AI is already helpful for routine, data-heavy engineering tasks, but wider gains depend less on model intelligence than on plumbing: fragmented data, closed legacy toolchains, and verification and governance gaps block higher-order automation. Firms expect a staged progression from assistive tools to trustworthy agentic workflows as infrastructure, auditability, and organizational capabilities mature.

Main Finding

Agentic AI offers sizable near-term productivity gains in engineering and manufacturing by automating structured, repetitive, and data‑intensive tasks, and its highest practical near‑term value comes from orchestrating multi‑step workflows across tools. Broader, higher‑stakes deployment is limited less by raw model capability and more by engineering‑specific frictions: fragmented/machine‑unfriendly data, legacy toolchains lacking APIs, strict security/regulatory constraints, and the need for verification, auditability, and human‑in‑the‑loop governance. Adoption is likely to follow a staged progression from low‑consequence assistance toward greater automation as trust, infrastructure, and verification mature.

Key Points

• Sample and perspective: Findings are drawn from 33 semi‑structured interviews across 28 organizations spanning large engineering firms (6), SMEs (14), AI developers (5), and CAD/CAM/CAE vendors (3).
• Task landscape:
  • Well‑posed near term: repetitive, high‑scale tasks and data‑intensive synthesis (e.g., design checks, parameter sweeps, documentation generation, code/markup assembly).
  • Higher‑value agentic use: workflow orchestration and multi‑tool coordination (agents sequencing CAD/CAE/CAM/ERP/simulation tools).
  • Safety‑critical contexts: AI currently advisory—humans retain final accountability; bounded autonomy is the norm.
• Principal technical constraints:
  • Fragmented, poorly structured, and proprietary engineering data formats (CAD files, simulation traces, shop‑floor logs).
  • Legacy tools with limited APIs and weak integration pathways.
  • Need for improved spatial and physical reasoning for many engineering problems.
• Governance, verification, and trust:
  • Reliability, verification, and auditability are central adoption requirements.
  • Human‑in‑the‑loop workflows and governance aligned with existing engineering reviews are widely favored.
• Organizational barriers:
  • Persistent AI literacy gaps and cultural heterogeneity across firms/teams.
  • Governance structures have not caught up to agentic capabilities (liability, certification, regulatory compliance).
• Enablers and breakthroughs needed:
  • Integration with traditional engineering tools and native data types.
  • Robust verification, validation, and audit frameworks suited to engineering outputs.
  • Secure, API‑accessible legacy toolchains and standardized data representations.
  • Upskilling and organizational change to embed AI into engineering processes.
• Strategic implication: staged adoption—start with low‑risk automation and build trust/infrastructure before delegating higher‑consequence tasks.

Data & Methods

• Empirical approach: qualitative, exploratory, state‑of‑practice study based on 33 interviews (30–60 minutes typically) using a semi‑structured protocol.
• Sampling: purposive recruitment of organizations with engineering/manufacturing activity and AI engagement, augmented by snowball referrals; included a mix of AI enthusiasts and skeptics to reduce bias.
• Analysis: recordings transcribed and analyzed inductively; themes surfaced iteratively across stakeholder groups; consensus coding for disagreements.
• Limitations: non‑random, relatively small qualitative sample; results reflect interviewee perspectives and industry practice at time of study rather than representative prevalence or causal inference.

Implications for AI Economics

• Productivity and total factor growth:
  • Short/medium run: automation of repetitive and data‑intensive tasks can raise engineer and shop‑floor productivity, potentially accelerating manufacturing output per worker.
  • Orchestration capabilities could unlock larger efficiency gains by reducing coordination frictions across design → simulation → manufacturing pipelines.
• Labor demand and skill composition:
  • Complementarity > substitution in many near‑term deployments: AI augments engineers (reducing drudgery, shifting tasks toward higher cognitive content).
  • Increased demand for higher‑skill roles (verification engineers, AI system integrators, data engineers) and for upskilling programs; possible displacement of lower‑skill routine tasks.
  • Distributional effects: SMEs may face higher adoption costs (integration, data cleaning) than large incumbents, potentially affecting market concentration.
• Investment and adoption economics:
  • Value capture likely to favor firms that control data pipelines, tool integration, and verification stacks—CAD/CAE vendors embedding agentic features could gain lock‑in advantages.
  • Upfront costs center on data engineering, API and legacy integration, security/compliance, and verification frameworks—these are likely binding frictions that slow diffusion more than model cost.
  • Public and private investment priorities: data standards, secure APIs, verification tooling, and workforce training are high‑ROI targets to accelerate adoption.
• Trade, reshoring, and regional competitiveness:
  • AI + automation reduces sensitivity to labor cost differentials—supporting reshoring/nearshoring trends and altering the geography of manufacturing.
  • National competitiveness may hinge on investments in data infrastructure, workforce development, and regulatory clarity.
• Regulation, liability, and externalities:
  • Strict regulatory and safety requirements raise compliance costs and slow full autonomy—this increases demand for auditability, certification processes, and liability frameworks.
  • Negative externalities (safety failures, accountability gaps) create a premium for provably verifiable systems and monitoring mechanisms.
• Research and measurement needs:
  • Economists should measure causal effects of agentic AI on productivity, employment composition, wages, and firm‑level investment decisions via field pilots and longitudinal data.
  • Cost‑benefit analyses should include integration and verification costs, not just model performance metrics.
• Policy implications:
  • Support standards for engineering data interoperability and secure tool APIs.
  • Fund development of verification/validation benchmarks for engineering agentic systems.
  • Subsidize workforce upskilling and transitional assistance to mitigate distributional harms.
  • Encourage pilot programs with rigorous evaluation to understand real‑world economic impacts before large‑scale automation.

Suggested next steps for researchers and policymakers: prioritize empirical pilots that combine agentic prototypes with investment in data/tool integration and independent evaluation, build sectoral data standards and verification benchmarks, and monitor labor market adjustments to target upskilling programs effectively.