How do you find the AI your team uses without telling you?

Ask, then read: give every colleague a way to report the AI tools they use, and read the records that already name them, from dependency files and your public pages to expenses and sign-ins. Register each tool as a system with an owner. The EU AI Act never says "shadow AI". It reaches these tools through the deployer definition in Article 3, point (4) and the AI literacy duty in Article 4.

Written and last checked 9 October 2026 against Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744. Article 4 and Article 111(2) were replaced; Article 3, points (3) and (4), Article 6(3), Article 25(1) and Article 26 were not amended.

Does the EU AI Act say anything about shadow AI?

Not in those words: neither Regulation (EU) 2024/1689 nor the 2026 amending act uses the term. What the Act has is roles. A deployer is "a natural or legal person, public authority, agency or other body using an AI system under its authority except where the AI system is used in the course of a personal non-professional activity" (Article 3, point (4)). A colleague drafting customer replies with a chatbot is working, so the one exception does not fit. Whether buying a tool makes you its deployer turns on the same words.

The definition does not settle whose authority a tool runs under when someone adopts it without approval, and recital 13 repeats the definition and adds nothing on whose authority counts. In practice the question can wait: a tool you have found can be approved, given an owner or stopped, and one you have not found cannot.

Which duties need the list?

Article 4(1), as replaced by Regulation (EU) 2026/1744, binds providers and deployers whatever the risk level. They "shall take measures to support the development of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf, taking into account their technical knowledge, experience, education and training and the context the AI systems are to be used in, and considering the persons or groups of persons on whom the AI systems are to be used". The same paragraph adds that the obligation "does not require providers or deployers to guarantee any specific level of AI literacy of any individual". Two of those inputs belong to the system, not the person: its context and the people it is used on. Neither can be weighed for a tool nobody has found. The AI literacy guide covers the rest of the article.

Article 26 binds deployers of high-risk systems. A tool sold for a purpose listed in Annex III, such as one intended "to analyse and filter job applications" (point 4(a)), is high-risk unless Article 6(3) takes it out, and Article 6(3) never does where the system performs profiling of natural persons. Its deployer "shall take appropriate technical and organisational measures to ensure they use such systems in accordance with the instructions for use accompanying the systems, pursuant to paragraphs 3 and 6" (Article 26(1)). It assigns human oversight to natural persons "who have the necessary competence, training and authority, as well as the necessary support" (Article 26(2)). At the workplace, Article 26(7) applies: "Before putting into service or using a high-risk AI system at the workplace, deployers who are employers shall inform workers' representatives and the affected workers that they will be subject to the use of the high-risk AI system. This information shall be provided, where applicable, in accordance with the rules and procedures laid down in Union and national law and practice on information of workers and their representatives." The deployer obligations guide covers the rest of Article 26.

For a high-risk system placed on the market or put into service before the date of application of Chapter III, Article 111(2), as replaced, applies the Regulation to its operators "only if, as from that date, those systems are subject to significant changes in their designs", without prejudice to Article 5 and outside the large-scale IT systems of Article 111(1). In any case, "the providers and deployers of high-risk AI systems intended to be used by public authorities shall take the necessary steps to comply with the requirements and obligations laid down in this Regulation by 2 August 2030" (Article 111(2), as replaced). Article 111(2) names one date of application of Chapter III where Article 113 sets several, two of them in its third paragraph, point (c), for Sections 1, 2 and 3, and the text does not say which one a given system meets.

A general-purpose tool pointed at the same task can change your role. Article 25(1), point (c) makes a deployer the provider of a high-risk system, with the provider obligations of Article 16, where "they modify the intended purpose of an AI system, including a general-purpose AI system, which has not been classified as high-risk and has already been placed on the market or put into service in such a way that the AI system concerned becomes a high-risk AI system in accordance with Article 6". When a deployer becomes a provider works through whether a new task is a new intended purpose.

Where do you look?

Five places, asking first:

1. Ask. A short form anyone can fill in: the tool's name and what they use it for. Say it feeds a register, not a disciplinary file.
2. Read the dependency files. package.json, requirements.txt and their equivalents name the libraries that call a model. A library is not a system, but it tells you which team to ask.
3. Read your own website. Each AI feature it describes is a statement you have already made, and a candidate for the register.
4. Read the spend and the sign-ins. Expense claims show paid subscriptions, and your identity provider shows the apps people sign in to with work accounts. Neither shows a free tool used in a browser, which is why asking comes first.
5. Ask by use. People who cannot name a tool recognise a use: writing code, drafting copy, summarising meetings, sorting applications. A list of uses also catches AI switched on inside software you already pay for.

One tool used for two purposes can be two systems: how many AI systems you have sets out how we count them.

What does Complipath do here, and what does it not?

Complipath suggests and a person decides. The AI inventory has five ways in, all live in the product:

• Intake link: anyone in your company can report an AI tool, with no account and no sight of the register. A named person reviews each report.
• Dependency reading: paste a package.json and it names the AI libraries in it, with the line each one sits on. Reading requirements.txt and go.mod is not built.
• Website reading: it reads your public pages and suggests the AI systems they describe, each anchored to a sentence you published.
• Category intake: twelve categories to tick when you do not know what to call what you run.
• Bulk registration: paste a list, one system per line, optionally with its business owner.

Nothing enters the register until a person adds it, and every entry starts unassessed. Complipath does not read your source code; spend and sign-in records are yours to read. On Starter, the AI literacy tracker then sends each person a briefing built from the law and your register, and records when they read it. Every live Complipath feature is on one page, and AI register software compared ranks the alternatives.

What this means for you

If you're a provider: a model library in your product's package.json may sit inside an AI system you develop and place on the market or put into service under your own name or trademark. That makes you its provider (Article 3, point (3)). Article 4 binds you as one too.

If you're a deployer: run the intake first, because it is the only step that finds a free tool used in a browser. Then classify each entry: what it owes beyond Article 4 follows from its classification, not from how it was found.

FAQ

Is using ChatGPT at work without approval shadow AI under the EU AI Act? The Act does not use the term. A colleague using an AI system for work is not in the one exception of Article 3, point (4), a personal non-professional activity. The definition does not say whose authority an unapproved tool runs under. Finding the tool and deciding whether it may be used settles the practical question.

Do we have to keep a register of every AI tool? No provision asks a deployer for a list of every AI system it uses, and the Article 49 registration concerns systems referred to in Annex III. But the Article 4 measures take account of the context each system is used in, and Article 26 attaches to each high-risk system, so a per-system list is where both start.

Does the AI literacy duty cover tools staff picked up on their own? Article 4(1) binds deployers for "their staff and other persons dealing with the operation and use of AI systems on their behalf", with no risk qualifier. Whether an unapproved tool is one you deploy is the open question in Article 3, point (4). Once you have found it and allowed it, Article 4 leaves none of its users out.

Can a package.json show every AI system we run? No. A dependency file names packages, not systems: one package can sit behind several systems, and a system can be called through an API, built in-house or bought as a service without appearing in it. Use the file to find which team to ask, then ask that team what each package is for.

Sources: Regulation (EU) 2024/1689 (EUR-Lex), Article 3, points (3) and (4), Article 4, Article 6, Article 16, Article 25(1), Article 26, Article 49, Article 111, Article 113, Annex III and recital 13; as amended by Regulation (EU) 2026/1744 (EUR-Lex), which replaced Article 4 in its entirety, Article 111(2) and Article 113, third paragraph, points (a) and (c). Article 3, points (3) and (4), Article 6(3), Article 25(1) and Article 26 were not amended.

Where this question meets the product: the AI inventory. Answering it in a customer's questionnaire: how to answer AI questions in a supplier questionnaire. The feature that does this in the app: every live Complipath feature. Tools ranked on it: Best AI governance software, mid-size firms.