ThingWorx + private AI
AI for PTC ThingWorx: Natural-Language Access to Things, Value Streams and ERP Context
Short answer
AI for PTC ThingWorx exposes Things, properties, services and value stream history to a private language model through the ThingWorx REST API, then joins that data to ERP production and maintenance records. Engineers ask in plain English, see the calls made, and the model can run on-prem alongside a self-hosted ThingWorx deployment.
- ERP
- PTC ThingWorx
- Industries
- Manufacturing, Automotive, Heavy Equipment, Aerospace
- Written for
- IIoT Architect
ThingWorx organises industrial data as Things built from Thing Templates and Thing Shapes, each with properties, services and events. Property changes are logged to value streams or streams, and Mashups sit on top for operators. Many plants have a rich model, years of history and a few dashboards, yet the people with questions rarely know which Thing holds the answer.
Because ThingWorx can be deployed on-prem, in a private cloud or as managed service, it is common in plants with strict data controls. That makes a public LLM unattractive but leaves a real need: operators and engineers want to ask what happened on a line without learning the model, writing a service, or building another Mashup.
The ERP holds the other half of the picture: orders, routings, materials, maintenance plans. ThingWorx can integrate with ERP through connectors and custom services, but the integration is usually point-to-point and built for one dashboard rather than for open questions.
This page covers how a private AI layer uses ThingWorx's own REST surface safely, how it stays read-only, and where ThingWorx's built-in analytics and Mashup tooling is the better tool.
What usually gets in the way
The problems we hear most from iiot architect teams running PTC ThingWorx.
Hard to know where the answer lives
Large ThingWorx models have hundreds of Thing Templates and thousands of Things. Newcomers cannot tell which Thing, property or stream answers a question.
Every new view needs a Mashup
Operator questions that do not fit an existing Mashup wait on a developer. Composer work for a one-off question is poor value.
ERP integration is dashboard-shaped
Connections to the ERP typically feed a specific screen, not open exploration across machine history and production orders.
Value stream queries are easy to get wrong
Historical queries need correct time windows, property names, and paging. Large unbounded queries load the platform and its persistence layer.
Security posture limits cloud AI
ThingWorx installs in regulated plants often cannot use SaaS AI services, so any assistant must be private.
Where AI earns its place in PTC ThingWorx
Each use case names the ERP objects it reads or writes, so your ERP team can judge the integration effort before anyone commits budget.
Ask about a line in plain English
Operator asks what stopped Line 3 last night. The agent finds relevant Things, queries property history and returns a timeline.
Touches: ThingWorx Things, properties, value streams; ERP work centers and production orders
Outcome: Answers in a minute or two instead of waiting for a developer to build a view.
Downtime reasons joined to orders
Match stop events to the production order running, the part, and the shift, then summarise the top causes.
Touches: Production orders, operation confirmations, shift calendars; ThingWorx event history
Outcome: Pareto of downtime by product and cause that previously required manual alignment.
Thing model discovery assistant
Search the model by meaning: which Things expose spindle temperature, and what units do they use.
Touches: Thing Templates, Thing Shapes, property definitions; equipment master
Outcome: Faster onboarding and fewer duplicate Mashups; descriptions flagged for owner review.
Maintenance triage with live context
For a condition alert, pull recent property trends, open maintenance orders and spare stock, then draft a note.
Touches: Maintenance orders, equipment, spare parts; ThingWorx alerts
Outcome: Technicians start with context; order creation still requires approval.
Genealogy question support
Combine machine parameter history with lot and serial data to show conditions a batch was produced under.
Touches: Lot and serial records, production orders; parameter streams
Outcome: Faster containment investigations; evidence packaged for quality review.
Service drafting for administrators
Draft ThingWorx service JavaScript or query snippets for an administrator to review and test in a dev instance.
Touches: Thing services, DataTables, Streams
Outcome: Less boilerplate; code is never deployed without review and test.
Throughput versus plan summaries
Compare counted output properties to planned quantity from the ERP and explain the gap.
Touches: Planned orders, routing standards; production counters
Outcome: A daily narrative for supervisors that references both sides of the data.
Reference architecture
An orchestration service calls ThingWorx REST endpoints with an application key restricted to read, and queries the ERP read-only. A private model plans bounded calls, and the UI shows each call.
- 1
ThingWorx connector
Uses REST endpoints for Things, properties and services with a dedicated app key and a narrowly permissioned user; value stream queries are always time-bounded.
- 2
Model index
A searchable catalogue of Thing Templates, Shapes and property metadata so the model finds the right Thing before querying.
- 3
ERP context
Read-only ERP access for orders, routings, maintenance and lots, aligned to Things by equipment or work center keys.
- 4
Private inference
Open-weight models on vLLM or Ollama in the customer's data center or private cloud.
- 5
Guardrails
Allow-listed services, read-only mode, logging of every call, and human approval for any state-changing action.
Integration notes for your ERP team
- Use the ThingWorx REST API with an application key bound to a user that only has read permissions on required entities.
- Allow-list the services the agent may call; reject arbitrary service execution.
- Always pass explicit start and end times and a row limit when querying value streams or streams.
- Index Thing Templates and Shapes first so the model resolves names without scanning all Things.
- Join to the ERP through an equipment or work center property stored on the Thing rather than name matching.
- Test against a non-production instance, since heavy history queries can load the persistence provider.
Deployment options
Alongside on-prem ThingWorx
Plants hosting ThingWorx themselves
Model server in the same network zone; no external calls; the simplest compliance story.
Private cloud tenancy
Teams running ThingWorx in their own cloud account
Orchestrator and GPUs in the same VPC, with private endpoints between components.
Air-gapped plant network
Defense and high-security sites
Model weights loaded offline, no internet egress, ERP replica or direct read-only link inside the enclave.
Compliance and data control
How the architecture supports your obligations. Certification and accountability stay with your organisation; the design keeps the evidence straightforward.
IEC 62443
Zone and conduit design keeps the AI tier read-only toward OT systems.
NIST SP 800-82
Segmentation and monitoring consistent with ICS security guidance for the added service.
ITAR / CMMC (customer obligation)
On-prem design lets the customer keep controlled data inside their boundary and scope the system within their own assessment.
ISO 9001 / AS9100 records
Parameter evidence referenced in quality records is retrieved read-only and logged.
How an engagement runs
Phase 1 . 1-2 weeks
Model review
- -Thing and template inventory
- -Priority question list
- -ERP key alignment
- -Security zone design
Phase 2 . 2-4 weeks
Connector build
- -Read-only app key and permissions
- -Model index
- -Bounded query tools
- -Test instance validation
Phase 3 . 3-5 weeks
Pilot
- -Operator and engineer trials
- -Evaluation questions
- -Call transparency UI
- -Audit logging
Phase 4 . 2-4 weeks
Rollout
- -Additional lines and Things
- -Monitoring and runbook
- -GPU sizing review
- -Governance for new services
Questions to ask any vendor, including us
A short list that separates real PTC ThingWorx AI work from a chatbot demo.
- Which ThingWorx permissions does the agent use, and are they read-only?
- How are value stream queries bounded to protect the platform?
- How does it find the right Thing in a large model?
- Can it run in our network with no internet egress?
- How is the ERP joined to Things reliably?
- Who approves anything that changes state?
- How are ThingWorx upgrades handled?
- When is a Mashup or ThingWorx Analytics the better tool?
Frequently asked questions
Can an LLM query ThingWorx directly?
Yes, through the REST API, but not with free-form access. A safe design exposes a small set of allow-listed tools, such as list Things by template or fetch property history for a bounded window, and the model chooses among them. Credentials are held by the orchestrator, not the model, and every call is logged.
Does this work with on-prem ThingWorx?
That is the most natural fit. The model server sits in the same network zone as ThingWorx and the ERP, so no data crosses the internet. You need GPU capacity sized for your concurrency, which depends on model choice and the number of simultaneous users.
How is this different from ThingWorx Analytics?
ThingWorx Analytics focuses on predictive modelling over your data. A natural-language layer is about exploration and question answering across Things and ERP data. They are complementary, and if your need is a specific prediction, the native analytics tooling may be the simpler path.
Can it generate ThingWorx services?
It can draft service JavaScript or query logic for an administrator, which is useful for boilerplate. Output should be reviewed and tested in a development instance, never deployed automatically. Generated code is treated as a suggestion with the same change control as hand-written code.
Will heavy queries slow the platform?
They can if unbounded, which is why every tool enforces time windows and row limits and the pilot runs against a test instance. Caching summaries and, where justified, replicating selected history to a separate store reduces load on the production persistence provider.
What about write access?
Read-only by default. Anything that would change a property, trigger a service with side effects, or create an ERP order is drafted for a human to approve. Control actions toward machines are outside the scope of a language-model assistant.
Is it worth it for a small deployment?
If you have a handful of Things and a few Mashups that cover the questions asked, probably not. The value appears when the model is large, the questions vary, and engineers spend real time hunting for the right Thing or joining machine data to ERP records.
Related guides
AI Agents for ERP, Running On-Prem
A practical guide to on-prem AI agents for ERP: what they can safely automate, where human approval belongs, and how to design the guardrails.
RAG + SQL + permissionsA Private LLM Grounded on Your ERP Data
How a private LLM answers questions on your ERP data: RAG plus text-to-SQL, role-based permissions inherited from the ERP, and where each fits.
Faster paperwork, same standardAI for NCR, CAPA, and 8D Drafting in AS9100 and ISO 9001 Quality Systems
AI drafts NCR, CAPA, and 8D reports from ERP quality data for engineer review, cutting drafting time under AS9100 and ISO 9001 without changing sign-off.
SAP PM/EAM + on-prem AIAI for SAP Plant Maintenance: Notifications, Work Orders, and Predictive Signals
AI grounded on SAP PM data that triages notifications, drafts work orders, standardizes failure codes, and flags predictive maintenance signals, on-prem.
On-prem AI, any ERP, A&DOn-Prem AI for ERP in Aerospace, Defense, and Electronics Manufacturing
A hub guide to on-prem AI across SAP, Infor LN, Costpoint, IFS, and Oracle EBS for aerospace, defense, and electronics manufacturers under ITAR, CMMC, and AS9100.
Opcenter Execution + private AIAI for Siemens Opcenter Execution MES, Running Inside Your Own Network
Add private, grounded AI to Siemens Opcenter Execution MES: ask questions over lots, work orders, nonconformances and genealogy without leaving your network.
Plan it with numbers
OEE Calculator (Overall Equipment Effectiveness)
Break your Overall Equipment Effectiveness into availability, performance, and quality, and see how far you sit from the 85% world-class benchmark.
Free ToolManufacturing Downtime Cost Calculator
Combine lost contribution margin, idle labor, and absorbed overhead into a defensible monthly and annual downtime cost - with recovery effects modeled.
Free ToolOn-Prem AI Security Hardening Checklist
A practical control checklist for securing self-hosted language models, covering model provenance, network isolation, data governance, host hardening, and audit readiness.
GuideIoT Sensor to ERP Data Integration for Manufacturing
Build reliable IoT sensor to ERP data pipelines for manufacturing. Covers MQTT, edge processing, data normalization, and ERP API ingestion patterns.
GuideEdge Computing for Manufacturing ERP Integration
Deploy edge computing for manufacturing ERP integration. Covers edge gateway selection, local processing, latency reduction, and offline-capable ERP operations.
GuideAir-Gapped LLM Deployment Patterns That Actually Work
Air-gapped LLM deployment patterns that work: offline model transfer, update workflows, monitoring without telemetry, and CMMC-ready architectures.
Talk it through with an engineer who knows PTC ThingWorx
Bring one real question your team cannot answer from the ERP today. We will map the data path, the model, and where it runs, and tell you honestly if AI is the wrong tool for it.