agentic AI on top of a shaky data foundation.</p><h3>The honest truth about where we are</h3><p>We’re in the “smart assistant” phase, not the “autonomous agent” phase. And that’s okay.</p><p>The real progress is happening in the plumbing. Better streaming engines. More reliable edge-to-cloud architectures. Standardized data models. Unified namespaces. These aren’t sexy topics, but they’re the foundation that will eventually make real agentic AI possible.</p><p>I’ve been part of evaluating next-generation IIoT platforms, running proof of concepts across multiple sites, testing edge connectivity, data streaming, cloud integration, and performance under stress. The platforms are getting better. But even the best ones still need significant customization for each site, each process, each compliance framework.</p><h3>What I’d actually recommend</h3><p>If you’re a manufacturing leader thinking about Agentic AI, here’s my honest advice:</p><ol><li><strong>Fix your data first.</strong> If you don’t have clean, contextualized, reliably streamed data from your equipment, no AI agent will save you. Invest in your IIoT architecture. Get your streaming engine right. Standardize your naming conventions.</li><li><strong>Start with “assisted” not “autonomous.”</strong> Let AI suggest, flag, and recommend. Keep humans in the loop for decisions. Especially in regulated industries, this isn’t just safer, it’s required.</li><li><strong>Pick one use case and go deep.</strong> Don’t try to deploy AI across your entire operation. Pick something concrete, like OEE monitoring with AI-powered root cause analysis, and prove it works at one site before scaling.</li><li><strong>Don’t skip the PoC phase.</strong> I’ve seen organizations rush to deploy and regret it. Run a proper proof of concept. Test with real equipment, real data sources, real network conditions. Not simulated demos.</li><li><strong>Think about cybersecurity early.</strong> The moment you connect OT systems to cloud-based AI, you open attack surfaces. This isn’t an afterthought. It’s a design requirement from day one.</li></ol><h3>Looking ahead</h3><p>Will we see truly agentic AI on the shop floor? Yes, eventually. But probably not the way vendors are selling it today. It will come gradually. First in low-risk, well-understood processes. Then expanding as trust builds and regulations adapt.</p><p>The factories of the future won’t be “lights out” overnight. They’ll be “lights dimmed,” with AI handling the routine and humans handling the exceptions. And honestly? That’s a pretty good outcome.</p><p>The best technology I’ve seen in 20 years of doing this isn’t the flashiest. It’s the one that actually works when the network drops, when the batch goes sideways, when the auditor shows up. Agentic AI will get there. Just not as fast as the marketing slides suggest.</p><p><em>Originally published at </em><a href="https://iiotblog.com/2026/05/18/agentic-ai-on-the-shop-floor-real-vs-hype-in-2026/"><em>https://iiotblog.com</em></a><em> on May 18, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1036a4a96e4a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Advising to Owning. The Shift Nobody Fully Prepares You For]]></title>
            <link>https://medium.com/@henry.costa/from-advising-to-owning-the-shift-nobody-fully-prepares-you-for-4aff22e7e510?source=rss-e2cc0bd3556a------2</link>
            <guid isPermaLink="false">https://medium.com/p/4aff22e7e510</guid>
            <category><![CDATA[manufacturing]]></category>
            <category><![CDATA[consulting]]></category>
            <category><![CDATA[industry]]></category>
            <category><![CDATA[career-advice]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <dc:creator><![CDATA[Henry Costa]]></dc:creator>
            <pubDate>Tue, 12 May 2026 17:19:04 GMT</pubDate>
            <atom:updated>2026-05-12T17:26:28.037Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-g40AEPsnj3tZQEo" /></figure><p>I spent about 16 years in consulting. Building practices, leading delivery teams, flying to client sites, presenting architectures, and handing over solutions. Then I moved to industry. Specifically, to a large global pharmaceutical company, leading IIoT architecture from the inside.</p><p>And honestly, it’s a completely different game.</p><p>I don’t mean the technology changes. OPC UA is still OPC UA. MQTT is still MQTT. What changes is everything around the technology. The pace, the accountability, the politics, the emotions.</p><p>So here’s what I’ve learned after a few years sitting on the other side.</p><h3>1. You Stop Leaving</h3><p>In consulting, you show up, do the work, deliver, and eventually move on to the next client. There’s always a fresh start around the corner.</p><p>In industry, there’s no “next client.” You live with your decisions. That architecture you designed? You’re still maintaining it two years later. That workaround you accepted under pressure? It’s now your daily headache.</p><p>This changes how you think. You become much more careful, sometimes slower, because you know you’ll be the one cleaning up the mess.</p><h3>2. You Lose the “Expert Halo”</h3><p>In consulting, when you walk into a room, people listen. You’re the external expert. You flew in. You have a fancy slide deck. There’s a built-in authority that comes with being an outsider.</p><p>In industry, you’re just another colleague. Nobody cares that you’ve done this 20 times before. They care about whether you understand their plant, their process, their constraints. And that takes time to earn.</p><p>For me, this was humbling. I went from being the person everyone looked to for answers, to the person who had to ask a lot of questions before anyone trusted my input.</p><h3>3. The Pace Is Different (Not Slower, Just Different)</h3><p>People in consulting often assume industry moves slowly. That’s not quite right.</p><p>Industry moves carefully. There are validation requirements, cybersecurity reviews, change control processes, GxP compliance gates. In pharma, you don’t just deploy something because it works. You deploy it because it’s proven, documented, and approved.</p><p>So the pace is not slow. It’s deliberate. And once you understand why, you actually start to appreciate it. I’ve seen too many consulting projects rush to go-live and leave the client with a fragile system.</p><h3>4. Politics Hit Different</h3><p>In consulting, organizational politics exist, but they are often easier to navigate. Since consultants usually work on temporary engagements, they are less tied to long-term internal structures and dynamics.</p><p>In industry roles, politics are much more embedded in day-to-day operations. Budget planning, headcount discussions, cross-functional priorities, and long-standing vendor relationships all play a role in how decisions are made. These factors cannot really be avoided, so learning how to work within them becomes important.</p><p>For me, this was one of the biggest adjustments. In consulting, most of my attention could stay on the technical solution itself. In industry, technical expertise is still important, but it is only part of the work. Alignment across teams, influencing stakeholders, and maintaining patience are just as critical for making progress.</p><h3>5. You Actually See the Impact</h3><p>This is the best part. And I mean it.</p><p>In consulting, you deliver something and move on. You rarely see how it performs over time. You don’t see the operator using your dashboard every morning. You don’t see the maintenance team catching a failure before it happens because of the alerting system you designed.</p><p>In industry, you see all of it. You walk the plant floor. You talk to the people using the system. You watch it evolve. And when something works, really works, it feels different. It feels real.</p><p>That moment when a site engineer says, “This actually helped us,” is worth more than any consulting milestone payment.</p><h3>6. You Realize How Much Consulting Misses</h3><p>This one might be unpopular, but I’ll say it anyway.</p><p>When I was in consulting, I thought we understood our clients deeply. We did workshops, interviews, assessments. We produced beautiful documents.</p><p>But from the inside, I now see how much we missed. The informal processes, the tribal knowledge, the real reasons certain decisions were made. Things that don’t show up in a workshop. Things people don’t tell consultants because they don’t trust them enough yet, or because they don’t even realize it’s important.</p><p>This doesn’t mean consulting is bad. It means there’s a gap. And the best consultants I’ve worked with, both as a peer and now as a client, are the ones who stay long enough and listen hard enough to close that gap.</p><h3>7. Your Definition of Success Changes</h3><p>In consulting, success is often tied to deliverables, timelines, and contract milestones. Did we finish on time? Did the client sign off?</p><p>In industry, success is adoption. Did people actually use what we built? Did it survive the first year? Did it scale to a second site? Did it reduce manual work or improve data quality?</p><p>I’ve seen plenty of “successful” consulting projects that delivered on time but were abandoned six months later. From the inside, you measure success very differently.</p><h3>So, Would I Go Back?</h3><p>Honestly, I miss some things about consulting. The variety, the pace, the constant learning from different industries and problems.</p><p>But I wouldn’t trade what I have now. Being embedded in a real operation, seeing the full lifecycle, understanding the true complexity of manufacturing at scale. That perspective is something consulting never fully gave me.</p><p>If you’re in consulting and thinking about making the switch, my advice is simple. Do it with your eyes open. Be ready to unlearn some habits. And be patient with yourself, because the first year is basically a reset.</p><p>And if you’re in industry wondering what consultants are really thinking when they show up at your site, well, now you know. We were doing our best. We just didn’t always have the full picture.</p><p><em>Originally published at </em><a href="https://iiotblog.com/2026/05/12/from-advising-to-owning-the-shift-nobody-fully-prepares-you-for/"><em>https://iiotblog.com</em></a><em> on May 12, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4aff22e7e510" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Software-Defined Automation (SDA) Is Coming. What It Means for Your IIoT Architecture]]></title>
            <link>https://medium.com/@henry.costa/software-defined-automation-sda-is-coming-what-it-means-for-your-iiot-architecture-851bc1502c6f?source=rss-e2cc0bd3556a------2</link>
            <guid isPermaLink="false">https://medium.com/p/851bc1502c6f</guid>
            <category><![CDATA[asd]]></category>
            <category><![CDATA[un]]></category>
            <category><![CDATA[smart-manufacturing]]></category>
            <category><![CDATA[iiot]]></category>
            <category><![CDATA[software-defined]]></category>
            <dc:creator><![CDATA[Henry Costa]]></dc:creator>
            <pubDate>Thu, 07 May 2026 14:23:36 GMT</pubDate>
            <atom:updated>2026-05-07T15:45:11.427Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*B0x4ANWYEFIH2EvX" /></figure><p>Software-Defined Automation (SDA) is one of those terms that sounds like it was invented in a marketing meeting. But this time, there’s real substance behind it. And if you’re building or evolving an IIoT architecture today, you should pay attention.</p><p>Let me break it down based on what I’ve actually seen so far, not what the slide decks promise.</p><h3>What Is Software-Defined Automation?</h3><p>For decades, industrial automation has been hardware-defined. You buy a PLC from a vendor, you get their software, their engineering environment, their protocols, their upgrade cycle. Everything is tightly coupled. If you want new functionality, you usually buy new hardware.</p><p>SDA changes that model.</p><p>The core idea is simple. Decouple the software from the hardware. The control logic, HMIs, gateways, and industrial applications become portable. They can run on edge devices, virtual machines, containers, or different compute platforms, wherever it makes sense operationally.</p><p>In many ways, it’s similar to what virtualization did to IT infrastructure years ago.</p><p>But here’s the reality. The plant floor doesn’t change because someone publishes a whitepaper.</p><h3>Why This Shift Matters</h3><p>Recently, I was involved in the architecture of a large-scale IIoT platform for a global manufacturing company. We initially selected a cloud-native IoT broker from a major hyperscaler. On paper, it seemed like the ideal solution: a managed service with native integrations, simplified deployment, and reduced operational overhead.</p><p>Then real-world scale entered the picture.</p><p>As data volumes and site adoption increased, the platform started running into throughput limitations, API constraints, and rapidly growing operational costs. We eventually had to redesign the messaging layer and move to an enterprise-grade MQTT architecture that provided more control and scalability.</p><p>That single change had a ripple effect across the entire architecture. We had to redesign ingestion pipelines, introduce custom integration layers, revisit data contracts, and rethink how information moved from the edge to enterprise systems.</p><p>That experience reinforced something important for me.</p><p>When an industrial architecture becomes too tightly coupled to a specific vendor ecosystem, even necessary changes become expensive and complex. Upgrades turn into migrations. Scaling becomes harder than expected. Expanding to new sites often requires custom work instead of repeatable deployment patterns.</p><p>That’s one of the reasons Software-Defined Automation is gaining traction. The move toward <strong>decoupled</strong>, <strong>portable</strong>, software-centric architectures simply makes sense for where industrial systems are heading.</p><h3>If You’ve Built a UNS, You’re Already Closer Than You Think</h3><p>This is the part many people miss.</p><p>If your organization has already been investing in Unified Namespace concepts, MQTT-based architectures, standardized data models, event-driven communication, or containerized edge services, then you are already moving in the same direction as many Software-Defined Automation principles.</p><p>Over the past several years, many industrial teams have been modernizing their architectures by standardizing information models, improving deployment practices, reducing dependency on proprietary integrations, and adopting more scalable software-centric approaches.</p><p>That shift toward portability, decoupling, abstraction from hardware, and infrastructure automation is fundamentally what SDA is trying to formalize.</p><p>In many ways, the industry has already been moving there. We just weren’t using the same terminology yet.</p><p>So if your organization is already building modern industrial data foundations and scalable integration architectures, you are not starting from zero. You are likely further along than you think.</p><h3>The Hardest Part Will Not Be Technology</h3><p>One of the biggest challenges with Software-Defined Automation will not be the technology itself. It will be organizational mindset and operational culture.</p><p>In many industrial environments, there is still a strong separation between traditional automation practices and modern software-centric approaches. Concepts like containerized deployments, infrastructure automation, and software-defined systems are often viewed as “IT concepts” rather than part of industrial operations.</p><p>There is also understandable hesitation around moving away from long-established hardware ecosystems and operational models that have been trusted for decades.</p><p>And honestly, those concerns are valid.</p><p>Industrial environments prioritize stability, reliability, safety, and predictable operations above everything else. When production downtime has major operational or financial impact, organizations naturally move carefully when introducing new architectural paradigms.</p><p>That’s why this transition will happen gradually, not overnight.</p><p>Software-Defined Automation will likely continue expanding first through monitoring systems, edge gateways, data integration layers, analytics platforms, and supervisory applications. As operational maturity, trust, and standardization improve, software-defined concepts will progressively move closer to core operational and control environments.</p><h3>What I’d Prioritize If I Were Building an IIoT Architecture Today</h3><p>If I were designing a greenfield IIoT architecture right now, SDA would absolutely influence the decisions I make.</p><p>Here’s what I would prioritize:</p><ol><li><strong>Build your UNS on open MQTT infrastructure. </strong>Not on proprietary messaging layers or cloud-specific brokers. Use a platform you control.</li><li><strong>Standardize your information model early. </strong>UDTs, topic structures, naming conventions, metadata standards. Without a strong data model, SDA just moves the mess somewhere else.</li><li><strong>Containerize as much as possible at the edge. </strong>Protocol converters, gateways, ingestion services, transformation layers. Containers make upgrades operational instead of project-based.</li><li><strong>Decouple ingestion from the cloud provider. </strong>Build adapters and bridges that you own. Vendor-native ingestion paths create long-term lock-in faster than most teams realize.</li><li><strong>Invest in CI/CD practices for OT systems. </strong>Version control, approval workflows, automated deployments, rollback strategies, and auditability are no longer optional in modern industrial environments.</li></ol><h3>My Take on Where the Market Is Going</h3><p>Here’s my view.</p><p>SDA is real, and the direction is absolutely important. But many vendors today are simply rebranding existing products with a new label.</p><p>True Software-Defined Automation requires open standards, interoperability, and the ability for customers to mix technologies without being trapped inside a single ecosystem.</p><p>Most vendors are still uncomfortable with that reality.</p><p>So yes, be excited about SDA. But also stay skeptical of the marketing.</p><p>Build your architecture around open standards. Invest heavily in your data model. Control your integration layer. Design for portability from the beginning.</p><p>The strongest IIoT architectures I’ve worked on were never the most proprietary ones. They were the ones where we maintained flexibility, ownership, and architectural independence.</p><p>SDA doesn’t replace those principles. It validates them.</p><p><em>Originally published at </em><a href="https://iiotblog.com/2026/05/07/software-defined-automation-sda-is-coming-what-it-means-for-your-iiot-architecture/"><em>https://iiotblog.com</em></a><em> on May 7, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=851bc1502c6f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Communicating Edge Contextualization and Unified Namespace to Leadership]]></title>
            <link>https://medium.com/@henry.costa/communicating-edge-contextualization-and-unified-namespace-to-leadership-dd262d097627?source=rss-e2cc0bd3556a------2</link>
            <guid isPermaLink="false">https://medium.com/p/dd262d097627</guid>
            <category><![CDATA[iiot]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <category><![CDATA[unified-namespace]]></category>
            <category><![CDATA[data-centric-architecture]]></category>
            <category><![CDATA[edge-contextualization]]></category>
            <dc:creator><![CDATA[Henry Costa]]></dc:creator>
            <pubDate>Wed, 29 Apr 2026 16:18:45 GMT</pubDate>
            <atom:updated>2026-04-29T16:30:43.855Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*b-Q9iBfqJFGnXhPO" /></figure><p>The team had a problem. Not technical this time. They needed to explain <strong>Edge Contextualization</strong> and <strong>Unified Namespace (UNS)</strong> to senior leaders who did not care about protocols or data models.</p><p>The audience knew their factories well. Directors, VPs, plant managers. But when the discussion shifted to MQTT topics or edge processing, attention dropped almost instantly. It happens in every company. Technical depth loses the room fast.</p><p>If they could not make the message clear, the project would stall. No budget. No alignment. No rollout.</p><p>So they made a decision. Stop explaining the technology. Start explaining the value.</p><h3>Why It Mattered</h3><p>A large food and beverage company was rolling out an IIoT platform across their global manufacturing network. Dozens of plants. Thousands of equipment tags. Multiple historians, different automation vendors, different naming conventions at every facility.</p><p>The architecture called for two things:</p><ol><li><strong>Contextualize data at the edge</strong>, meaning add meaning to raw machine signals before sending them anywhere.</li><li><strong>Publish everything into a Unified Namespace (UNS)</strong>, a single, organized structure where all operational data lives and can be consumed by anyone who needs it.</li></ol><p>Walker Reynolds, who coined the term, describes UNS as “a real-time single source of truth for data in an industrial or manufacturing environment, semantically organized like the business and built to be open.” That’s a great definition, but try saying that to a VP of Operations who just wants to know why their OEE numbers don’t match across plants.</p><p>Both concepts are critical. But if leadership didn’t understand why, the project would never get the budget, the site access, or the cooperation it needed.</p><h3>The Grocery Store Analogy</h3><p>Here’s what finally worked.</p><p>The lead architect asked the group: “Imagine you walk into a grocery store and all the products are just thrown on the floor. No aisles, no labels, no categories. The milk is next to the laundry detergent. The bread is somewhere in the back, maybe. How long would it take you to find what you need?”</p><p>Everyone laughed. “That’s ridiculous,” someone said.</p><p>“Exactly. That’s what our data looks like right now.”</p><p>Every plant has its own naming conventions. One facility calls a mixing tank “MX-101.” Another calls the exact same type of equipment “MIXER_A.” A third uses a code that only the local automation engineer understands. When that data lands in the cloud, nobody knows what’s what. It’s a grocery store with no aisles.</p><p>So what does a Unified Namespace do? It’s the store layout. It says: “All dairy goes in Aisle 3. All cleaning products go in Aisle 7.” It gives every piece of data a logical address that makes sense across the entire company, not just one plant.</p><p>And Edge Contextualization? That’s the work that happens before the product even reaches the shelf. Someone at the warehouse puts the label on, sorts it, and makes sure it goes to the right aisle. If you skip that step, you’re just dumping boxes on the floor again.</p><p>That analogy stuck. People started using it in their own meetings. “Are we putting this in the right aisle?” became a real question in architecture reviews.</p><h3>The Real-Life Example That Sealed It</h3><p>Analogies are nice, but leaders want proof. So the team showed them a real scenario from one of the pilot sites, a bottling plant that ran multiple packaging lines.</p><p>About 8,000 tags were streaming from two production lines. Temperature, pressure, fill levels, line speed, batch IDs, equipment states. At the edge, a gateway collected all of that and a data modeling tool added context, things like which production line, which building, which plant, and what unit of measure.</p><p>Before contextualization, the data looked like this:</p><ul><li>Tag_4021: 4.8</li><li>Tag_4022: 72.1</li><li>Tag_4023: TRUE</li></ul><p>Useless to anyone who wasn’t the local automation engineer.</p><p>After contextualization, the same data looked like this:</p><ul><li>PlantUS02/Bldg_North/Line_03/Filler/Pressure: 4.8 bar</li><li>PlantUS02/Bldg_North/Line_03/Pasteurizer/Temperature: 72.1 °C</li><li>PlantUS02/Bldg_North/Line_03/Filler/Running: TRUE</li></ul><p>Now anyone in the organization, from a data scientist to the VP of Quality, could look at that and immediately understand what it meant, where it came from, and what it referred to. No phone calls. No tribal knowledge. No guessing.</p><p>One operations director looked at the before-and-after and said: “So we’ve been sending garbage to the cloud this whole time?” Not garbage exactly, but yes, data without context is just noise.</p><p>This is what consulting companies call the shift from “application-centric” to “data-centric” architecture. Instead of each system owning its own silo, you organize everything around the data itself. The UNS becomes the backbone, and every system, whether it’s your MES, your ERP, your historian, or your analytics platform, just subscribes to what it needs.</p><h3>The Three Sentences That Worked</h3><p>After a few rounds of these conversations, the team landed on three sentences that work every time someone asks for an explanation of UNS and Edge Contextualization:</p><ol><li><strong>“A Unified Namespace is like a common language for all your plant data, so every site speaks the same way.”</strong></li><li><strong>“Edge Contextualization is the translator that adds meaning to raw machine data before it leaves the plant.”</strong></li><li><strong>“Together, they make sure that when data arrives in the cloud, it’s ready to use, not a puzzle you need to solve.”</strong></li></ol><p>That’s it. No mention of MQTT, no mention of brokers, no mention of Sparkplug or topic trees. Just outcomes.</p><p>I big IT company puts it well in their recent research: UNS provides “a single source of truth for all industrial data, eliminating silos and enabling real-time, standardized data flows from the shop floor to the executive suite.” That’s exactly what we were trying to say, just in fancier words.</p><h3>What We Learned Along the Way</h3><p>A few things became clear during this process:</p><ul><li><strong>Start with the “why,” not the “how.”</strong> Leaders don’t care about protocols. They care about speed, cost, and risk. Frame everything around those three.</li><li><strong>Use the same vocabulary the business uses.</strong> If they say “recipe,” don’t say “batch context payload.” If they say “line,” don’t say “edge node.”</li><li><strong>Show, don’t tell.</strong> A 2-minute demo of before-and-after data is worth more than a 30-slide deck. Many teams learn this the hard way after a few workshops that go nowhere.</li><li><strong>Don’t hide the complexity. Just don’t lead with it.</strong> When someone asks “how does it work under the hood,” give them the honest answer. But don’t open with the hood up.</li><li><strong>Connect it to something they already measure.</strong> In this case, OEE was the magic word. When the team explained that contextualized data could help compare OEE across plants with confidence (because the data was finally standardized), the conversation shifted from “why should we care” to “when can we start.”</li></ul><h3>The Honest Opinion</h3><p>Here’s something that’s true but not everyone agrees with: most IIoT projects fail not because of technology, but because the technical team can’t explain what they’re doing in plain language.</p><p>Brilliant architectures get killed in steering committees because the presenter uses 40 acronyms in 10 minutes. And simpler, less elegant solutions get funded because someone tells a clear story.</p><p>If you can’t explain your architecture to a plant director in under 3 minutes, you don’t understand it well enough yet. Or, more likely, you haven’t thought enough about what it means for them.</p><p>This team spent almost as much time preparing how to communicate their architecture as they did designing it. That’s not a waste of time. That’s the job.</p><p>There’s an old Reddit thread where an industrial controls engineer asks, genuinely frustrated: “What is the actual advantage of IoT in an industrial context over existing technology?” That question gets asked in every steering committee, just in different words. If your technical team can’t answer it in plain English, you’ve already lost the room.</p><h3>A Quick Note on Naming Conventions</h3><p>One of the biggest practical challenges was naming. Every plant had its own way of naming equipment, tags, areas, and lines. When the team started defining the UNS hierarchy, it became a political conversation, not a technical one.</p><p>“Why should we change our naming? We’ve been using this for 15 years.”</p><p>Fair point. So nobody was forced to change their local systems. Instead, the group agreed on a common namespace that sits on top. The edge gateway maps local names to the global structure. The plant keeps its naming. The enterprise gets consistency. Everyone wins.</p><p>That compromise took weeks to negotiate, but it was worth it. Without it, the project would have been stuck in an endless loop of “my naming convention is better than yours.”</p><p>This is actually a well-known pattern in UNS implementations. A vendor MQTT broker’s UNS essentials series describes it as building a “live, shared data environment where machines, systems, and even AI can communicate and collaborate in real-time.” But getting to that shared environment means getting humans to agree first. That’s always the harder part.</p><h3>Wrapping Up</h3><p>For anyone working on IIoT or Industrial AI, UNS, or data architecture in manufacturing, don’t underestimate the communication side of the work. The technology is important, obviously. But getting people on board is what makes it real.</p><p>And if you ever find yourself in a room full of non-technical leaders, just remember the grocery store. It works every time.</p><p><em>Originally published at </em><a href="https://iiotblog.com/2026/04/29/communicating-edge-contextualization-and-unified-namespace-to-leadership/"><em>https://iiotblog.com</em></a><em> on April 29, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dd262d097627" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Real-Time Decisions: Lessons From Mining Operations]]></title>
            <link>https://medium.com/@henry.costa/real-time-decisions-lessons-from-mining-operations-432c0403d376?source=rss-e2cc0bd3556a------2</link>
            <guid isPermaLink="false">https://medium.com/p/432c0403d376</guid>
            <category><![CDATA[mining]]></category>
            <category><![CDATA[iiot]]></category>
            <category><![CDATA[edge-computing]]></category>
            <category><![CDATA[real-time-data]]></category>
            <category><![CDATA[alarm-management]]></category>
            <dc:creator><![CDATA[Henry Costa]]></dc:creator>
            <pubDate>Mon, 20 Apr 2026 17:32:28 GMT</pubDate>
            <atom:updated>2026-04-20T18:44:42.778Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*vuIqC4MmzaXADUd7" /></figure><p>I never planned to work in mining. It just happened. Early in my career, a project came up at a metals and mining operation in Brazil, and I said yes. That decision gave me some of the most intense, real lessons about what real-time data actually means when the stakes are high.</p><p>Most people in IIoT talk about real-time data in clean, air-conditioned settings. Dashboards, charts, nice conference rooms. Mining is nothing like that. It’s loud, dusty, massive, and the equipment is brutal. But the principles I learned there still guide how I think about real-time decision making today.</p><h3>The Environment Changes Everything</h3><p>The first thing that hit me was the scale. We’re talking about conveyor belts stretching hundreds of meters, crushers the size of small buildings, and furnaces running 24/7. When something goes wrong in a mining plant, it doesn’t wait for your approval to escalate. A blocked crusher or an overheated bearing doesn’t send a polite email. It just breaks.</p><p>That’s where I first really understood that real-time data isn’t about having more data. It’s about having the right data at the right moment, in a form that someone can actually act on. In a mining control room, operators don’t need 500 tags on a screen. They need three or four signals that tell them: “Something is about to go wrong. Here. Now.”</p><h3>When Seconds Actually Matter</h3><p>In pharma or automotive, real-time often means minutes. Maybe even hours, depending on the process. In mining, especially in smelting or crushing operations, real-time can literally mean seconds. A temperature spike in a furnace can damage refractory lining in minutes. A conveyor misalignment can shred a belt worth tens of thousands of dollars before anyone walks over to check.</p><p>So the systems we built had to be fast, but more importantly, they had to be simple. The operators weren’t engineers with laptops. They were experienced, practical people who needed clear alerts, not complex dashboards. That taught me something I carry to every project now. If your real-time system requires a training course to understand, it’s not real-time. It’s just fast data that nobody uses.</p><h3>The Historian Wasn’t Enough</h3><p>At that site, we had a traditional historian collecting process data. Thousands of tags, sampled every few seconds. On paper, it looked great. In practice, it was mostly used for post-mortem analysis. Something broke, and then someone would pull up the trend to figure out what happened.</p><p>That’s not decision making. That’s archaeology.</p><p>The real shift came when we started building simple rules on top of the streaming data. Not AI, not machine learning, just practical threshold logic combined with rate-of-change calculations. For example, if a motor’s vibration increases by more than 15% in under two minutes, flag it. Don’t wait for it to cross an absolute limit. Watch the trend, not just the number.</p><p>This sounds obvious now, but back then, most systems were designed around static alarm limits. High-high, high, low, low-low. That’s it. Nobody was watching how fast things were changing, which in mining, is often more important than the value itself.</p><h3>People Make the Decisions, Not Systems</h3><p>One of the biggest lessons from mining was about people. The best operators I worked with had decades of experience. They could hear a motor struggling before any sensor flagged it. They knew the smell of overheated rubber on a conveyor belt. No system replaces that.</p><p>What we learned to do was build tools that supported their instincts, not replaced them. Give them a simple, clean view of what’s happening. Let them confirm what they already suspect. And most importantly, don’t bury them in false alarms.</p><p>False alarms were a huge problem. At one point, the control room was getting hundreds of alarms per shift. The operators just started ignoring them. It’s a well-known problem in process industries, called alarm fatigue. In mining, the consequences of ignoring the wrong alarm can be catastrophic.</p><p>So we spent weeks rationalizing alarms. Removing duplicates, adjusting thresholds, grouping related alerts. It wasn’t glamorous work, but it made a real difference. The operators started trusting the system again. And when an alarm went off, they actually looked at it.</p><h3>Dust, Heat, and the Reality of Edge Computing</h3><p>Another thing mining taught me is that your architecture has to survive the environment. We had servers in rooms where the dust would clog filters in days. Temperatures that pushed hardware to its limits. Network connections that would drop because a truck drove over a cable tray.</p><p>This is where I first appreciated the importance of edge computing before it was even called that. We needed local processing because we couldn’t rely on a stable connection to a central server. Data had to be buffered, processed locally, and forwarded when the connection came back. If the edge couldn’t handle it alone for a few hours, the design was wrong.</p><p>That lesson applies everywhere today. Whether you’re in a pharmaceutical plant with strict network segmentation or a remote food processing facility with limited bandwidth, the edge has to be self-sufficient. Mining taught me that the hard way.</p><h3>Simple Beats Sophisticated</h3><p>Here’s my honest opinion, and I know it’s not popular in the IIoT world. The most effective real-time systems I’ve seen in mining were the simplest ones. Not the ones with the fanciest dashboards or the most advanced analytics. The ones where an operator could glance at a screen and know exactly what to do.</p><p>I’ve seen projects where teams spent months building beautiful visualization layers, AI-driven predictive models, and complex event processing engines. And then the operators went back to their old paper checklists because the new system was too complicated to use during a shift.</p><p>The mining environment forced us to keep things simple because there was no room for complexity. The operators needed answers, not options. And that discipline, building for the person who has to act, not the person who has to present, is something I wish more IIoT projects would adopt.</p><h3>What I Took With Me</h3><p>I eventually moved on from mining to other industries. Automotive, pharma, food and beverage. But the lessons from those early projects stuck with me:</p><ul><li><strong>Real-time means actionable, not just fast.</strong> If nobody acts on the data, it’s just noise.</li><li><strong>Watch the rate of change, not just the value.</strong> How fast something is changing often matters more than where it is right now.</li><li><strong>Alarm rationalization is not optional.</strong> Too many alarms equals no alarms.</li><li><strong>Design for the operator, not the engineer.</strong> The person making the decision at 3 AM needs clarity, not complexity.</li><li><strong>The edge must survive on its own.</strong> Never assume the network will be there when you need it.</li><li><strong>Simple systems get used. Complex systems get bypassed.</strong></li></ul><p>Mining is a tough industry. The conditions are extreme, the margins are tight, and the tolerance for downtime is almost zero. But that’s exactly what makes it such a great teacher. It strips away all the nice-to-haves and forces you to focus on what actually works.</p><p>Every time I start a new IIoT project, no matter the industry, I think back to that dusty control room. That’s where I learned what real-time really means.</p><p><em>Originally published at </em><a href="https://iiotblog.com/2026/04/20/real-time-decisions-lessons-from-mining-operations/"><em>https://iiotblog.com</em></a><em> on April 20, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=432c0403d376" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Meets UNS: Practical LLM Uses on Industrial Data]]></title>
            <link>https://medium.com/@henry.costa/ai-meets-uns-practical-llm-uses-on-industrial-data-bca6a98c58bf?source=rss-e2cc0bd3556a------2</link>
            <guid isPermaLink="false">https://medium.com/p/bca6a98c58bf</guid>
            <category><![CDATA[iiot]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[un]]></category>
            <category><![CDATA[unified-namespace]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Henry Costa]]></dc:creator>
            <pubDate>Fri, 17 Apr 2026 21:20:37 GMT</pubDate>
            <atom:updated>2026-04-17T21:33:05.843Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*UZmEj82nHBD_8Y4T" /></figure><p>A few years ago, if you told me we’d be using large language models on top of plant data, I would’ve said: “Cool demo. Not sure it survives contact with a real site.”</p><p>Now I’m more convinced. Not because LLMs magically understand manufacturing. They don’t. I’m convinced because we found a few practical, boring-looking use cases where LLMs helped real people do real work faster, as long as we put them on top of a solid Unified Namespace (UNS) and treated them like an assistant, not an oracle.</p><p>Here are the practical ways we used LLMs on top of industrial data. Sanitized. Truth-only. No fairy tales.</p><h3>First, the unsaid truth. LLMs don’t fix messy data</h3><p>The best “AI project” I’ve seen in a plant wasn’t an AI project. It was cleaning up naming standards, tag models, and context.</p><p>If your data is scattered across:</p><ul><li>a historian with inconsistent tag names</li><li>a SCADA system where everything is “Tag_1234”</li><li>spreadsheets people swear are “the source of truth”</li><li>PLC tags that mean one thing at Site A and another thing at Site B</li></ul><p>Then an LLM will just produce confident nonsense faster.</p><p>A UNS helps because it forces you to answer hard questions:</p><ul><li>What is this asset?</li><li>What is this measurement?</li><li>What is the unit?</li><li>What is the state model?</li><li>What is the relationship between this line, this machine, and this batch?</li></ul><p>When we had that, the LLM became useful. When we didn’t, it was mostly a chatbot with good manners.</p><h3>Use case 1: Plain-English search across tags and assets</h3><p>This was the first win, and it was almost embarrassingly simple.</p><p>People don’t want to search for:</p><p>PlantA/Area3/Line2/Filler01/Status/Mode</p><p>They want to type:</p><p><em>“Show me the filler mode changes from yesterday”</em></p><p>or</p><p><em>“Which machine caused the most downtime on Line 2 this week?”</em></p><p>So what we did was combine:</p><ul><li>a UNS with consistent hierarchy and naming</li><li>a searchable catalog of tags, assets, and descriptions</li><li>an LLM that translated a user question into a structured query</li></ul><p>Not “LLM, please answer from vibes.” More like:</p><ol><li>Identify the asset(s)</li><li>Identify the signals (status, speed, downtime, alarms)</li><li>Identify the time window</li><li>Generate a query against the historian or time-series store</li><li>Return results with links, units, and timestamps</li></ol><p>The LLM wasn’t the brain. It was the translator.</p><p><strong>What changed operationally:</strong> controls engineers and reliability folks found data faster without remembering exact tag paths. That alone saved real time during investigations.</p><p><strong>Hard rule we kept:</strong> the model never “estimated” values. It only returned values actually queried from the source system.</p><h3>Use case 2: Faster downtime triage with narrative summaries</h3><p>Downtime triage is where humans lose time. Not because they’re slow. Because the data is noisy.</p><p>A typical “micro-stop” story looks like:</p><ul><li>machine state flickers</li><li>alarm list explodes</li><li>operator notes are… creative</li><li>timestamps don’t line up perfectly</li><li>some events are missing</li></ul><p>We used the LLM to generate a <em>draft narrative</em> like:</p><ul><li>what happened first</li><li>what changed next</li><li>what alarms appeared around that window</li><li>when the machine returned to steady state</li></ul><p>But we constrained it heavily:</p><ul><li>The LLM received a structured event timeline (from the UNS context plus alarms and states).</li><li>The output had to cite exact timestamps and signal names from that timeline.</li><li>If the timeline was incomplete, it had to say “unknown” instead of guessing.</li></ul><p><strong>Example of what this looked like in practice:</strong></p><p>“ <em>Between 10:14:22 and 10:16:05 the machine state moved from Running to Starved. A ‘Downstream Blocked’ alarm appeared at 10:14:30. Speed dropped from 220 to 0 over 12 seconds. The system returned to Running at 10:16:05.</em> “</p><p>That’s the kind of thing a human can write. But writing it 15 times a day gets old.</p><p><strong>What changed operationally:</strong> quicker first-pass triage. People still did root cause analysis. We just reduced the “blank page” problem.</p><h3>Use case 3: Operator and technician “what do I do now?” assistant</h3><p>This one surprised me. The value wasn’t in answering deep technical questions. It was in helping people navigate documentation and procedures.</p><p>In most plants, procedures exist. The problem is:</p><ul><li>they’re hard to find</li><li>they’re written in a way nobody talks</li><li>people don’t know which procedure applies to which asset variant</li></ul><p>Because the UNS already links:</p><ul><li>asset identity</li><li>model/type</li><li>firmware versions (sometimes)</li><li>line context</li><li>associated documents</li></ul><p>We could let a technician ask:</p><p><em>“What’s the restart procedure for this skid after a power dip?”</em></p><p>or</p><p><em>“Where is the calibration SOP for this transmitter?”</em></p><p>And the assistant would respond with:</p><ul><li>the most relevant procedure link</li><li>the specific section</li><li>a short summary in plain English</li></ul><p><strong>Important constraint:</strong> the assistant did not invent steps. It only summarized existing approved documents. If there was no approved doc, it said so.</p><p><strong>What changed operationally:</strong> less hunting. Less tribal knowledge dependency.</p><h3>Use case 4: Data quality and naming standard enforcement</h3><p>Nobody wakes up excited to enforce naming standards. But everyone suffers when they’re ignored.</p><p>We used the LLM as a reviewer.</p><p>When new tags or assets were added, the LLM checked:</p><ul><li>naming convention compliance</li><li>missing descriptions</li><li>unit mismatches (psi vs bar, C vs F)</li><li>duplicated signals that “look the same but aren’t”</li></ul><p>This worked well because the UNS gives structure. The LLM could say:</p><p><em>“This tag name doesn’t match the pattern for this asset class.”</em></p><p>or</p><p><em>“Description is empty. Add what this measures and where.”</em></p><p>It’s like spellcheck. For engineers.</p><p><strong>What changed operationally:</strong> cleaner UNS over time. Fewer “mystery tags.”</p><h3>Use case 5: Converting engineering speak into business speak</h3><p>Here’s an unpopular opinion. Most “AI in manufacturing” fails because teams try to impress leadership with tech, instead of translating impact.</p><p>We used the LLM to help convert plant metrics into a plain summary for non-technical stakeholders:</p><ul><li>“Top 3 constraints this week”</li><li>“Where we lost time”</li><li>“What improved”</li><li>“What we’re doing next”</li></ul><p>But again, we forced it to base statements on actual numbers coming from the data, not generic advice.</p><p><strong>What changed operationally:</strong> better alignment. Less time in meetings explaining basics.</p><h3>What didn’t work (at least not for me)</h3><p>A few things looked good in demos and didn’t hold up:</p><ol><li><strong>“Ask the LLM anything about the plant”</strong>: Too broad. Too many edge cases. Too much hallucination risk.</li><li><strong>Letting the model directly recommend parameter changes</strong>: In regulated or safety-conscious environments, this is a non-starter unless it’s tightly controlled, validated, and reviewed. And even then, good luck.</li><li><strong>Skipping governance</strong>: If you can’t answer “where did this data come from” and “who approved this procedure,” the whole thing becomes a liability fast.</li></ol><h3>The practical pattern that worked</h3><p>The pattern that consistently worked looked like this:</p><ol><li><strong>UNS provides structure and context</strong> (assets, relationships, consistent naming)</li><li><strong>LLM converts intent into a structured action</strong> (query, retrieve, summarize)</li><li><strong>System returns real results</strong> (time-series, events, docs)</li><li><strong>LLM summarizes with constraints</strong> (no guessing, cite timestamps, show units)</li><li><strong>Human remains accountable</strong> (review, decisions, changes)</li></ol><p>Not flashy. Very effective.</p><h3>Final thought</h3><p>LLMs on top of a UNS can be genuinely useful. But only if you treat the UNS as the foundation and the LLM as a helper.</p><p>If someone asks me, “Should we start with AI or with UNS?” I’ll give the boring answer every time.</p><p>Start with the UNS. Then add AI where it removes friction.</p><p>Because in manufacturing, “pretty good” isn’t good enough. Not when downtime costs money, and mistakes cost more than money.</p><p><em>Originally published at </em><a href="https://iiotblog.com/2026/04/17/ai-meets-uns-practical-llm-uses-on-industrial-data/"><em>https://iiotblog.com</em></a><em> on April 17, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bca6a98c58bf" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Industrial AI Explained. Why Data Is the Real Starting Point]]></title>
            <link>https://medium.com/@henry.costa/industrial-ai-explained-why-data-is-the-real-starting-point-bd697c908130?source=rss-e2cc0bd3556a------2</link>
            <guid isPermaLink="false">https://medium.com/p/bd697c908130</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[smart-manufacturing]]></category>
            <category><![CDATA[iiot]]></category>
            <category><![CDATA[un]]></category>
            <category><![CDATA[industrial-ai]]></category>
            <dc:creator><![CDATA[Henry Costa]]></dc:creator>
            <pubDate>Thu, 16 Apr 2026 13:05:32 GMT</pubDate>
            <atom:updated>2026-04-16T13:18:04.468Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DgVyWR3bf64EJoKN" /></figure><p>There’s a lot of noise around AI in manufacturing right now. Every vendor deck, every conference keynote, every LinkedIn post seems to promise that AI will transform your plant.</p><p>And honestly, most of the time, I just smile and scroll past.</p><p>Not because AI doesn’t matter. It does. But because the way people talk about it is disconnected from what actually happens on a plant floor.</p><p>So let’s break it down in a simple, honest way. What Industrial AI really means. What it looks like in practice. And why it’s both powerful and frustrating at the same time.</p><h3>First, Let’s Be Honest About the Hype</h3><p>When people talk about AI in manufacturing, they usually imagine something futuristic. A control room where algorithms run everything. Machines predict failures on their own. Operators just watch dashboards.</p><p>That’s not reality.</p><p>What I’ve seen is very different. Plants struggling to get clean data from a single line. Teams that don’t agree on what “downtime” means. Historians full of data, but no one knows which tags are still valid.</p><p>That’s the real starting point. Not AI models. Data chaos.</p><h3>What Industrial AI Actually Is</h3><p>Industrial AI is just artificial intelligence applied to manufacturing. That includes machine learning, computer vision, analytics, and sometimes generative AI.</p><p>The goal is simple. Use data from machines, sensors, and people to make better decisions.</p><p>That shows up in real use cases like:</p><ul><li>Predictive maintenance</li><li>Quality inspection</li><li>Process optimization</li><li>Energy management</li><li>Production planning and scheduling</li></ul><p>For example, computer vision systems can detect defects on a production line 24/7. In some cases, they reduce scrap in just a few months. Predictive models can detect early signs of equipment failure and prevent downtime before it happens.</p><p>But here’s what often gets skipped.</p><p>Industrial AI is not about the algorithm. It’s about the data behind it.</p><p>If you want to predict a pump failure, you need vibration, temperature, flow, and current data. But more importantly, you need context. Which pump. Which line. Which product. Which shift.</p><p>Without that, the model is just guessing.</p><p>That’s why I always say. AI is the last mile. The first 99 miles are data.</p><h3>Why Industrial AI Is Different</h3><p>Industrial AI is not the same as enterprise AI.</p><p>It operates in a very different environment.</p><ul><li><strong>Real-time constraints</strong>. Decisions often need to happen in milliseconds.</li><li><strong>Physical consequences</strong>. A wrong decision can impact safety, quality, or production.</li><li><strong>Messy data</strong>. Signals are noisy, incomplete, and inconsistent.</li><li><strong>Legacy systems</strong>. PLCs, SCADA, MES, and historians, many of them old and hard to integrate.</li><li><strong>Regulation</strong>. In industries like pharma, everything must be validated, traceable, and auditable.</li></ul><p>So when someone says “just plug AI into the plant,” the real question is. Plug it into what?</p><h3>The Foundation Nobody Talks About</h3><p>Here’s the uncomfortable truth.</p><p>Most companies are not ready for Industrial AI.</p><p>Not because they lack ambition. But because they haven’t built the data foundation.</p><p>In many projects, the first months are spent on things like:</p><ul><li>Connecting machines and PLCs</li><li>Standardizing tag naming</li><li>Cleaning and validating data</li><li>Building reliable data pipelines</li><li>Defining a consistent data model</li></ul><p>It’s not exciting work. But it’s the work that makes everything else possible.</p><p>You can’t build AI on top of broken data. It doesn’t matter how advanced the model is.</p><p>Think of it like building a house. You don’t pour the foundation after the walls are up. But that’s exactly what happens in many AI projects.</p><h3>The Role of the Unified Namespace</h3><p>One of the biggest enablers for Industrial AI is the Unified Namespace, or UNS.</p><p>UNS creates a single, structured, real-time layer where all industrial data lives. Instead of pulling data from dozens of systems, everything flows through one consistent model.</p><p>This changes everything.</p><ul><li>Data becomes easier to access</li><li>Context is preserved</li><li>Systems can publish and subscribe in real time</li><li>AI models don’t need heavy data wrangling</li></ul><p>Without this, teams spend most of their time just preparing data.</p><p>With it, data is already usable.</p><p>That’s what makes AI scalable across sites.</p><h3>Architecture. From Machine to Model</h3><p>A typical Industrial AI setup follows a simple flow.</p><ul><li>Data is collected from machines and systems</li><li>It is cleaned, contextualized, and structured</li><li>It feeds dashboards, alerts, and AI models</li><li>Models are trained in the cloud</li><li>Models are deployed at the edge for real-time decisions</li></ul><p>This hybrid setup balances speed and scalability.</p><p>It also allows plants to act locally while learning globally.</p><h3>Why Most Projects Struggle</h3><p>Even with all the potential, many Industrial AI projects fail to scale.</p><p>The reasons are usually the same:</p><ul><li>Data quality issues</li><li>Lack of clear ROI</li><li>Integration challenges with legacy systems</li><li>Skills gap between IT and OT</li><li>Resistance from operators and teams</li></ul><p>The technology is not the hardest part.</p><p>The combination of data, process, and people is.</p><p>If you skip that, the project won’t last.</p><h3>Where Industrial AI Is Going</h3><p>Things are changing fast.</p><p>Cloud platforms and edge computing are making it easier to process industrial data at scale. Tools are improving. Architectures are becoming more standardized.</p><p>And new capabilities are emerging.</p><p>For example, operators interacting with systems using natural language. Asking simple questions like, “Why did this line stop?” and getting clear answers based on real data.</p><p>This is already happening in early forms.</p><p>Not perfect yet. But moving in that direction.</p><h3>How to Get Started</h3><p>If you’re thinking about Industrial AI, keep it simple.</p><ul><li>Start with data. Fix connectivity and structure first.</li><li>Pick one use case. Something small and measurable.</li><li>Involve operators early. They understand the process better than anyone.</li><li>Be honest about your data. Clean it before building models.</li><li>Think long-term. This is a capability, not a one-time project.</li></ul><p>Most importantly, don’t rush to AI.</p><p>Build the foundation. Then scale.</p><h3>Final Thought</h3><p>Industrial AI is real. It delivers value. I’ve seen it work in real plants, with real people.</p><p>But it only works when the foundation is there.</p><p>So the next time someone tells you AI will transform your factory overnight, just ask one question.</p><p>What does your data layer look like?</p><p>That’s where the real conversation starts.</p><p><em>Originally published at </em><a href="https://iiotblog.com/2026/04/16/industrial-ai-explained-why-data-is-the-real-starting-point/"><em>https://iiotblog.com</em></a><em> on April 16, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bd697c908130" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[IIoT Total Cost of Ownership: Cloud, Licenses, and the Hidden Costs]]></title>
            <link>https://medium.com/@henry.costa/iiot-total-cost-of-ownership-cloud-licenses-and-the-hidden-costs-c63bba4f882a?source=rss-e2cc0bd3556a------2</link>
            <guid isPermaLink="false">https://medium.com/p/c63bba4f882a</guid>
            <category><![CDATA[data-egress]]></category>
            <category><![CDATA[total-cost-of-ownership]]></category>
            <category><![CDATA[iiot]]></category>
            <category><![CDATA[azure]]></category>
            <category><![CDATA[aws]]></category>
            <dc:creator><![CDATA[Henry Costa]]></dc:creator>
            <pubDate>Tue, 14 Apr 2026 18:52:00 GMT</pubDate>
            <atom:updated>2026-04-14T19:11:23.876Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ka4YlwKou-jYSQoY" /></figure><p>Let’s talk about the real cost of Industrial IoT. Not the version you see in vendor slide decks, but what actually impacts your budget once you move beyond a pilot and start scaling.</p><p>If you are working on IIoT initiatives, especially in a multi-site or regulated environment, this is the part that tends to catch teams off guard. The cost is not just cloud or licenses. It is everything that builds around it over time.</p><h3>The Cloud Looks Cheap at First. Then Reality Kicks In</h3><p>Cloud-native IIoT is often sold as simple and flexible. You send data, you run analytics, and you pay for what you use.</p><p>That part is true. But incomplete.</p><p>“What you use” includes more than storage and compute. The biggest surprise is usually data egress. Getting data out of the cloud, whether for analysis, integration with MES, or regulatory reasons, can quickly become expensive.</p><p>High-frequency manufacturing data makes this worse. Streaming data into the cloud is often cheap. Pulling it back out, or moving it across regions, is where costs start to grow fast. Those small charges per gigabyte do not stay small when you scale.</p><h3>Licensing Is Simple. Until It Isn’t</h3><p>Most platforms present a clean pricing model. Something like a few dollars per device per month. That works well for pilots.</p><p>But once you move into production, things change.</p><p>You start needing customization. Legacy systems need to be integrated. Dashboards need to be adapted to real operations. That means development work, and that cost adds up quickly.</p><p>Integration is another layer. Connecting MES, historians, or ERP systems is not a small effort. API management, testing, and validation all bring additional cost. And in many cases, platform pricing is not as flexible as it looks. You may not be limited by devices, but by payload size, storage, or analytics features. That often forces upgrades earlier than expected.</p><h3>The Costs That Do Not Show Up in the Proposal</h3><p>The biggest impact on your budget usually comes from areas that are not highlighted upfront.</p><p>Legacy integration is one of them. Making older PLCs or DCS systems work with modern platforms requires gateways, middleware, and time. In regulated environments, it also requires qualification and documentation.</p><p>Maintenance is another. IIoT platforms need updates, patches, and sometimes hardware changes. Over time, these small efforts become a steady cost.</p><p>Data management is often underestimated. As your deployment grows, so does the volume of data. Storage, processing, and analytics costs can exceed expectations by a large margin.</p><p>Security and compliance add another layer. Cybersecurity tools, audits, and validation activities are not optional in many industries. They are ongoing and must be planned from the start.</p><h3>Cloud Platform Pricing. The “Unit Trap”</h3><p>Cloud providers have their own patterns that are not always obvious.</p><p>With services like AWS IoT Core, cost is tied to activity. Every message, every rule execution, every update adds to the bill. As devices become more active, costs can grow faster than expected.</p><p>On the Microsoft Azure side, pricing often works in tiers. You may end up paying for capacity you are not fully using, or hitting limits that force you to upgrade sooner than planned.</p><p>In both cases, premium support is another factor. It is often required in enterprise environments, and it can add a noticeable percentage to your total cost.</p><h3>The Cost of Keeping Everything Running</h3><p>There is also a cost that is easy to overlook. The cost of operating the platform.</p><p>Managing devices, tracking usage, optimizing configurations, and handling lifecycle events takes effort. Without automation, this quickly becomes a full-time job for multiple people.</p><p>At scale, this operational overhead can represent a significant portion of your total spend. It is not just about technology. It is about people and processes needed to keep everything stable.</p><h3>What This Means in Practice</h3><p>When you look at IIoT cost, you need to think beyond the initial setup.</p><p>A pilot with a few devices will almost always look affordable. The real question is what happens when you scale to thousands of devices, across multiple sites, with real operational requirements.</p><p>That is where data volume increases. Integration becomes more complex. Compliance requirements become stricter. And costs start to behave differently.</p><h3>My Honest Take</h3><p>Most IIoT projects underestimate total cost of ownership. Not by a small margin.</p><p>In many cases, the gap is around 30 to 50 percent. In regulated environments, it can go even higher.</p><p>The issue is not bad technology. It is incomplete planning.</p><p>That is why it is critical to model real scenarios early. Not just a pilot. A scaled version of your future state. Include data flows, integrations, security, and operations.</p><p>It is not about being cautious. It is about making sure your project can survive when it grows.</p><p>Because in IIoT, scaling is where the real story begins.</p><p><em>Originally published at </em><a href="https://iiotblog.com/2026/04/14/iiot-total-cost-of-ownership-cloud-licenses-and-the-hidden-costs/"><em>https://iiotblog.com</em></a><em> on April 14, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c63bba4f882a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Five Conversations That Changed How I Approach IIoT Projects]]></title>
            <link>https://medium.com/@henry.costa/five-conversations-that-changed-how-i-approach-iiot-projects-454f54e532f3?source=rss-e2cc0bd3556a------2</link>
            <guid isPermaLink="false">https://medium.com/p/454f54e532f3</guid>
            <category><![CDATA[digital-transformation]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[industrial-iot]]></category>
            <category><![CDATA[lessons-learned]]></category>
            <category><![CDATA[iiot]]></category>
            <dc:creator><![CDATA[Henry Costa]]></dc:creator>
            <pubDate>Mon, 06 Apr 2026 15:21:24 GMT</pubDate>
            <atom:updated>2026-04-06T15:25:37.558Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*tG6jZ1y13mUbD0iO" /></figure><p>Some of the biggest changes in how I work didn’t come from reading a whitepaper or going to a conference. They came from conversations. Quick ones, sometimes. A comment over coffee, a pushback during a design review, a frustrated operator explaining why nobody trusts “the new system.”</p><p>After 20+ years working on industrial data and manufacturing projects, I can point to five conversations that shaped how I think about this work.</p><h3>1. “You’re solving a problem nobody asked you to solve.”</h3><p>This one came early. I was maybe three years into manufacturing IT, excited about connecting everything. I had put together this great architecture, data flowing from the shop floor all the way up to the ERP. I was proud of it.</p><p>A plant manager looked at it, paused, and said, “That’s nice. But who asked for this?”</p><p>He wasn’t being rude. He just wanted to know which operator, which supervisor, which engineer had said, “I need this data.” And I didn’t have a good answer. I had built something because I could, not because someone needed it.</p><p>That conversation taught me to always start with the problem. Not the architecture. Not the platform. The problem. Who is struggling, and why? If you can’t answer that clearly, you’re not ready to build anything.</p><h3>2. “We already have a Historian. Why do we need another system?”</h3><p>I heard this one at a large manufacturing site during a workshop. An automation engineer was genuinely confused about why we were adding a new cloud data layer when they already had a system collecting thousands of data points.</p><p>And honestly, it was a fair question.</p><p>The real answer wasn’t that their system was bad. It was doing exactly what it was built to do, storing time-series data for process control. But the new needs, things like quality predictions, efficiency dashboards, comparing performance across sites, required something different. They needed data with context, not just raw numbers.</p><p>That conversation pushed me to get better at explaining <em>why</em> a shared data layer matters, in plain words. Not “we need semantic interoperability.” More like, “Right now, if someone asks what the batch yield was last Tuesday, three people give three different numbers. We need one answer everyone trusts.”</p><p>If you can’t explain it to the engineer who’s been running that plant for 15 years, you don’t understand it well enough yourself.</p><h3>3. “Stop showing me dashboards. Show me decisions.”</h3><p>This came from a VP of Operations during a project review. We had just set up a bunch of real-time dashboards, live data streaming in, color-coded alerts, the works. The whole team was proud.</p><p>The VP looked at it and said, “It’s pretty. But what decision does this help me make at 7 AM on a Monday?”</p><p>That hit hard. Because he was right. We had built visibility without purpose. Lots of data on the screen, but no clear action tied to it.</p><p>After that, I started building every dashboard, every report, every screen around one question: what’s the decision? If the answer is “it’s just nice to know,” that’s a warning sign. Nice-to-know doesn’t survive the first budget cut.</p><p>Now, whenever I design something, I think in terms of decisions per shift, not data points per second.</p><h3>4. “You IT guys always leave after go-live.”</h3><p>An operator told me this during a site visit, about two years after a system had gone live. He wasn’t angry, just tired of it. He said the system worked fine for a few months, then things started breaking. Small things. A data mapping changed, a dashboard stopped updating, a new product was added but nobody fixed the setup. And no one came to help.</p><p>That conversation changed how I think about project scope. I stopped treating launch day as the finish line. Now I push hard for support plans, clear ownership, simple guides, and training that actually works, not just a document nobody opens.</p><p>The truth is, most projects don’t fail at launch. They fail six months later, quietly, when the people who built it have moved on and the people who use it don’t know how to keep it running.</p><p>If you want your project to last, plan for the boring part. The upkeep. The fixes. The handoffs. That’s where real value lives.</p><h3>5. “I don’t care about the technology. I care about my people.”</h3><p>This one came from a site director at a manufacturing plant. We were talking about a big modernization effort, replacing old control systems, adding new computing at the edge, rolling out analytics. Big scope, big budget.</p><p>I was deep into the technical details when she stopped me and said, “I don’t care about the technology. I care about my people. Will they understand it? Will they trust it? Will it make their day easier or harder?”</p><p>That was a turning point. Because she was absolutely right. You can build the most elegant system in the world, but if the operator on the night shift doesn’t trust the data, or the maintenance tech can’t figure out what went wrong, or the engineer feels like the system was forced on them, it won’t work.</p><p>Since then, I always make change management and training a core part of the work, not something we think about at the end. I spend time on the floor with operators before designing anything. I ask them what’s annoying, what’s slow, what they wish they had. And then I try to build that.</p><p>Technology is the easy part. People are the hard part. And the important part.</p><h3>So, What Did I Actually Learn?</h3><p>Looking back, these five conversations all point to the same thing. This work is not a technology problem. It’s a people problem wrapped in technology.</p><p>The best systems I’ve seen weren’t the most advanced. They were the ones where someone took the time to understand the real problem, explain the solution in simple words, build for decisions instead of data, plan for what happens after launch, and put people first.</p><p>That’s it. No magic framework. No fancy method. Just listen more, assume less, and never forget that every sensor, every data point, every dashboard exists to help a person do their job better.</p><p>If you’re early in your journey, my advice is simple. Go have more conversations. Not with vendors. Not with analysts. With the people on the floor. That’s where the real insights are.</p><p><em>Originally published at </em><a href="https://iiotblog.com/2026/04/06/five-conversations-that-changed-how-i-approach-iiot-projects/"><em>https://iiotblog.com</em></a><em> on April 6, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=454f54e532f3" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>