Assistant
PTC/User 2026 Recap: From Digital Thread to AI Agents — A Practical Playbook
At PTC/User 2026 in Las Vegas, one theme cut through every keynote, panel, and hallway conversation: manufacturers are urgently trying to turn AI prototypes into real, scalable engineering value. Curiosity is high, but mature execution remains rare. Many engineering organizations are still stuck in pilot purgatory — three things almost working, none in production, none scaling.
Sommy Boucansaud and I shared a session on this — From Digital Thread to AI Agents: Practical Approaches for Engineering Productivity — drawing on more than two decades of enterprise search and knowledge management work at Sinequa with some of the world’s most demanding manufacturers, and on the agentic AI deployments layered on that foundation over the past few years (roughly since neural search and modern LLMs made genuinely agentic systems viable in production). Customers like Cummins, Boeing, Airbus, and Mitsubishi Power shaped the patterns presented in the session: the architectural choices, the governance discipline, and the sequencing that separates the 5% of pilots that scale from the 95% that don’t.
But the most valuable hours of the week were the ones spent listening.
Why the Knowledge Layer Is Reshaping Engineering AI
The single most important line heard all week wasn’t said by ChapsVision. It was said in PTC’s AI Strategy session, almost in passing:
“Intelligence resides in data, not agents.”
Six words. They explain why most enterprise AI pilots stall.
PTC’s strategy team described a three-layer platform underneath everything they’re building — a relational data layer for first-class entity relationships, vector stores for semantic retrieval, and a semantic layer that maps natural language to each customer’s specific vocabulary. In their architecture, the agent is the visible part. The data underneath is what actually does the reasoning.
That’s the same thesis the ChapsVision session was built around — what we call the knowledge foundation. Watching a roomful of engineering leaders hear that conclusion from PTC, then again from us, then again from PTC’s Joseph June and François Lamy in the AI trust panel later that day, was clarifying. Different vendors, different products, same answer.
The argument has shifted. We’re no longer debating whether you can run enterprise AI without a knowledge layer. The argument now is what kind of knowledge layer, and how you operationalize it.
The Reality Check: Most AI Pilots Fail in the Same Place (It’s a Data Problem)
Despite massive investment, 95% of enterprise AI pilots never reach production — a statistic both PTC and ChapsVision cited from independent industry sources. The failure pattern is remarkably consistent: the pilot was built without a governed retrieval layer underneath it.
Without that foundation, agents face three predictable problems:
- Hallucinations: confidently wrong answers, because the model has no grounding in the customer’s actual content.
- Security gaps: surfacing documents the user shouldn’t see, because permissions weren’t inherited at the index.
- Lack of cross-system reasoning: returning a partial answer because the data needed to answer fully lives in a different system that the agent can’t see.
Sommy walked through the Sinequa Windchill security connector as one example of the unsexy but load-bearing engineering this requires — inheriting source-system permissions at the index, federating early-binding and late-binding security models, respecting ITAR and EAR labels orthogonally to ACLs. None of that work lives in the model. All of it determines whether your agents are deployable in a regulated environment.
As Joseph June, PTC’s AI Strategy Owner, put it during the trust panel: “Trust comes from how you architect the system, not from which model you pick.”
VisionCast ON-DEMAND
VisionCast: Agentic AI You Can Actually Trust
See how enterprises move beyond AI experiments to trusted, scalable AI agents. Watch the on-demand session.
Inside a Real Deployment: Agentic AI in Manufacturing Engineering
At the center of the session was a case study from one of the world’s largest engine manufacturers — Cummins — whose team has been an extraordinary partner in pressure-testing this approach over the past two years.
Their team inventoried their AI candidate use cases across the engineering organization and matched each one to the simplest pattern that could solve it. The result surprised the room: only a small minority required full agentic orchestration. Most landed as document experts, subject experts, or pre-defined workflows. The bulk of the ROI came from the simpler patterns.
In one published benchmark, the contrast was even sharper: a leading general-purpose copilot landed in the 30th percentile for accuracy on complex engineering questions, while the same questions answered through a grounded enterprise knowledge layer cleared the 90th percentile. That’s the difference between a failing grade and an A — and it’s almost entirely a function of the data foundation underneath, not the model on top.
Cummins Builds a Knowledge Fabric with Sinequa
Read the case studyHow Manufacturers Can Replicate This Success
The session closed with four concrete actions for organizations ready to move beyond AI pilots — focused on the workflow, the retrieval foundation, and the governance required for scale.
The clear message: organizations that build on a proven retrieval platform scale agentic AI faster and more reliably than those starting from scratch with each new pilot. The PTC ALM team illustrated exactly this pattern in their breakout, with their Requirements and Test Case Assistants delivering meaningful engineering productivity gains using AI assistants — not full agents — and a clear sequencing roadmap that ships the simpler patterns first and layers agentic capabilities on top. That’s the right cadence: ship the simpler patterns, build the foundation, then go agentic.
Key Takeaways
- ROI is real. Agentic AI is already delivering measurable engineering productivity, not just theoretical promise — and the simpler patterns (document experts, subject experts, workflows) are where most of it shows up first.
- Most problems aren’t agent problems. Match architecture to use case. The “everything is an agent” framing has crowded out the more useful framing — most engineering questions are answerable by a well-grounded retrieval pattern.
- Trust is architectural. Source-system ACLs, citations, observability, and human-in-the-loop are baked in from Day 1 — not retrofitted. Buyers fixated on which LLM are usually missing where the actual reliability comes from.
- MCP is becoming the lingua franca. PTC committed across the portfolio; ChapsAgents has been MCP-native from day one. Anyone selling you a closed system in 2026 is selling you a 2024 architecture.
- Inventory the questions that drain hours. The recurring engineering questions that should take minutes and take hours are your AI roadmap — prioritized by pain.
Conclusion
PTC/User 2026 made it clear: engineering AI is shifting from experimentation to execution. The organizations that succeed will be the ones that treat agentic AI not as a novelty, but as a strategic capability — anchored in a governed knowledge foundation, operationalized through open agent ecosystems, and matched carefully to the problems engineers actually face every day.
As the Cummins case demonstrated, scaling AI agents inside a complex manufacturer isn’t just possible — it’s happening today, in production, across design reviews, procurement workflows, and compliance investigations. And for companies ready to move beyond pilots, the path forward is clearer than ever: start with focused, high-value use cases, invest in the foundation first, and govern from Day 1.
At Sinequa by ChapsVision, this has been our mission for over two decades: to help the world’s most data-rich organizations operationalize their knowledge, govern AI that engineers can trust, and deploy intelligent systems that truly understand work. With the rise of agentic AI in manufacturing, that mission has never been more important — or more achievable.
The companies that get this right in 2026 are the ones that quietly started in 2024. For everyone else, the opportunity is now.
Thanks to everyone who came by the session, stopped at the booth to see the Digital Thread 360° demo, or pulled us aside in the hallway. The conversations were the best part of the week. If you’d like the slides, the CIMdata commentary referenced, or a follow-up conversation, get in touch.
See Enterprise Agentic AI in action
Request a demoShare
