Hybrid AI in manufacturing: what goes private, what goes managed
A task-by-task guide to routing telemetry, inspection, engineering and market work between private models in your cloud and managed models such as Claude or GPT.
Read articleHybrid AI for Manufacturing and Industrial Operations
Hybrid AI that routes every request. Private models in your own cloud or at the plant edge for process recipes, designs, IoT sensor telemetry and quality data. Leading managed models like Claude or GPT for market research, supplier news and public regulation. Answers that reason like your engineers and quality teams, at a cost you can forecast.
Built for
Manufacturing AI rarely fails for lack of good models. Control breaks down for five reasons, and most plants, suppliers and OEMs recognise at least one of them.
Reason 1Shadow AI
Drawings, process recipes and failure logs get pasted into public chatbots because the approved tool is too weak or too slow.
The cost: A process-IP and export-control risk you can’t see or audit.
Reason 2Stalled Pilots
IT and OT security won’t let sensor data, designs or quality records leave the network, so AI stays on brochures and public documents.
The cost: You pay for AI that never touches the machines that matter.
Reason 3Generic Output
General-purpose models ignore your equipment history, work instructions and FMEAs, so engineers second-guess what comes back.
The cost: The time AI was meant to save disappears in the double-checking.
Reason 4Runaway Cost
Machines stream data around the clock and cameras inspect every part. Per-token pricing turns every alarm, image and log line into a cost finance can’t forecast.
The cost: The more lines you connect, the less you can afford to.
Reason 5Audit Scrutiny
OEM customers, certification auditors and regulators now ask how AI influenced a quality decision, which data it saw and who approved it.
The cost: Untraceable AI puts certifications, customer approvals and deliveries at risk.
Aktuara keeps process, design and machine data private, uses Claude or GPT only where it helps, reasons like your engineers, and turns runaway token spend into a cost you can plan.
See how hybrid routing worksContract manufacturer? The same five problems show up with a twist: every customer’s designs must stay separate, and customers audit how you use AI. See the contract manufacturing page.
Hybrid AI is the core of Aktuara. Every request passes a policy check before any model sees it. Process, design, IoT telemetry and quality data runs on LLMs on GPUs in your own cloud or at the plant edge, with no per-token fees. Work with no confidential data, such as market research, supplier news or tracking public regulation, can use Claude or GPT, billed per token for that share only. When in doubt, it stays private.
A RequestFrom an engineer, workflow or machine
Policy CheckConfidential data?
Private Where It MattersLLMs on GPUs in your cloud or at the edge
Managed Where It HelpsPer token, public work only
One AnswerLower cost per answer
Process recipes, engineering designs, sensor telemetry and quality records are handled only by open-source models you control. When in doubt, a request stays private.
Managed models such as Claude, GPT or Grok read market news, supplier filings and public regulation without ever seeing your data. You pay per token only for this share of the work.
Routine work such as alarm triage and work-instruction lookup is handled by small, fast models. Large models are reserved for root-cause analysis and complex drafting.
| Task | Route | What the Model Sees |
|---|---|---|
| IoT sensor streams and edge gateways | Private AI | Live data from connected machines, PLCs and IoT gateways, processed in your cloud or at the plant edge |
| Camera vision and video analytics | Private AI | Line, cell and site camera video, processed on edge GPUs; never sent to a managed model |
| Predictive maintenance alerts | Private AI | Sensor telemetry, equipment history and maintenance logs |
| Defect classification from line images | Private AI | Plant-floor images, inspection results and your defect catalogue |
| Engineering change request summaries | Private AI | Drawings, BOMs, routings and change history |
| Process and recipe analysis | Private AI | Recipes, setpoints, yield and SPC data |
| SOP and work instruction Q&A | Private AI | Your SOPs, work instructions, FMEAs and safety procedures |
| Supplier quality and 8D reports | Private AI | Non-conformances, audit findings and supplier contracts |
| Market and supplier research | Managed AI | Public company news, filings and trade press, never your volumes or prices |
| Public regulation and standards tracking | Managed AI | Public regulatory texts such as the EU Machinery Regulation and AI Act |
| Non-proprietary documentation | Managed AI | Public datasheets, generic product descriptions and public marketing copy |
See how Aktuara answers: an engineer asks, and every answer separates observation from interpretation, states its confidence, cites your own SOP and shows which side of the hybrid route handled it.
Vibration and temperature trend from a press line
Private AIConfidential data: stays in your cloud
| Signal | Reading |
|---|---|
| Vibration | 7.8 mm/s (baseline 2.1) |
| Temperature | +11 °C over 6 days |
| Load | 92% of rated |
| Last service | 412 days ago |
Is this bearing heading for failure, and when should we act?
Image from the end-of-line camera at weld station W3
Private AIConfidential data: stays in your cloud
| Field | Value |
|---|---|
| Station | Weld seam W3 |
| Part | Bracket BR-220 |
| Detected | 3 surface pores |
| Same shift | 4 similar flags |
Classify this defect against our defect catalogue.
Engineering change request with drawings and BOM changes
Private AIConfidential data: stays in your cloud
Summarise this change request and everything it affects.
Yield drop on an injection moulding line
Private AIConfidential data: stays in your cloud
| Measure | Value |
|---|---|
| Yield | 91.2% (target 97%) |
| Main scrap | Short shots, 64% |
| Melt temperature | +6 °C drift |
| Material | New resin lot |
Why did yield drop, and what should we check first?
Operator question at a stopped robot cell
Private AIConfidential data: stays in your cloud
What does our procedure say for this fault?
Public news and filings on your key suppliers
Managed AIPublic sources only: managed model allowed
What has changed for our key suppliers this quarter?
Illustrative examples with synthetic data. These are not real machines, parts, companies or model output; they show the format of Aktuara’s answers. Document names and sections are examples of your own SOPs.
Special Focus: Camera Vision
Most plants already have cameras on lines, robot cells and loading bays, and almost nobody watches the footage. Vision-language models now turn that video into answers you can search, count and act on. It is where AI adds the most value on the plant floor today. Because footage shows your process and your people, Aktuara runs it only on private models at your plant edge.
Illustrative example with synthetic data. The engineer reviews each clip before anything changes.
Video is processed on GPU servers on site or in your own cloud. It is never sent to a managed model, and raw footage never leaves your infrastructure.
No facial recognition and no scoring of individual workers. Faces can be blurred at the source, and retention follows your own policy.
Detections are suggestions for your inspectors, engineers and safety team. Designed to support GDPR and EU AI Act obligations, with documentation for your DPIA and works council.
General-purpose assistants give every company the same answers. As more engineering, maintenance and quality work passes through AI, the process know-how that sets your plants apart becomes available to anyone with a subscription.
| Aspect | Generic AI | Aktuara |
|---|---|---|
| Learns from | Public data, the same for every customer | Your IoT sensor data, equipment history, recipes, FMEAs and work instructions |
| Reasons like | The vendor’s house style | Your engineers, maintenance planners and quality teams |
| Judged by | The vendor’s benchmarks | Your experts, on real machines, defects and changes |
| Improves | When the vendor updates everyone | Every time your engineers review an answer |
| Uses | One vendor’s model for everything | Private models for confidential data, managed models where they help |
| Runs on | The vendor’s servers | Your own cloud account or plant-edge servers |
Your Competitor Can License the Same Model Tomorrow. They Can’t License Your Process, Your Machine History or Your Engineers’ Judgement.
Your process, your AI.
Its knowledge, reasoning and quality standard come from your engineers. Nothing you build trains anyone else’s model.
Confidential by design.
Every model that sees process, design or machine data runs in your cloud account or at your plant edge. We operate the platform without any access to your data.
Ready for audits and customers.
Experts rate answers on real machines and parts, every answer cites its source, and every change is versioned and tested. A change goes live only when it meets your standard.
The strongest model for each task.
Private models for confidential data, Claude or GPT for public research. Agentic RAG pulls in the exact work instruction or maintenance record. New models are adopted only after testing on your data.
Costs less as lines are connected.
Small models first, capacity that follows your shifts and batch runs, and no per-token fees on confidential work. Managed AI is paid for only where it earns its place.
A complete hybrid AI stack, operated for you. Each capability runs where it belongs: private for confidential data, managed where public sources help.
Several large models on call, each chosen per task, so no single model or vendor limits your work.
Private + ManagedFast, efficient SLMs handle routine work first, such as alarm triage and work-instruction lookup, keeping each answer quick and low-cost.
PrivateModels tuned on your work instructions, maintenance logs and approved reports, so answers use your plant’s own language.
PrivateSupervised agents gather drawings, BOMs, maintenance records and inspection data into one working draft, such as a change summary or 8D report, with an expert approving each outcome.
Private + ManagedAnomaly detection over vibration, temperature, pressure, energy and PLC data, connected read-only through OPC UA, MQTT, your historian or your IoT platform.
PrivateSupplier news, market prices, trade press and public regulation such as the EU Machinery Regulation, read from public sources and kept current.
ManagedDefect classification from line cameras and phone photos, trained on your own defect catalogue and checked against your inspectors’ decisions.
PrivateDrawings, work instructions, maintenance logs, scanned checklists and supplier certificates read accurately, with low-quality pages flagged for review.
PrivateModels run on GPU servers at the plant, close to lines and robot cells, when latency, connectivity or OT security require it.
PrivateGovernance built in from day one, the way controlled changes on your plant floor already run.
PrivatePrivate runs in your cloudManaged public sources onlyPrivate + Managed routed per request
Every AI workload has a break-even point. Below it, paying per token is cheaper; above it, owned capacity wins. Manufacturing data never sleeps: sensors stream around the clock and cameras inspect every part, so much of it crosses that line early. Aktuara puts each workload on the right side of the line, and moves it when it crosses.
Illustrative. Monthly cost (vertical) against monthly usage (horizontal); dots mark each break-even point.
At very low volumes, pay-per-use can be cheaper. Our break-even calculator shows where your organisation sits. Cost is one reason for hybrid AI, not the only one: process trade secrets and export-controlled designs often justify private AI before the numbers do.
Building hybrid AI for your plants yourself means hiring a team, building a platform and passing IT, OT security and quality reviews before a single engineer benefits. With Aktuara, you skip straight to the part that matters: your engineers using it.
Your private environment is running in hours, while a build-it-yourself project typically spends its first months on hiring and infrastructure. A supervised pilot with your experts follows within weeks.
No AI engineers, MLOps specialists or GPU experts to recruit in one of the tightest talent markets. Your IT and OT teams stay focused on the systems they already run.
Upgrades, monitoring, scaling, security patches and model releases are handled around the clock. No infrastructure for your team to babysit.
Illustrative comparison of typical phases. Your timeline depends on scope and your own review processes.
XePlatform is our production AI operations platform, already running private and managed AI side by side inside customers’ own cloud accounts. Underneath is platform engineering on Kubernetes: a wide ecosystem of interconnected open-source tools for model serving, scaling, observability and security, integrated, tested and operated as one platform. Aktuara is the manufacturing and industrial operations product built on it, so the engineering underneath is proven before your first pilot begins.
Most plants start with one of these workflows. Every answer shows where it came from and which route handled it.
Sensor trends compared with your own equipment history, so planners get an early, explained warning and fix the machine in a planned stop.
Images from line cameras and phones classified against your defect catalogue, so inspectors confirm instead of starting from scratch.
Change requests summarised with every affected drawing, BOM line, routing and open order, ready for the change board to decide.
Operators and technicians ask in plain language and get the exact step from your approved procedure, with the section cited.
Aktuara supports engineering, maintenance and quality work. It does not control machines or make decisions: a qualified person reviews every output, and nothing changes on a machine, line or robot automatically.
Explore workflows, example outputs and team scenariosYour teams work with Aktuara inside your own cloud account or at your plant edge. We run the platform from outside, with no access to what is inside.
Process, design, telemetry and quality data: never, including prompts, outputs, documents and logs. Operational health metrics: only these, which you can inspect or switch off. Managed AI: only the public sources and public-only tasks you route to it, under a monthly cap you set. See the full data-flow breakdown.
Start with one workflow, often one line or one maintenance team. Your IT, OT security and quality teams review in parallel, with documentation we supply.
Aktuara is installed in your cloud account, data centre or plant-edge servers, in the region you choose, in Europe or North America.
SOPs, work instructions, maintenance history and quality records are indexed, each source traceable. Machine data is connected read-only. You choose which tasks may use managed AI.
Engineers and technicians use Aktuara on real alarms, defects and change requests, in parallel with today’s process, and rate every answer.
The configuration your engineers and quality team approved goes live. Rolling back takes one step.
Manufacturers rightly ask for proof. Here is what you can inspect before you decide, and what your pilot hands you at the end.
No per-token fees for confidential work, however many sensors, images and documents you process. Estimate your break-even point with our calculator.
Before the pilot starts, we record how your team works today. Then we measure the difference.
About hybrid, private and managed AI for manufacturing and industrial operations.
Hybrid AI uses different models for different tasks. In Aktuara, anything involving process recipes, engineering designs, sensor telemetry or quality data stays on private, open-source models in your own cloud or at your plant edge, with small models answering first and larger models when needed. Tasks with no confidential data, such as market research, supplier news or tracking public regulation, can use managed models such as Claude, GPT or Grok.
Private-only means your teams miss out on the most capable models for public research. Managed-only sends process and design data to a vendor and bills every alarm, image and document by the token. Hybrid gives each task the right model: confidentiality where it matters, frontier capability where it helps, and the lower cost for each workload.
No. Every request passes a policy check before any model sees it. Anything involving process, design, telemetry, quality or customer data, or where the check is unsure, stays private. You decide which task types may use managed AI at all, and you can switch it off entirely.
No. Aktuara works on top of the systems you already run. Your MES, ERP, historian, IoT platform, maintenance system and PLM stay the systems of record; Aktuara reads from them and helps your teams analyse, summarise and draft, grounded in your own SOPs.
No. Aktuara is decision support with human oversight. Machine data is connected read-only, nothing is written to PLCs, robots or lines, and a qualified person reviews every output and makes the decision. Your organisation defines the intended use of each workflow.
Yes, read-only. Aktuara connects to machine and sensor data through standard interfaces such as OPC UA and MQTT, your historian or your existing IoT platform, and processes it in your cloud or on GPU servers at the plant edge. Nothing is ever written back to PLCs, robots or devices, and IoT data never reaches a managed model.
No. Aktuara’s camera vision detects events such as defects, blocked aisles, robot cell stops and zone entries, not identities. There is no facial recognition and no scoring of individual workers; faces can be blurred at the source, and video is processed only on private models in your own infrastructure. We provide documentation for your data-protection impact assessment and works council.
In Aktuara, sovereign or private AI is the part that handles confidential data. The models run inside your own cloud account, data centre or plant-edge servers, under your control, instead of on a vendor’s servers. Aktuara delivers it as an operated platform, so manufacturers get the benefits without building an AI team.
Running AI on confidential plant data is far more than provisioning a server with a GPU. It needs infrastructure provisioning as code, model serving, retrieval over your own SOPs, connectors to plant systems, staging environments, release engineering with evaluation gates and rollback, observability, autoscaling, backup and security policy, all kept running every day. Aktuara delivers that complete stack, built on XePlatform, inside your own infrastructure and operated for you, so a supervised pilot can start within weeks without new hires.
Aktuara is designed to support quality management expectations such as ISO 9001 and IATF 16949, industrial security practices such as IEC 62443, and obligations under NIS2, the EU AI Act, the EU Data Act, GDPR and export-control rules for technical data. Confidential data is processed only inside your own infrastructure, every request and approval is logged, and we provide documentation for your risk assessments and audits. Compliance itself depends on how your organisation uses the platform.
Open-source general, vision and time-series models run inside your infrastructure for all confidential work, alongside your own engineering models. New models are adopted only after they pass evaluation on your own data. Managed models such as Claude, GPT or Grok are used only for tasks with no confidential data.
Yes. Aktuara runs on the major public clouds in the region you choose, in Europe or North America, in your private cloud or data centre, or on GPU servers at the plant edge, built on open standards so you can move without rebuilding.
A fixed platform subscription plus your own infrastructure costs, with no per-token fees for confidential work. Managed AI is billed per token only for the tasks you route to it, under a monthly cap. Small models first, capacity that follows your shifts and batch runs, and shared hardware keep the cost of each answer low, which matters when sensor and image volumes run around the clock.
The environment is set up in hours and your SOPs loaded within days. A supervised pilot with your engineers typically runs about four weeks, while your IT, OT security and quality teams complete their review in parallel.
A task-by-task guide to routing telemetry, inspection, engineering and market work between private models in your cloud and managed models such as Claude or GPT.
Read articleMachines stream data around the clock and cameras inspect every part. Here is where the break-even point sits, and why owned capacity wins.
Read articleIf every manufacturer uses the same assistant, the process knowledge and engineering judgement that set a strong plant apart start to disappear.
Read articleExplore a supervised pilot using your own machine data, documents, infrastructure and evaluation criteria. Tell us the workflow that costs your engineers the most time.