An AI writes to your PLC: who releases the change?
AI agents write into control projects and generate CAD parts. Compiling is not a test. A six-step release procedure for manufacturing companies.
In short
When an AI agent writes into a PLC project, an error-free compile proves only syntax and data types, not machine behaviour. The agent belongs in a copy with no right to download. Release takes six steps: draft in a copy, task and limits, four-eyes review, test, documented release, fallback version. Safety program and safety parameters stay locked for the agent.
Auf Deutsch lesen: deutsche Fassung
Public demonstrations show the following: an AI agent receives a task in plain language. Through an interface, a so-called MCP server, it accesses the engineering software of the controller manufacturer. It creates variables, generates data blocks and writes program code. In a second kind of demonstration it generates a part in the CAD system by the same route. At the end there is always the same check: the project compiles, the model rebuilds. Done.
I run a manufacturing company, audit management systems and help companies build them or develop software themselves. From that experience I say: the technology is impressive, and it will arrive in companies. What interests me is the route by which the agent's code reaches the machine. The demonstrations say nothing about that route.
Compiling is not a test
An error-free compile says exactly one thing: the syntax is correct and the data types fit. Whether the axis moves in the right direction, whether an interlock engages at the right point and whether a time suits the process, it does not say.
The same applies to the CAD part. A model that rebuilds without errors is a body with dimensions. Whether it can be clamped, whether the tolerances suit the function and whether the wall thickness holds, nobody has checked.
A pattern for illustration, freely chosen: the agent is to change a dwell time from two to five seconds. It finds the variable, changes the value, the project compiles without errors. The same variable serves as a monitoring time in a second place. The compiler does not report that. A person who knows the plant would have looked.
Five questions before the first use
Before an agent may write into a control project, I want to see five answers.
- Who makes the change? The agent works under a user account. The project then shows the name of an employee who never wrote the line. Define how the agent's changes are marked.
- Who releases? A qualified person who knows the plant. The agent that writes the code does not sign it off itself.
- What is in the change log? The task in its wording, the blocks changed, the state before and after, the reviewer, the date.
- How is it tested? In simulation or on a test controller before the version goes to the machine. The test cases are fixed before the run.
- Who may execute code in the CAD account? As I understand it, scripts can run through such interfaces with the rights of the logged-in account. The agent can then do everything the designer can, including overwriting or deleting files. Give it an account of its own with access to one working folder and nothing else.
Which rights and limits an agent needs before it works on its own:
Securing AI agentsWhen a change becomes a substantial modification
Until 20 January 2027 the Machinery Directive 2006/42/EC applies. On that day it is repealed (Article 51(2) of Regulation (EU) 2023/1230), and the Machinery Regulation applies (Article 54). The term substantial modification does not appear in the text of the Directive. The Regulation defines it in Article 3(16) and expressly includes digital modifications.
In substance: it means a modification after placing on the market or putting into service that the manufacturer neither foresaw nor planned and that affects safety because a new hazard is created or an existing risk increases. A second condition is added. Guards or protective devices become necessary whose integration modifies the safety control system, or additional protective measures for stability or strength.
The consequence is in Article 18: whoever carries out a substantial modification is considered a manufacturer. They bear the manufacturer's obligations under Article 10 and must apply the conformity assessment procedure.
For practice I read it like this: a changed dwell time in the process sequence does not, as a rule, reach this threshold. It becomes critical when a change touches speeds, forces, ranges of movement, interlocks or operating modes. That assessment belongs in a documented evaluation by a qualified person, before the change. The safety program and safety parameters are locked for the agent. Full stop. And in case of doubt, a specialist body for machinery safety confirms the classification.
One more detail from Annex III of the Regulation. Under point 1.1.9, machinery must collect evidence of interventions in its software and its configuration. Point 1.2.1(f) requires a tracing log of the versions of the safety software. For new machinery the change log thus becomes part of the design. In your company you should already keep it today.
Which obligations additionally apply to manufacturers of products with digital elements:
Cyber Resilience Act for manufacturersAnd the AI Act?
The agent that drafts code at the desk is a development tool. As I understand it, that does not make it a component of the machine. The AI Act, Regulation (EU) 2024/1689, starts elsewhere. Under Article 6(1) an AI system is high-risk if it serves as a safety component of a product from Annex I, or is itself such a product, and that product requires a third-party conformity assessment. The Machinery Regulation is listed in Annex I Section B as point 21. In its own Annex I Part A, points 5 and 6, it names safety components and embedded systems with self-evolving behaviour based on machine learning that ensure safety functions.
Which manufacturing AI counts as high-risk and which does not:
AI in CNC manufacturing and the EU AI ActArticle 6(1a) draws the line: AI systems used exclusively for non-safety-related aspects such as automation, performance optimisation or quality control are not considered safety components.
Article 25(1) governs when deployers or third parties become the provider of a high-risk AI system. That applies, for example, to a substantial modification of such a system (point b) or when someone changes the intended purpose of an AI system so that it becomes high-risk (point c). For machinery there is a special feature. Under Article 2(2), only Article 6(1), Article 60a and Articles 102 to 112 apply to high-risk AI in the area of Annex I Section B. The substantive requirements come through delegated acts on Annex III of the Machinery Regulation, which apply from 2 August 2028 (Article 8, third and fourth paragraphs, of the Machinery Regulation). Whoever builds a learning component into a safety function therefore lands first at Article 18 of the Machinery Regulation.
That is my reading as a practitioner. On legal questions a specialist should always take a look before you decide on that basis.
The release procedure in six steps
- Draft in a copy. The agent works in a copy of the project or in simulation. It has no connection to the running controller and no right to download.
- Record the task and its limits. What is to be changed, which blocks are off limits? The safety program and safety parameters are always excluded.
- Four-eyes review. A qualified person reads the comparison between the old and the new state line by line. They assess whether the change touches safety or the range of movement and whether a substantial modification comes into question.
- Test. Test cases fixed in advance run in simulation or on the test controller, then on the machine in setup mode. The result is filed with a date.
- Documented release. Name, date, version, result of the assessment. Only then is the program downloaded, and a human does the download.
- Fallback version. The last released version is kept ready as a backup. Someone has practised restoring it at least once.
For CAD parts the same procedure applies in a small form: draft in the working folder, check by the designer, release through the normal drawing process.
How halting and restarting have to be regulated, matching the fallback version:
Emergency stop for AI: halt and restartWhere this connects to ISO/IEC 42001
The standard knows the building blocks of this procedure. I transfer them to the agent as a deployed AI system.
- A.6.2.4: The organisation defines and documents how it verifies and validates and by which criteria. Compiling alone is not a criterion.
- A.6.2.5: There is a deployment plan, and the requirements are met before deployment.
- A.6.2.8: The organisation determines in which phases events are logged.
- A.3.2: Roles and responsibilities are assigned. Who commissions, who reviews, who downloads.
In an audit I ask to see the last change an agent was involved in: task, comparison, test record, release, fallback version. If one of them is missing, the procedure is not finished yet. Network protection and safety-related controllers have their own bodies of rules, keywords IEC 62443 and functional safety.
What to do on Monday
Ask in maintenance and design who is already experimenting with such interfaces. Lay down that agents work only in copies and that nobody downloads a version that has not gone through the six steps. Hope is not a strategy. A one-page procedure is.
This is my interpretation from practice. For legal advice, please turn to the appropriate professionals.
Primary sources
- Regulation (EU) 2023/1230 (Machinery Regulation), consolidated version of 27 July 2026, Article 3(16), Article 8, Article 18, Article 51, Article 54, Annex I Part A points 5 and 6, Annex III points 1.1.9 and 1.2.1 (EUR-Lex, accessed on 19 November 2026)
- Directive 2006/42/EC (Machinery Directive) (EUR-Lex, accessed on 19 November 2026)
- Regulation (EU) 2024/1689 (AI Act), consolidated version of 27 July 2026, Article 2(2), Article 6(1) and (1a), Article 25, Annex I Section B point 21 (EUR-Lex, accessed on 19 November 2026)
- ISO/IEC 42001:2023, Annex A, controls A.3.2, A.6.2.4, A.6.2.5 and A.6.2.8 (iso.org)
Release procedure for changes an AI agent makes to a control project
Draft in a copy
The agent works in a copy of the project or in simulation. It has no connection to the running controller and no right to download.
Record the task and its limits
What is to be changed, which blocks are off limits? The safety program and safety parameters are always excluded.
Four-eyes review
A qualified person reads the comparison between the old and the new state line by line. They assess whether the change touches safety or the range of movement and whether a substantial modification comes into question.
Test
Test cases fixed in advance run in simulation or on the test controller, then on the machine in setup mode. The result is filed with a date.
Documented release
Name, date, version, result of the assessment. Only then is the program downloaded, and a human does the download.
Fallback version
The last released version is kept ready as a backup. Someone has practised restoring it at least once.
Frequently asked questions
May an AI agent write directly into the running controller?+
Technically that can be set up. I advise against it. The agent belongs in a copy of the project or in simulation, with no connection to the machine and no right to download. A checked and released version is downloaded by a human.
Is it enough if the project compiles without errors?+
No. The compile checks syntax and data types. It says nothing about the behaviour of the machine. That takes test cases fixed before the change and a qualified person who reads the comparison between the old and the new state.
From when does the Machinery Regulation apply, and what does it mean for software changes?+
Regulation (EU) 2023/1230 applies from 20 January 2027 (Article 54). It defines substantial modification in Article 3(16) and expressly includes digital modifications. Whoever carries out a substantial modification is considered a manufacturer under Article 18.
Does a change made by the agent turn our company into a manufacturer?+
That depends on the change. Which tool wrote it plays no part. What matters under Article 3(16) is whether a new hazard is created or a risk increases and additional protective measures therefore become necessary. A qualified person assesses that before the change and records it in writing.
Author & expert review: Lars Zimmermann · ISO/IEC 42001 Senior Lead Auditor & Senior Lead Implementer · ISO/IEC 27001 Lead Auditor & Lead Implementer (PECB)
Last updated: 19 November 2026. Researched and reviewed to the best of our knowledge; not a substitute for individual legal advice.
Questions about your own case?
In a free 15-minute intro call we assess where you stand on ISO 42001, ISO 27001 and the EU AI Act, honestly and without a sales pitch.