Agentic AI in Engineering and Manufacturing: Industry Perspectives on Utility, Adoption, Challenges, and Opportunities

This qualitative study of over 30 industry interviews reveals that while agentic AI offers significant potential for orchestrating complex manufacturing workflows, its widespread adoption is currently hindered more by fragmented data, legacy toolchain limitations, and organizational governance gaps than by model capabilities, necessitating a staged progression toward automation grounded in robust verification and human-in-the-loop frameworks.

The Gist

The Big Picture: The "Super-Intern" That Needs a Map

Imagine the engineering and manufacturing world as a massive, bustling construction site. For decades, the workers (engineers) have been building skyscrapers, bridges, and rockets using blueprints, calculators, and specialized tools.

Now, a new type of worker has arrived: Agentic AI.

Think of this AI not just as a smart calculator, but as a super-intern. This intern is incredibly fast at reading manuals, organizing files, and suggesting ideas. However, unlike a human engineer, this intern cannot yet "think" like a master builder. It can't look at a wobbly bridge and intuitively know why it might fall without being told exactly what to look for.

This paper is the result of the authors interviewing 33 experts (from NASA to small machine shops) to answer one question: "How do we get this super-intern to actually help us build things without causing a disaster?"

---

1. What Can the AI Do Right Now? (The "Paperwork" Phase)

The paper finds that the AI is currently great at the boring, repetitive stuff that humans hate doing.

• The Analogy: Imagine you are a chef. You love cooking the main course, but you hate chopping 500 onions or organizing the spice rack. The AI is the perfect sous-chef for the chopping and organizing.
• Real-world examples:
  • Reading the Fine Print: Engineers spend hours reading 500-page requirement documents. The AI can read them in seconds and say, "Hey, if we change this bolt, it breaks that electrical wire."
  • Data Entry: Filling out forms, copying data from old PDFs to new spreadsheets, and checking if a part number matches a drawing.
  • Finding the Needle in the Haystack: If a company has 20 years of data scattered across hard drives, emails, and dusty filing cabinets, the AI can find the specific part you need much faster than a human digging through boxes.

The Verdict: The AI is a fantastic assistant for "data-heavy" and "repetitive" tasks. It frees up humans to do the creative, high-level thinking.

---

2. What Can't It Do Yet? (The "Black Box" Problem)

The AI struggles with the "hard stuff" where safety is critical and the rules are complex.

• The Analogy: Imagine asking the super-intern to design the engine for a jet plane. It might suggest a design that looks cool and follows the rules, but it might miss a subtle vibration that causes the engine to explode at 30,000 feet.
• The Problem:
  • No "Gut Feeling": Human engineers have "spatial reasoning." They can look at a 3D model and visualize how parts fit together, how heat moves, or how metal bends. The AI is still bad at this "3D thinking."
  • The Black Box: If a traditional computer program fails, you can look at the code and see exactly where it went wrong. If the AI fails, it's like a magic trick. You get a result, but you don't know why it happened. In engineering, if you can't explain why a bridge is safe, you can't build it.
  • Safety First: Because of this, no one is letting the AI drive the car alone yet. It can suggest the route, but a human must keep their hand on the wheel.

---

3. The Three Big Roadblocks (Why We Aren't Using It Everywhere)

The paper identifies three main reasons why companies aren't just flipping a switch and letting the AI take over.

A. The "Messy Attic" (Data Problems)

• The Analogy: Imagine trying to teach a robot to cook, but all your recipes are written in crayon on napkins, some are in a different language, and the rest are stuck in a locked safe.
• The Reality: Engineering data is a mess. It's scattered across old computers, PDFs, and handwritten notes. It's often "machine-unfriendly." Before the AI can learn, companies have to spend months cleaning up their data. Also, because this data is often secret (like military designs), it can't be sent to the cloud to be processed by big AI companies. It has to stay in a "digital bunker" (on-premise servers).

B. The "Old Tools" (Legacy Software)

• The Analogy: Imagine trying to plug a modern, high-speed electric car into a 1970s gas pump. The connection just doesn't fit.
• The Reality: Many engineering tools (CAD software, manufacturing machines) were built decades ago. They were designed for humans to click buttons with a mouse, not for AI to talk to them via code. The AI wants to say, "Open this file, change this number, and save it." But the old software says, "I don't speak that language; you have to click the menu manually."

C. The "Trust Gap" (Culture and Fear)

• The Analogy: You wouldn't let a stranger drive your family car just because they have a great GPS. You need to know they are safe, responsible, and that you can stop them if they go off a cliff.
• The Reality: Engineers are trained to be risk-averse. They need to know exactly why a decision was made. If an AI suggests a design, the engineer needs to be able to prove to a regulator (or a judge) that it's safe. Right now, the AI is too "guessy" (probabilistic) for high-stakes jobs. Also, many engineers don't know how to use these tools properly yet, leading to either over-trusting them or ignoring them completely.

---

4. The Future: How Do We Fix It?

The paper suggests that to make Agentic AI truly useful, we need a few "breakthroughs":

1. Standardized Handshakes: We need a universal language (like a standard USB port) so that AI can talk to any engineering software without needing custom code for every single machine.
2. The "Truth Detector": We need new ways to verify that the AI isn't hallucinating. We need tools that can say, "I am 99.9% sure this design is safe, and here is the proof."
3. Better 3D Brains: AI needs to get better at understanding space, physics, and how things fit together, not just reading text.
4. Human-in-the-Loop: The best future isn't AI replacing engineers; it's AI acting as a "co-pilot." The AI does the heavy lifting and the boring work, and the human engineer acts as the captain, making the final call and taking responsibility.

The Bottom Line

The paper concludes that AI is not a magic wand that will solve everything tomorrow.

It's more like a powerful new engine that we are trying to install in an old car. The engine is amazing, but the car's wiring is old, the fuel is messy, and the driver is nervous.

To get the full benefit, companies need to:

• Clean up their data (organize the attic).
• Update their software tools (fix the wiring).
• Train their people (teach the driver).
• And most importantly, keep a human in the driver's seat until the AI proves it can handle the most dangerous turns safely.

The goal isn't to replace the engineer; it's to give them a super-powerful assistant so they can build better, safer, and faster things.

Technical Summary

Based on the paper "Agentic AI in Engineering and Manufacturing: Industry Perspectives on Utility, Adoption, Challenges, and Opportunities," here is a detailed technical summary.

1. Problem Statement

The engineering and manufacturing sectors face a critical convergence of challenges: a projected shortage of 1.9 million manufacturing engineers by 2033, an aging workforce leading to the loss of tacit knowledge, and a structural shift toward resilient, regionalized supply chains. While Artificial Intelligence (AI) is identified as a key solution to augment productivity, the industry struggles to move beyond basic automation to Agentic AI—systems capable of planning multi-step actions, interacting with external tools, and adapting to feedback.

The core problem is that current AI adoption in engineering is constrained not by model capability, but by ecosystem readiness. Key barriers include:

• Data Fragmentation: Engineering data is scattered across unstructured formats (PDFs, legacy databases) and is often "machine-unfriendly."
• Legacy Toolchains: Critical engineering tools (CAD, CAE, CAM, ERP) lack programmatic interfaces (APIs) required for agents to orchestrate workflows.
• Trust and Governance: The probabilistic nature of Large Language Models (LLMs) conflicts with the deterministic, safety-critical requirements of engineering (e.g., aerospace, defense), where explainability and auditability are non-negotiable.
• Security: Strict regulatory constraints (ITAR, GDPR, CMMC) prevent the use of cloud-based AI models for proprietary data.

2. Methodology

This is an exploratory, qualitative state-of-practice study conducted by researchers from MIT and industry partners.

• Data Collection: The authors conducted 33 semi-structured interviews with 28 organizations across four distinct stakeholder groups:
  1. Large Enterprises: (e.g., NASA, GE Vernova, major defense contractors).
  2. Small/Medium Enterprises (SMEs): (e.g., additive manufacturing firms, contract manufacturers).
  3. AI Developers: (e.g., LangChain, C3 AI, Synera).
  4. CAD/CAM/CAE Vendors: (e.g., Autodesk, PTC, Siemens).
• Participants: Included C-suite executives, CTOs, heads of AI, and frontline engineers.
• Analysis: An iterative, inductive approach was used to identify recurring themes regarding task landscapes, barriers, and enablers. The study intentionally included both AI-enthusiastic and skeptical organizations to reduce confirmation bias.

3. Key Contributions

The paper makes three primary contributions to the field:

1. Synthesis of Industry Perspectives: It characterizes how agentic AI is currently understood and adopted, moving beyond theoretical potential to practical reality.
2. Task Landscape Mapping: It delineates the "engineering task landscape," distinguishing between well-posed tasks (repetitive, data-intensive, structured) suitable for near-term AI and challenging tasks (safety-critical, spatial reasoning, multi-physics) that require future breakthroughs.
3. Identification of Breakthrough Requirements: It surfaces the specific technical, infrastructural, and socio-technical breakthroughs needed to unlock broader utility, shifting the focus from "better models" to "better integration."

4. Key Results and Findings

A. The Task Landscape for Agentic AI

• Near-Term Opportunities (Well-Posed Tasks):
  • Repetitive/High-Volume Work: Requirements processing, traceability matrix creation, and data entry (e.g., filling ERP systems from PDFs).
  • Data-Intensive Synthesis: Retrieving historical mission data, generating documentation, and cross-referencing component datasheets.
  • Process Orchestration: Coordinating multi-step workflows like Request for Quotation (RFQ) processes between OEMs and suppliers.
• Challenging Future Tasks:
  • Safety-Critical Decision Making: AI currently serves as an advisory tool (Human-in-the-Loop) rather than an autonomous decision-maker in high-stakes scenarios.
  • Spatial and Physical Reasoning: Current models struggle with 3D geometric understanding, CAD-to-CAM translation, and multi-physics reasoning (e.g., thermal + mechanical + electrical coupling).

B. Primary Barriers to Adoption

• Data Constraints:
  • Fragmentation: Data exists in silos (personal drives, SharePoint, ad-hoc files) with inconsistent naming conventions.
  • Representation: Critical specs are locked in unstructured PDFs or "tribal knowledge" in experts' heads, making them inaccessible to AI.
  • Quality: Historical designs often contain ad-hoc compromises; training on them risks teaching AI to replicate past errors.
• Technical & Infrastructural Gaps:
  • Legacy Integration: Most engineering tools lack REST APIs or headless operation modes, forcing agents to rely on brittle UI automation.
  • Verification Gap: Engineering requires deterministic, repeatable outputs. The probabilistic nature of LLMs creates a "black box" problem that hinders trust in safety-critical applications.
  • Spatial Reasoning: Models lack the ability to reason about 3D space, assembly constraints, and manufacturability.
• Organizational & Cultural Barriers:
  • AI Literacy Gap: Few engineers possess the skills to build, configure, or critically evaluate AI systems.
  • Security/Compliance: Regulations (ITAR, CMMC) often mandate air-gapped, on-premise deployments, limiting access to state-of-the-art cloud models.
  • Cultural Heterogeneity: A spectrum exists from risk-averse legacy firms to fast-moving startups, creating different adoption speeds and trust levels.

C. Enablers and Adoption Strategies

• Human-in-the-Loop (HITL): Non-negotiable for safety. AI is viewed as an augmentation tool that accelerates human decision-making, not a replacement.
• Alignment with Existing Governance: Successful adoption integrates AI into established engineering review processes (e.g., PDR, CDR, TRR) rather than bypassing them.
• Trust Building: Achieved through "communities of practice," internal AI champions, and demonstrating reliability on low-stakes tasks before scaling to high-stakes ones.
• Architecture: A shift toward on-premise, edge, or air-gapped deployments to satisfy security requirements.

5. Significance and Future Directions

The paper argues that the path to Agentic AI in engineering is not a linear progression of model improvement but a staged organizational transformation.

Key Breakthroughs Needed:

1. Standardized Agent Interfaces: Protocols (like Model Context Protocol - MCP) to allow agents to interact with diverse engineering tools programmatically.
2. Robust Verification Frameworks: New paradigms for testing probabilistic systems in safety-critical contexts (e.g., simulation-based validation, formal methods for uncertainty).
3. Physics-Informed Architectures: AI models that natively understand 3D space, physical constraints, and multi-domain interactions.
4. Data Transformation Pipelines: Automated systems to convert unstructured engineering artifacts (PDFs, tribal knowledge) into structured, machine-readable formats.
5. Governance Integration: Frameworks that seamlessly extend existing engineering review boards to include AI-generated outputs.

Conclusion:
The study concludes that while AI offers immense potential to address workforce shortages and efficiency gaps, its utility is currently limited to structured, low-consequence tasks. Widespread adoption of Agentic AI depends less on model capabilities and more on infrastructure readiness (data quality, API access), verification standards, and cultural trust. The most successful organizations are those treating AI adoption as a systemic change management challenge, prioritizing data infrastructure and workforce literacy over simple tool procurement.