SEC-01 / AI Governance

AI vendor due diligence for industrial software and OEM equipment

Most AI in a plant arrives through vendors: MES, CMMS, QMS, vision integrators and OEM machines. Due diligence should cover what the vendor does with your data, how models change, where they run, how they are reached remotely, what is inside them, and what the contract lets you enforce.

7 min read Updated CISOsOT engineersPurchasing leadersCIOsPlant managers

Key takeaways

  • Ask in writing whether your data is used to train models for other customers, and get the answer into the contract.
  • Model updates are changes. Require advance notice, release notes and a way to defer or roll back on production lines.
  • Prefer on-premise or edge options for anything touching restricted data or Level 1/2 control, and treat vendor cloud links as new conduits.
  • Ask for an SBOM and, increasingly, an AI bill of materials. Ask how the product development process aligns with IEC 62443-4-1 and the component with 62443-4-2.
  • Scale the questionnaire to the risk tier. Twenty questions for medium tier, the full set for high tier.

Mid-market manufacturers rarely build AI. They buy it, often without meaning to. The MES upgrade comes with an optimization module. The new injection molding machine has "adaptive process control". The vision integrator trained a model on your parts and will "keep improving it remotely". Each of these is a vendor relationship with AI inside, and each deserves a few pointed questions before the purchase order goes out.

This page sets out what to ask, why it matters on a plant floor, and which answers should change your decision or your contract.

Does this apply to you?

Yes if any of these are true:

  • You are buying or upgrading MES, CMMS, QMS, ERP, historian or analytics software with AI features.
  • You are buying machines, robot cells or vision systems that use machine learning.
  • A vendor or integrator wants remote access to retrain or tune a model.
  • A vendor wants your process data, images or documents to "improve the service".

Scale the effort to the risk tier of the use case. For low-tier SaaS features, five data questions may be enough. For an edge AI box writing setpoints to a PLC, use the full set.

1. Data use and training on your data

This is the question that protects your know-how. Ask:

  • Is any customer data (process data, images, documents, prompts, outputs) used to train or improve models? Yours only, or shared across customers?
  • Is training opt-in, opt-out or not configurable?
  • Where is data stored and processed (country, cloud region, sub-processors)?
  • How long is it retained? Can you get it deleted, with confirmation, at contract end?
  • Can the vendor's staff see your data? Under what controls?
  • If the model is fine-tuned on your data, who owns the resulting model weights? Can they be used for other customers?

Plant-floor reality: A vision model trained on thousands of your defect images encodes a lot about your process and your failure modes. A PdM model trained on your compressors is less sensitive. A model tuned on your recipe parameters is very sensitive. Match the contract strength to what the model learns.

2. Model updates and change notification

In a plant, an unannounced change in behavior is a quality and safety issue, not just an IT issue.

  • How often are models updated? Who triggers it (vendor, you, automatic)?
  • Will you get advance notice and release notes describing changes in behavior or performance?
  • Can you defer updates to a maintenance window, test them on one line first, and roll back?
  • How does the vendor validate a new model before release? Can you see the test results relevant to your use case?
  • Does the model learn continuously in production? If so, how is drift bounded?

Ask for a clause requiring advance notice of material model changes and a right to defer for production systems. Feed every accepted update through your MOC process.

3. Where it runs: cloud, on-premise or edge

  • Is there an on-premise or edge deployment option? What features are lost?
  • If cloud: what data leaves the site, at what frequency, through which connection?
  • Can the system keep running safely if the cloud connection drops? What is the fallback?
  • Does inference need internet access, or only updates?

For anything at Purdue Level 1 or 2, or anything using restricted data, prefer edge or on-premise, with any cloud link going through your IT/OT DMZ. Our edge vs cloud AI page covers the trade-offs.

4. Remote access

Vendor remote access is one of the most common routes into OT networks. AI vendors often want more of it, not less, to monitor and retrain models.

  • What remote access is required, by whom, how often?
  • Does it go through your remote access solution, or the vendor's own tunnel or cellular modem?
  • Is it user-initiated and session-based, with MFA and logging?
  • Can you see and record what the vendor does in a session?

Insist that access goes through your controlled path. See secure remote access. An embedded 4G router in a machine cabinet is a new conduit and should be treated as one.

5. SBOM and AI-BOM

  • Can the vendor provide a software bill of materials (SBOM) in a standard format such as SPDX or CycloneDX?
  • Which AI models are used, which versions, and where do they come from (in-house, open source, third-party API)?
  • Does the product call third-party AI services (for example a large language model API)? Which ones, and with what data?
  • How are vulnerabilities in model libraries and runtimes tracked and patched?

An "AI-BOM" is not yet a settled standard. Ask for a plain list of models, versions, sources and update method. That alone tells you a lot about the vendor's maturity. See SBOM in the glossary.

6. IEC 62443 and secure development

For products going into OT, IEC 62443 gives you a common language:

  • IEC 62443-4-1 covers the vendor's secure product development lifecycle.
  • IEC 62443-4-2 covers technical security requirements for components (embedded devices, host devices, network devices, software applications) and security levels.

Ask:

  • Is the vendor's development process certified or assessed against 62443-4-1? By whom, and when?
  • Is the product certified against 62443-4-2? For which component type and security level?
  • If not, how does the vendor align, and what evidence can they share?
  • How does the AI component fit into the zone and conduit model of your system? What ports and protocols does it need?

Certification is not mandatory, and many smaller vendors lack it. Lack of certification is not a deal-breaker for a low-tier use case. For high-tier OT use, expect at least credible alignment and evidence.

7. Regulatory obligations: AI Act and CRA

As of November 2026:

  • EU AI Act. Most vendors of AI systems are "providers" and you are a "deployer". For high-risk systems, providers carry the heaviest obligations (risk management, data governance, technical documentation, conformity assessment). Following the Digital Omnibus on AI (Regulation (EU) 2026/1744), these apply from 2 December 2027 for Annex III systems and 2 August 2028 for AI in products under Annex I legislation. Ask the vendor how they classify the product and what documentation and instructions for use they will provide. Watch for the provider role shifting to you if you or your integrator substantially modify the system or rebrand it (Article 25).
  • Cyber Resilience Act. Applies to products with digital elements placed on the EU market, which includes much industrial software and connected equipment. Manufacturers' vulnerability and incident reporting obligations have applied since 11 September 2026; most other obligations apply from 11 December 2027. Ask how the vendor is preparing, how they will notify you of actively exploited vulnerabilities, and what support period they commit to. See our Cyber Resilience Act page.

Watch out: A vendor statement that the product is "AI Act compliant" means little before the relevant obligations apply and before harmonized standards are settled. Ask what they classify the product as and what evidence they will hand over.

8. Contract clauses

The minimums we would want in any contract for medium or high-tier AI:

Clause What it says
Data use No training on your data for other customers without explicit written opt-in
Data location and deletion Named regions, sub-processor list, deletion on exit with certificate
Model ownership Ownership or exclusive use of models fine-tuned on your data
Change notification Advance notice of material model or feature changes, release notes, right to defer and roll back
Security Vulnerability notification within agreed timelines, patch support period, SBOM on request
Remote access Only through customer-controlled path, logged, MFA
Performance Agreed metrics for your use case, acceptance criteria at SAT
Regulatory cooperation Vendor provides documentation needed for your AI Act, CRA and NIS2 obligations
Incident cooperation Support for investigations, logs retained for an agreed period
Exit Data return format, transition support

Smaller vendors may push back. Prioritize data use, change notification and remote access.

Effort

  • Low tier: 30 minutes with a short data-use checklist.
  • Medium tier: half a day including vendor call.
  • High tier: 2 to 5 person-days across IT, OT, legal and the use-case owner, plus vendor response time of two to four weeks. Start before the purchase order, not after.

Common mistakes

  • Asking after the PO. Leverage disappears.
  • Sending the IT SaaS questionnaire to an OEM. It asks the wrong questions and the OEM ignores it.
  • Accepting "we do not train on your data" in a sales email. Get it in the contract.
  • Ignoring firmware updates. For embedded AI, firmware is the model update.
  • Allowing the vendor's own modem. It bypasses every zone and conduit you designed.

Where to start

  • Add an AI tick box and a link to the OT AI vendor security questionnaire to your purchase requisition.
  • Pick your three biggest AI-enabled vendors from the AI use-case inventory and send them the data-use and change-notification questions this month.
  • Ask legal for a standard AI addendum with the clauses above.
  • Add "AI features, data flows and remote access documented" to FAT and SAT checklists.

This page is general guidance, not legal advice. Confirm regulatory and contractual positions with your counsel.

Sources and further reading

  1. ISA/IEC 62443 Series of Standards (ISA)
  2. Regulation (EU) 2024/1689 (Artificial Intelligence Act), EUR-Lex
  3. Regulation (EU) 2024/2847 (Cyber Resilience Act), EUR-Lex
  4. CRA reporting obligations (European Commission)
  5. CISA and partners: Principles for the Secure Integration of AI in Operational Technology (Dec 2025)

This guide is general information for manufacturers, not legal advice. Confirm regulatory obligations with qualified counsel.

FAQ

Frequently asked questions

What questions should I ask a vendor about AI training on my data?

Ask whether customer data, including process data, images and user prompts, is used to train or improve models; whether that is opt-in or opt-out; whether data is shared across customers; where it is stored; how long it is kept; and whether you can have it deleted on exit. Ask for the answers in the contract or data processing terms, not a sales email.

Is IEC 62443-4-2 certification required for AI products in OT?

No law requires it in general, but it is a strong signal. IEC 62443-4-1 covers the vendor's secure development lifecycle and 62443-4-2 covers technical security capabilities of components. Ask whether the product or the vendor's process is certified, by whom, and for which security level. If not certified, ask how they align and request evidence.

What is an AI bill of materials?

An AI-BOM extends a software bill of materials to AI components: which models are used, their versions and sources, third-party or open-source models, key libraries, and sometimes training data provenance. Formats are still maturing. For a mid-market buyer, a plain list of models, versions, sources and update method is a useful start.

Who is responsible under the EU AI Act, the vendor or the manufacturer using the system?

Usually the vendor is the provider and you are the deployer, with different duties. That can change. If you put your own name on the system, substantially modify a high-risk system or change its intended purpose, you can take on provider obligations. Integrators who retrain models for your line should be clear in the contract about who is the provider.

Related guides

Know where your plants stand

Score your AI governance, OT security and data readiness in about 10 minutes each. Results stay in your browser.