By Jag Gattu, CEO, UptimeAI
This post originally appeared on LinkedIn.
My first post on this topic highlighted the shift we’re seeing in how many tools we build in house vs. buy, and where we draw the line. At UptimeAI, our line sits at the transition point from retrieval to reasoning.
I took a step back to see where others in the market are drawing the line. While I found general excitement about the potential of BIY (Claude Code and the likes), I found that leaders are being appropriately calculated in their consideration of these projects. There isn’t just one line that they’re drawing for when to build versus buy a solution, but many, and the lines link back to the same challenges that have plagued digital projects in our industry for the past decade.
Here are the three biggest challenges with BIY agentic AI projects in the process industries.
1. Governance
In a recent blog post, Verdantix highlighted that AI sprawl is becoming a C-suite problem. Sprawl is a second-order cost to unlimited BIY development. Every vibe-coded agent is also one more tool requiring governance — and a new source of shadow AI, data leakage, and regulatory exposure if it’s not accounted for. Even Amazon has acknowledged internally that “the AI boom is driving duplication and fragmentation, not less of it.” Building a tool and then governing it responsibly, indefinitely, in a live plant are two different jobs, and the governance doesn’t get easier just because the build got faster. Industrial software companies have been operating in this area for decades and have the data governance and cybersecurity infrastructure already in place to deploy agentic AI projects fast without risk.
2. Scale
According to McKinsey’s 2025 State of AI report, less than 10% of organizations currently experimenting with building AI agents are scaling them. BIY pilot development is easier than production deployment. You can have a demo working reliably on clean, well-behaved offline data in a matter of hours. Standing up something that holds up to the challenges of real operations data — fragmented, inconsistently modeled ERP, MES, historian, and SCADA data, spanning reliability, process, and operations teams — is another. Agents built on top of fragmented data don’t fix fragmented decision-making, they automate and amplify it. That’s true whether you built the agent or bought it. The difference is whether you’re building the curation, integration, and verification layer from scratch, or standing on one that’s already been proven across other plants like yours.
3. Impact
In a recent LNS Research blog post, Research Analyst Vivek Murugesan emphasizes the importance of “leading with business problems and not technology itself.” Are you looking for a personal productivity tool or a way to make margin-impacting operations decisions? A tool that surfaces the right document or summarizes a work order can be easily built, and it is useful, but it isn’t the same as a system that provides operators with confident decisions to high-stakes calls. Margin is lifted through improvements to availability, quality, and reliability. Personal productivity can support improvements to these areas, but it won’t move the needle on margin the way the components of operational excellence can. An enterprise deployment of a Root Cause Agent, for example, requires automated expert reasoning, robust data handling, and continuous learning and improvement loop. That’s not retrieval, it’s judgment, and it’s proven to provide rapid technology ROI through improvements to diagnosis time, accuracy, and avoided downtime.
Draw YOUR right lines
The challenges raised above aren’t reasons to avoid BIY altogether. There are problems where in-house solutions are sufficient, like our sales enablement tool I mentioned in my last post. Retrieval has become easy to build and scale in-house. But when making build v. buy decisions around reasoning and judgement—the agentic AI capabilities required to make impactful decisions in a regulated industrial environment—trusted partners can effectively overcome the governance, scale, and impact hurdles.
