Compliance and data rules for AI in manufacturing
What are the compliance rules for AI in manufacturing?
Manufacturing AI compliance splits across three zones. On the shop floor, worker personal data under the DPDP Act and network segmentation govern what leaves the plant. In enterprise systems, customer confidentiality on drawings governs what reaches a model. Exported machinery brings the EU AI Act.
Manufacturing AI compliance splits across three zones with different rules. On the shop floor, worker personal data under the DPDP Act 2023 and operational-technology segmentation govern what may leave the plant. In enterprise systems, customer confidentiality clauses on drawings and specifications govern what may reach a model at all. For anything sold into the EU, the AI Act governs safety-related components.
Confusing those three zones is the usual reason a manufacturing AI project stalls in legal review. This article separates them, gives you the architecture rule each one forces, and covers the quality-system obligation that surprises teams putting a model into an inspection decision.
Zone one: the shop floor and operational technology
The plant network is not the office network, and treating it as one is both a security failure and, increasingly, a contractual one. Industrial control systems sit behind segmentation designed so that a compromise in enterprise IT cannot reach a programmable logic controller. Any AI project that wants sensor, vision or machine data has to cross that boundary deliberately.
The workable pattern is one-directional: data leaves the control zone through a broker or historian into a read-only store in the enterprise zone, and nothing in the AI system holds a write path back to a controller. When a recommendation needs to reach the line, it reaches a human or a supervisory system, not the controller.
Worker data is personal data
Biometric attendance, CCTV of work areas, wearable safety monitors and productivity tracking all produce personal data about identifiable workers, and the DPDP Act 2023 applies to it. The employment-related grounds in the Act are narrower than teams assume, so treat notice, stated purpose and retention periods as mandatory rather than optional. Our walkthrough of DPDP Act 2023 and AI covers the obligations in detail.
Where a model may recommend and where it may not act
A model that predicts a bearing failure and raises a work order is an advisory system. A model that interlocks a machine or overrides a safety function is a safety-related control function, and that is governed by functional safety engineering rather than by data protection law. The line matters because crossing it changes the discipline, the evidence and the sign-off. Keep AI advisory until you have deliberately engaged a functional safety process.
Zone two: enterprise systems, drawings and trade secrets
The most common compliance failure we see in manufacturing AI has nothing to do with statute. It is a contract breach. Customer drawings, bills of material, process recipes and tolerances are usually covered by a confidentiality agreement that restricts disclosure to third parties, and a managed model endpoint is a third party.
Before any document intelligence or knowledge assistant touches customer technical data, someone needs to read the master agreements and answer three questions: may this data be processed by a named subprocessor, must it stay in a stated region, and is there an explicit prohibition on use for model training. Those answers determine whether you can use a hosted endpoint with a no-training term or whether you need self-hosted agentic AI inside your own perimeter, which starts at $31,500 or ₹20.8 lakh plus infrastructure.
Alongside that sit statutory record duties that are easy to forget because they are routine: GST e-invoicing and e-way bill records attached to despatch are legal documents, and an AI system may draft or reconcile them but should not silently amend a filed record. The general argument for keeping sensitive workloads within national or corporate boundaries is set out in Sovereign AI in India.
Zone three: the product, the customer and the EU AI Act
If you export machinery or components into the European Union, the EU AI Act reaches you through the product rather than through your data. The Act classifies AI systems intended to be used as safety components of products covered by Union harmonisation legislation, machinery among them, as high risk, with obligations covering risk management, data governance, technical documentation, logging, human oversight and accuracy. The consolidated text and its annexes are published at artificialintelligenceact.eu.
That matters even for a plant in Pune or Coimbatore, because the obligation attaches to the product placed on the EU market. An AI-based vision system you embed in a machine you export is in a different regulatory category from the same model used internally for your own quality control.
The rules mapped to architecture decisions
Compliance in manufacturing is mostly settled by four or five architecture decisions taken in the first fortnight. This is the mapping we work from.
| Zone | Data | Rule that applies | Architecture decision it forces |
|---|---|---|---|
| Control and shop floor | PLC tags, machine telemetry, vision frames | Network segmentation; no write path from IT to OT | One-way export through a broker or historian; advisory output only |
| Workforce | Biometric attendance, CCTV, wearables, productivity data | DPDP Act 2023 notice, purpose limitation, retention | Pseudonymise before analytics; keep raw video out of model context |
| Quality | Inspection images, accept and reject decisions, calibration records | Quality management system traceability and documented information | Version and log every model decision with the image and the operator override |
| Customer technical data | Drawings, BOMs, recipes, tolerances | NDA and master agreement terms; subprocessor and training restrictions | Named vendor with no-training term, or self-hosted inference in your perimeter |
| Despatch and finance | E-invoice, e-way bill, GRN, supplier invoices | GST record accuracy; audit trail | AI drafts, a rule or a person commits; immutable log of both |
| Exported product | Embedded AI in machinery sold into the EU | EU AI Act high-risk obligations for safety components | Risk management file, logging, human oversight designed in from day one |
The quality-system obligation teams forget
If a model participates in an accept or reject decision, it has become part of your quality system, and a quality system requires that the method be documented, that its performance be demonstrated, and that decisions be traceable. An auditor will ask how you validated the model, what happens when it is uncertain, who can override it, and whether you can produce the record for a specific part six months later.
That is why we launch inspection models in shadow mode: the model scores every part while human inspectors continue to decide, and the disagreement log becomes the validation evidence. Computer vision for documents and quality inspection covers what these models do reliably and where they do not. Keeping a human in the loop on reject decisions is the default we recommend for safety-critical parts.
What auditors and customers ask for
Supplier audits and internal reviews converge on a short list. Prepare these once and reuse them.
- A register of AI systems in use, each with owner, purpose, data classes touched and whether any decision is automated
- A vendor and subprocessor list with region, contract terms and whether your data may be used for training
- Validation evidence for any model in a quality or safety-adjacent decision, including the shadow-mode comparison
- A decision log holding input, model version, prompt or weights version, output and any human override
- Worker data notices in the languages your plant actually uses, with retention periods stated
- An OT boundary diagram showing that no AI component holds a write path to control systems
- A rollback plan for a model that starts drifting after a product or supplier change
Frameworks help here rather than adding work. The NIST AI Risk Management Framework is voluntary, widely accepted by industrial customers, and maps cleanly onto the register and validation artefacts above. The log schema we use is described in AI audit trails.
When compliance caution becomes the wrong choice
Air-gapping everything is the manufacturing version of over-correction. Plenty of useful manufacturing AI touches no customer drawing and no worker identity at all: supplier invoice matching, maintenance manual search over your own documents, natural language reporting on production volumes. Running those on a managed endpoint with a no-training clause is proportionate, and insisting on a GPU estate for them buys risk reduction you did not need at a cost you did not have to pay.
The same applies to governance sequencing. A full AI policy, register and risk framework written before a single pilot exists tends to describe systems you never build. Write the register when you have two systems in it, and let the first ProofRun tell you which controls are real.
What this costs to get right
Compliance work is distributed through a manufacturing AI build rather than sitting in one line, and in our experience it adds roughly a fifth to the first project. A three-week ProofRun at $6,250 or ₹4,00,000 is where we usually settle the contract, residency and segmentation questions against one real workflow. Where the platform underneath is the constraint, a Legacy-to-AI Modernization programme starts at $31,500 or ₹22.4 lakh and custom ERP and CRM development at $28,000 or ₹18.4 lakh. Starting prices for every programme are on the pricing page, and the sectors we work in are listed on the industries hub. After launch, the AI add-on to a Care Plan at $750 or ₹40,000 per month keeps evals, cost monitoring and prompt regression running.
Related reading
What AI costs in manufacturing puts numbers against the controls described here, Legacy systems in manufacturing: modernise or replace? covers the platform question underneath, and AI governance for mid-size companies is the proportionate version of a governance programme. If you want the constraints written down before scoping, start on the contact page.
In manufacturing, the compliance question is not whether you may use AI; it is which zone the data came from and whether the model is allowed to act on the answer.
Frequently asked questions
Does the DPDP Act apply to shop-floor worker data?
▾
Yes. Biometric attendance, CCTV of work areas, wearables and productivity tracking all produce personal data about identifiable workers. Notice, a stated purpose and a retention period are required, and the employment-related grounds in the Act are narrower than most teams assume, so pseudonymise before analytics.
Can we send customer drawings to a hosted AI model?
▾
Only if the master agreement allows it. Drawings, bills of material and process recipes are usually covered by confidentiality terms restricting third-party disclosure, and a hosted endpoint is a third party. Check subprocessor, region and no-training clauses first, or run inference inside your own perimeter.
Does the EU AI Act affect an Indian manufacturer?
▾
It can, through the product rather than the data. AI systems used as safety components in machinery placed on the EU market fall into the high-risk category, bringing risk management, documentation, logging and human oversight obligations. Internal-only quality models used in an Indian plant are treated differently.