Vision, method, and evidence
The factory as a learning system.
Agents connect orders, knowledge, and processes on your existing systems. They coordinate tools and complete authorized work.
Research date and latest source review: September 30, 2026
Where agents take manual work off the shop floor
- Access sourcesMake required data available.
- Build toolsDevelop connectors and deterministic checks.
- Verify resultsCheck independently and connect your systems.
- Operate workflowAssign exception responsibility.
Specialist agents and deterministic tools connect data, interfaces, and workflows. Independent verification and assigned exception handling are part of operation.
In development
Drawing data without retyping
Dimensions, tolerances, and materials are retyped by hand from drawings into routings today. Agents extract them with an evidence crop; deterministic checks flag deviations for approval.
Trialled on 11 drawing sheets, not yet in use at customers; this does not establish general accuracy.
In development
Replan supplier date changes faster
When a supplier moves a date, planning recalculates by hand today. The planning preview shows stock and capacity over six months; the planner confirms every change.
Demonstrable state with invented sample data, not yet in use at customers.
Published customer case
A documented customer entry point
At a CNC contract manufacturer, order entry fell from 20 to 2 hours weekly. The complete workflow includes ERP handoff.
The result applies to the published workflow; savings elsewhere require separate evidence.
Industrial adoption through 2031
- Through September 2027
Complete work across systems
- From isolated assistance to agents completing cross-system work.
- Entry points: order intake, supplier exceptions, and quality documents.
- Evidence
- 84 percent of respondents report value [E1].
- About 20 percent of cases have scaled [E1].
- Fraunhofer: tool orchestration [E2]; PackSys: incremental integration [E4].
- Confidence
- Medium-high direction; lower confidence in Swiss-SME adoption speed.
- Requires data access, responsibility, and permissions.
- Counter-signal
- Faster extraction; exceptions consume savings.
- Revise if manual touches/outstanding exceptions do not decrease across the workflow.
- October 2027 through September 2029
Shared context and cooperating agents
- Shared context connects orders, planning, quality, and service.
- People direct goals/exceptions; tools and solvers compute and act.
- Evidence
- Manufacturing-X, NIST, and ERP-agent offerings support this direction [E5, E6, E8].
- Confidence
- Medium; requires common meanings, permissions, and cross-functional responsibility.
- Selecting a particular model cannot replace these requirements.
- Counter-signal
- Each new application rebuilds data foundations.
- Reassess architecture/economics when second/third workflows reuse little context or tooling.
- October 2029 through September 2031
Validated adaptation in operation
- More factories learn from outcomes and simulate alternatives.
- Adapt workflows/selected physical operations within agreed limits.
- Evidence
- NIST: twin validation/uncertainty; Siemens: adaptive factories [E5, E7].
- Announcements cannot establish universal operation.
- Confidence
- Medium direction; low confidence in broad dependable physical autonomy.
- Validate the process model for its specific decision.
- Counter-signal
- Simulation and real outcomes diverge.
- Stop mechanisms breaching quality/safety constraints; retain the connected-factory target picture.
These horizons describe adoption, organizational change, and breadth of operation. Suitable manufacturers can implement elements from any horizon today.
What the connected factory can achieve
An order ready to manufacture
IntakeDrawingsMaster dataWork instructions
Start: Verify ERP handoff, then trace requirements through to the operation.
An explainable delivery decision
MaterialSuppliersCapacityScheduling
Start: Connect a changed date to affected orders and feasible options.
Quality and expertise where needed
MeasurementsDocumentationExperience
Start: Cite certificate/answer sources; use corrections in the next workflow.
Better decisions from outcomes
EventsRulesProcess models
Start: Test scenarios; compare approved improvements against observed outcomes.
More can be built today than the catalog shows. Missing packaging requires development; assess technical feasibility separately.
| Capability | Outcome and entry point | Connection / responsibility |
|---|---|---|
| Order intake and documents | Capture the purchase order as a checked ERP result. | Master data and exceptions; order team approves deviations. |
| Work instructions and MES | Current instruction reaches the correct operation. | Drawing, routing, revision; manufacturing engineering reviews. |
| Drawings and tolerances | Identify conflicts before manufacturing or quoting. | Evidence/dossier checks; engineering accepts critical facts. |
| Master data and standards | Consistent identities and units for the chosen workflow. | Agents propose mappings; accountable experts confirm. |
| Quoting and costing | Reuse comparable jobs and supported costs. | Routings/cost rates; estimators approve prices. |
| Shop floor and exceptions | Resolve work arising from deviations. | Deterministic metrics/tasks; shift supervisors respond. |
| Lead times and material | Show affected dates and unknown inputs. | Check ERP durations and the reference baseline first; buyers and planners decide. |
| Capacity and scheduling | Provide feasible date alternatives. | Solver checks constraints; planner approves. |
| Quality and test documents | Assemble documents from real measurements. | Template, lot, test data; quality signs off. |
| Quality analytics and causes | Test hypotheses against measurements. | Traceability/experiments; correlation alone cannot establish causation. |
| Maintenance and assets | Reduce avoidable loss with useful signals. | Failure history and value ceiling; asset owners respond. |
| Knowledge and expertise | Find cited answers and comparable cases. | Sources, permissions, corrections; experts maintain accepted knowledge. |
| Procurement, invoices, and service | Complete routine work through authorized transactions. | Installed ERP first; approve transaction write rights separately. |
| Integration and orchestration | Connect systems, tools, and agents traceably. | Events, permissions, evaluation; process/service responsibility. |
| Sovereign operation | Select suitable data and model operation. | Assess data location, inference, support, license, workload TCO separately. |
| Digital twins and adaptation | Simulate alternatives; improve using outcomes. | Distinguish dashboards, process models, validated simulation, bounded control. |
Availability and evidence
Documented status; each customer installation needs its own verification. Agree on the first workflow based on your operation and interfaces.
- Order entry, work instructions, and shop floor
- in use: Published examples/documented workflows; customer-specific integration.
- Planning, lead time, orchestrator, and knowledge
- in development: Further connected workflows; no single finished package yet.
- Root-cause analysis and service
- planned: Define outcome and initial implementation per workflow.
- General operating system for an adaptive factory
- not evidenced: Target picture; universal complete deployment remains unproven.
Start with one complete workflow
- Choose outcomeWhat should reliably get done?
- Assign responsibilityDefine goals, authority, and exception handling.
- Build connectionsResolve required sources, data, and standards within the workflow.
- Verify workflowMeasure output, residual work, error costs, and system effects.
- Next connectionReuse proven context, tools, and accepted corrections.
Sources, method, and limits
- Observed facts, vendor assertions, and forecasts are identified separately.
- Global surveys cannot establish representative Swiss-SME figures.
- Deloitte counts respondents and use cases with different denominators [E1].
- KPI potential cannot establish realized returns; Orcha-3 validates virtual chains [E2].
- Widespread production deployment of Orcha-3 remains unproven.
- Examples in development are labelled as such.
- Customer results remain limited to published cases.
- METR: perceived speed/coding experiments cannot establish universal productivity multipliers [E10].
- Data location alone cannot establish local inference or comprehensive compliance.
- Review support, retention, license, updates, and privacy separately.
- AI-Act duties follow risk; industrial AI is not universally high-risk [E9].
- No legal advice or blanket assurance for Swiss companies.
- Boris Scharinger, Industrial AI: Pilot to Profit (2026).
- Chapter 5: PDF pp. 115-126; chapter 9: pp. 140-150.
- Chapter 12: pp. 161-163; chapter 20: pp. 202-226.
- Chapter 21: pp. 227-235; chapter 24: pp. 244-272.
- Part IV: 19 enterprise functions; taxonomy, no factorIQ benchmark.
- Dates identify publication/update; undated pages use a September 30, 2026 retrieval date.
- All sources were last reviewed September 30, 2026.
- Company-wide data cleansing can follow the first selected workflow.
- Goals, everyday responsibility, and real-workflow evidence determine the entry point.
- New manuscripts remain under editorial review; only published articles are linked.
- [E1] Deloitte: AI in Manufacturing 2026 (2026-06-30)
- [E2] Fraunhofer IWU: Orcha-3 and agentic AI (German source) (2026-09-08)
- [E3] Swissmem: Swiss industrial AI requirements (German source) (2024-05-17)
- [E4] Swissmem: PackSys on digital workflows and AI (German source) (2026-04-14)
- [E5] NIST: trustworthy digital twins for manufacturing (2026-07-20)
- [E6] VDMA: Manufacturing-X and sovereign data exchange (2026-09-30)
- [E7] Siemens: industrial AI, copilots and adaptive factories (2026-01-14)
- [E8] SAP: Joule and cross-system agents (2026-09-30)
- [E9] European Commission: risk-based AI Act (2026-09-30)
- [E10] METR: measuring developer productivity with changing AI tools (2026-02-24)
Published further reading
- Getting AI started in manufacturing
- Engineered AI and process automation
- Ask your machines
- Which AI tools pay off for manufacturing SMEs?
- AI agents in order intake
- The language model comes last
These iosys articles explain the methods we use.
Arrange a call →