Key takeaways
- Ask about value first. AI that cannot point to OEE, scrap, energy or working capital outcomes is a cost, not a strategy.
- The biggest AI risks in a plant are usually OT security, IP leakage and unvalidated decisions on quality or safety, not science-fiction scenarios.
- A good answer names an owner, a list, a number and a date. A weak answer names a vendor or a pilot.
- Under NIS2, management bodies of in-scope entities must approve and oversee cybersecurity risk measures and follow training. AI in OT is part of that.
- Ask for a one-page quarterly AI and OT risk report, not a 40-slide deck.
Boards and owners of mid-market manufacturers are hearing two stories about AI at once. One says competitors are using it to cut scrap and downtime and you will fall behind. The other says it brings new cyber, legal and IP risk. Both are partly true. The board's job is not to pick a story but to make sure management has a grip on both value and risk, in proportion to the size of the business.
This page gives a set of questions to ask, grouped into strategy and value, risk and control, and capability. For each, it shows what a good answer and a weak answer sound like. It closes with what NIS2 means for management accountability and a template for quarterly reporting.
Does this apply to you?
If you sit on the board of, own, or chair a manufacturer with 100 to 2,500 employees, yes. Family-owned and private-equity-backed businesses are no exception. In fact, they often have the leanest IT and OT teams and the most concentrated know-how, which makes both the value and the risk sharper.
If your company is in scope of NIS2 in the EU, there is also a legal dimension, covered below.
Strategy and value
1. Which AI use cases are live today, and what has each delivered against a baseline?
- Good answer: "We have 14 registered use cases, 4 in production. Vision inspection on Line 3 cut escapes by an amount we measured against six months of baseline. PdM on compressors has a cost avoidance estimate reviewed by finance. Two pilots were stopped because they did not beat the baseline."
- Weak answer: "We are running several exciting pilots with leading vendors."
2. How does AI connect to our operational priorities: OEE, scrap, energy, working capital, labor availability?
- Good: Each use case maps to a plant KPI with an owner.
- Weak: AI is a separate "innovation" budget line with no KPI owner.
3. What are we deliberately not doing with AI, and why?
- Good: "We are not letting AI release product or write setpoints on safety-critical equipment until validation and fallback are proven. We are not deploying productivity monitoring of individuals."
- Weak: "Nothing is off the table."
4. Where is AI arriving through vendors and OEMs rather than our own decisions?
- Good: "Our register shows 9 of 14 use cases came through vendor software or equipment. Purchasing now flags AI features and we send a standard questionnaire."
- Weak: "We only use AI where we have chosen to."
Risk and control
5. Do we have a complete register of AI in the company, including vendor and OEM features, with an owner for each?
- Good: A named register owner, a count by risk tier, every row has a business owner, and MOC and purchasing feed new entries. See AI use-case inventory.
- Weak: "IT keeps track of that."
6. Could an AI system, or the vendor connection behind it, give an attacker a path into our control systems?
- Good: "AI devices in OT are on our asset inventory, sit in defined zones, and vendor access goes through our controlled remote access path with MFA and logging. We reviewed this in the last 12 months."
- Weak: "The vendor says it is secure." or "Our OT network is air-gapped." (It rarely is.)
7. Which AI use cases can act on production, quality release, safety or people without a human checking each decision?
- Good: A short list, each with a documented risk assessment, limits enforced outside the AI, a manual fallback, and approval through MOC.
- Weak: "None, a human is always in the loop." (Ask how many decisions per shift that human reviews.)
8. How are we protecting recipes, process parameters, drawings and code from leaking into AI tools?
- Good: "We have an approved enterprise GenAI tool, a published list of what never goes into AI tools, blocking of public tools on company devices, and training completion above our target."
- Weak: "We told people not to use ChatGPT."
9. What would we do if an AI system caused a quality escape, a line stop or a cyber incident tomorrow?
- Good: AI is covered in the incident response plan, there is a manual fallback per production use case, and the team has walked through a scenario. See OT incident response.
- Weak: "We would call the vendor."
10. What regulatory obligations apply to our AI, and when?
- Good: A dated summary. For example, as of November 2026: AI literacy duty in force since February 2025; Article 50 transparency obligations from 2 August 2026; Annex III high-risk obligations, including employment uses, from 2 December 2027 and AI in products under Annex I legislation from 2 August 2028, following the Digital Omnibus on AI (Regulation (EU) 2026/1744). Plus a view on whether any of our systems fall in those categories.
- Weak: "We are compliant with the AI Act." (Most obligations do not apply yet, so this claim cannot be tested.)
Note: Customer requirements often arrive before regulation. Automotive, aerospace and medical customers increasingly ask about AI in supplier questionnaires. A good answer to question 10 includes customer obligations, not only laws.
Capability
11. Who owns AI governance, how much of their time does it take, and who do they report to?
- Good: A named executive sponsor, a small cross-functional group with operations, IT, OT, quality, EHS and HR, meeting monthly, with authority to stop a deployment. See roles and RACI.
- Weak: "Our IT manager handles it."
12. Do we have the OT security and data skills to run AI on the shop floor, or are we fully dependent on vendors?
- Good: An honest view: "We have one controls engineer with OT security training and rely on an integrator for the rest. We plan to add one data engineer next year."
- Weak: "Our vendors handle that."
13. What is our data foundation like: can we trust the historian and MES data that AI depends on?
- Good: A data readiness score per site and a plan to fix the worst gaps.
- Weak: "We have lots of data."
14. What training have management and staff had on AI and cybersecurity this year?
- Good: Board and management have had a session; operators and engineers have had role-based training; completion is tracked.
- Weak: "We sent round a policy."
NIS2 and management accountability
NIS2 (Directive (EU) 2022/2555) applies to medium-sized and larger entities in listed sectors, which include several manufacturing sub-sectors. Article 20 requires the management bodies of essential and important entities to approve the cybersecurity risk-management measures, oversee their implementation, and follow training. It also allows member states to hold management bodies liable for infringements.
As of November 2026, most EU member states have transposed NIS2 into national law, though a few had not completed it and details differ by country, including registration, reporting routes and how management liability works. Check the law in each country where you operate. Our NIS2 page has the basics.
AI is not singled out in NIS2. But AI systems in OT, vendor remote connections and cloud data flows are part of the attack surface that the management body must oversee. The practical consequence for boards: AI and OT risk should be reported through the same channel as cybersecurity, and the board minutes should show it was considered.
Watch out: "Approve and oversee" means more than noting a report. Ask for the risk measures, challenge them, and record decisions. A board that only receives slides may struggle to show oversight if something goes wrong.
What to ask management to report
A one-page quarterly AI and OT risk report should include:
| Item | Example |
|---|---|
| AI use cases by tier and status | 14 registered: 2 high, 5 medium, 7 low; 4 in production |
| Value delivered | KPI movement per production use case, with baseline |
| High-tier decisions taken | Approved closed-loop energy optimization on Furnace 2 with limits |
| Incidents and near misses | 1 model drift event on vision, caught by challenge parts |
| OT security posture | Vendor remote access paths reviewed; open findings |
| Regulatory and customer changes | New customer AI clause; AI Act dates as of this quarter |
| Capability and training | Training completion; open roles |
| Asks of the board | Budget, risk appetite decisions |
The board AI risk briefing gives you a ready template.
Effort
- Preparing the first board briefing: 3 to 5 days of management time if a register exists; longer if not.
- Quarterly updates: half a day once the register and governance group are running.
- Annual deep dive: a two-hour board session, ideally with an outside expert.
Common mistakes
- Asking only about risk. Management hears "stop", and value goes elsewhere.
- Accepting vendor assurances as evidence.
- Treating AI and OT security as separate topics. In a plant, they share the same networks, vendors and people.
- Delegating NIS2 entirely to IT. The duties sit with the management body.
Where to start
- Put AI and OT risk on the next board agenda with a 30-minute slot.
- Send management questions 5, 6, 7 and 11 in advance and ask for a written answer.
- Ask for a one-page register summary by risk tier.
- If governance does not exist yet, ask management to present a 90-day plan at the following meeting.
This page is general guidance, not legal advice. Confirm regulatory and director-liability obligations with your counsel.
Sources and further reading
- Directive (EU) 2022/2555 (NIS2), EUR-Lex
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), EUR-Lex
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), EUR-Lex
- NIST AI Risk Management Framework
- 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 should a board ask about AI in a manufacturing company?
Start with: what are our AI use cases and what value has each delivered; which of them can affect production, quality, safety or worker data; who owns each; how we control vendors and OEMs who bring AI in; whether our OT network could be reached through an AI system; and whether we have the people to run this. Ask for evidence, not reassurance.
Are board members personally liable for cybersecurity under NIS2?
NIS2 requires management bodies of essential and important entities to approve cybersecurity risk-management measures, oversee their implementation and follow training, and requires member states to ensure they can be held liable for infringements. The details depend on national transposition law. As of November 2026, most EU member states have transposed NIS2, but the specifics vary. Confirm your position with counsel.
How often should a board review AI risk?
For most mid-market manufacturers, a one-page AI and OT risk update each quarter, with a deeper review once a year, is proportionate. Add ad hoc reporting for material incidents, major AI investments, or any decision to let AI act automatically on production, quality release or worker management.
Does a mid-market manufacturer need a board-level AI expert?
Not necessarily. It helps to have at least one board member comfortable with technology and operations risk, and to invite outside expertise for an annual deep dive. More important is that management reports clearly and that the board knows which questions to ask and what a credible answer looks like.
Related guides
AI Governance
AI governance roles and RACI for mid-market manufacturers
Mid-market manufacturers need a small AI governance group drawn from roles they already have, a clear RACI across the AI lifecycle, and a firm split between what plant…
7 min read
Frameworks
NIS2 for manufacturers: scope, duties and what to do first
NIS2 (Directive (EU) 2022/2555) brings many medium and large manufacturers into EU cybersecurity law for the first time. If you make medical devices, electronics…
8 min read
AI Governance
How to build an AI use-case inventory in a manufacturing plant
An AI use-case inventory is a register of every AI system you build, buy or have embedded in equipment, with an owner, a risk tier and enough detail to act on. In a…
7 min read
OT Security
OT incident response with AI in the mix: playbooks and tabletops
OT incident response starts with safety and keeping the process in a known state, not with wiping machines. Plan for manual operation, include AI failures alongside…
8 min read