Low-code conversational AI slashes routine admin time in one factory: supplier-email drafting fell from about 12–15 minutes to 2–3 minutes and maintenance-data retrieval time dropped roughly 50%, but gains depended on ERP access, data quality and user readiness.

Main Finding

A low-code conversational-AI layer, embedded into an industrial maintenance-support workflow, produced clear task-level efficiency gains when tied to specific, verifiable records and human validation points. Two configured agents—ManuBot (maintenance-history queries/updates) and MailBot (recurrent supplier-email handling)—cut routine times substantially (MailBot: ~12–15 min → 2–3 min per message; ManuBot: ~50% reduction in data retrieval/query time). Gains depended on data access, ERP integration, data quality, platform permissions, and user readiness.

Key Points

• Intervention: Low-code conversational-AI integrated with existing spreadsheets, email routines and partial information systems; two agents deployed:
  • ManuBot — querying/updating maintenance-history records.
  • MailBot — drafting/structuring recurrent supplier emails.
• Measured effects:
  • MailBot reduced supplier-email preparation from ~12–15 minutes to ~2–3 minutes each.
  • ManuBot halved maintenance-data retrieval and query times on average.
  • Users reported better visibility of historical malfunctions, more structured email routines, and easier access to recurrent-event information.
• Where benefits were strongest:
  • Task-bounded, language-intensive, recurrent work tied to concrete records and templates.
  • Workflows that retained human validation checkpoints (reduced hallucination risk).
• Key constraints and frictions:
  • Incomplete ERP integration limited the agents’ grounding and scope.
  • Poor or inconsistent data structure reduced reliability of outputs.
  • Platform permissions and differing user digital readiness limited uptake and uniform gains.
• Organizational lesson: Low-code deployments enable fast configuration and iterative testing with users, but organizational complementaries (data, integration, governance, training) drive realized value.

Data & Methods

• Research design: Single-case, workflow-embedded field study with sequential phases — baseline diagnosis, prototype testing, and implementation-stage evaluation.
• Evidence sources:
  • Direct workflow observations and time-motion/task comparisons.
  • User feedback and structured interviews.
  • Usage logs and records from the implemented low-code agents.
  • Comparative task-time measurements (pre/post) for specific activities (email prep, data retrieval).
• Implementation details:
  • Low-code development used to connect LLM-based conversational agents with existing spreadsheets and email pipelines (not full ERP integration).
  • Human-in-the-loop design: outputs required human review before finalization.
• Limitations:
  • Single organizational case — external validity limited.
  • Effects measured at task-level and implementation-stage; longer-run organizational impacts not observed.
  • Quantitative evidence centered on time reductions and qualitative user reports rather than randomized causal inference.

Implications for AI Economics

• Micro-productivity and task-level gains:
  • Gen-AI can deliver large per-task time savings in language- and record-intensive support tasks; aggregate economic value depends on task frequency and staff cost. Simple ROI: (time saved per task × task frequency × wage rate) − implementation/maintenance costs.
• Heterogeneity and complementarity:
  • Benefits vary across users (digital literacy, experience) and tasks; younger/less-experienced or routine-focused staff may capture different gains. Economists should model heterogeneity and complementarities (training, governance, data systems).
• Adoption cost structure:
  • Low-code platforms lower up-front software development costs and speed deployment, reducing adoption friction. However, true costs include investments in data cleaning, ERP integration, access permissions, and governance mechanisms — these can dominate.
• Human-in-the-loop as an economic boundary:
  • Keeping human validation reduces operational risk and constrains full automation; this implies reallocation (task augmentation) rather than outright labor displacement in many support roles.
• Measurement recommendations for empirical work:
  • Focus on task-level, high-frequency measures (time-per-task, error/quality rates) and instrument with system logs.
  • Capture heterogeneity (worker experience, permission levels) and integration variables (ERP access, data quality metrics).
  • Where possible, use randomized rollouts or staggered adoption to identify causal effects and spillovers (reassignment of time to other tasks, effects on decision quality).
• Policy and firm implications:
  • Firms should weigh low-code AI adoption against necessary complementary investments (data infrastructure, integration, training, governance).
  • Policymakers and managers should anticipate re-skilling needs and design monitoring/guidance to avoid over-reliance on unconstrained LLM outputs.
• Directions for future economic research:
  • Multi-site studies to assess generalizability and sectoral differences.
  • Longitudinal studies to measure persistent productivity effects, task reallocation, and wages.
  • Cost–benefit analyses that incorporate data-integration and governance expenses, not only model/runtime costs.

Summary takeaway: Low-code conversational AI produces measurable, economically meaningful productivity improvements for bounded support tasks when integrated with organizational data and human validation — but the net economic value crucially depends on data access, system integration, governance, and workforce complementarities.