How do you build an AI register for the EU AI Act?

Last reviewed: 28 September 2026

General information, not legal advice. This guide explains regulation and good practice as of 28 September 2026. Have your own legal counsel review your register and your classifications.

What is an AI register, and how is it different from an AI tool inventory?

An AI register (also called an AI system inventory) is a controlled list of the AI your organisation uses or builds, kept up to date and reviewed. An "AI tool inventory" usually means only the list of tools; a register adds the decisions and the evidence: who approved each tool, for what, with which data, under which AI Act role and risk category.

In practice, most SMBs have three kinds of AI to record:

1. 1Stand-alone AI tools such as ChatGPT, Microsoft Copilot, Claude or Gemini.
2. 2AI features inside software you already use, such as CV screening in an applicant tracking system, AI summaries in a CRM, or transcription in a meeting tool.
3. 3AI you build or brand yourself, such as a customer chatbot on your website or a model inside your product. This is where you may become a provider, not just a deployer.

The AI Act defines an AI system as "a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions" (Article 3(1)). If you are unsure whether something qualifies, record it anyway and mark the AI Act role as "unclear" until you have checked it against the Commission's guidelines on the AI system definition.

Does the EU AI Act require an AI register?

Not for most companies, and not by that name. The registration duties in the AI Act concern the EU database for high-risk AI systems (Article 71), not an internal company register:

• Providers of high-risk AI systems listed in Annex III must register themselves and their systems in the EU database before placing them on the market or putting them into service (Article 49(1)).
• Deployers that are public authorities, EU institutions or bodies, or persons acting on their behalf, must register their use of Annex III high-risk systems (Article 49(3); Article 26(8)).
• A private company that only uses AI tools has no duty under the AI Act to register them anywhere.

So why keep one? Because several duties that do apply to ordinary deployers are hard to meet without knowing what AI you use:

Timeline status (checked 28 September 2026). The Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force on 27 July 2026. According to the Commission's AI Act page, it set the high-risk rules for Annex III areas (including biometrics, critical infrastructure, education, employment, migration, asylum and border control) to apply from 2 December 2027, and for AI in regulated products such as lifts or toys from 2 August 2028. It also rewrote Article 4 so that deployers must "take measures to support the development of AI literacy" without having to "guarantee any specific level of AI literacy of any individual". Check the Commission page again before you rely on any date: the Omnibus changed the high-risk timeline in 2026, and further guidance is still being published.

Which fields should an AI register have?

Start with the twelve required fields, then add the rest as your register matures. Field names below are the CSV column headers. Use one row per tool per use: if HR and marketing use the same tool for different purposes with different data, give them separate rows.

Use ISO dates (YYYY-MM-DD) and fixed values in the controlled columns so you can filter and count. Save the file as UTF-8 so names with ä, ö or å survive.

What does a filled-in register look like?

The CSV below has the header row and six example rows. Every example row starts with EXAMPLE- in register_id and says "EXAMPLE ROW" in notes. Delete them and add your own. The vendor facts in rows 1 to 3 reflect each vendor's public documentation on 28 September 2026 (sources below); check them again before you rely on them. Rows 4 to 6 use placeholder vendors.

register_id,tool_name,vendor,account_type,ai_type,business_purpose,business_owner,departments,approx_users,status,ai_act_role,ai_act_risk_category,annex_iii_area,art5_check_done,art50_disclosure,personal_data,highest_data_class_allowed,file_uploads_allowed,vendor_trains_on_inputs,data_location,dpa_in_place,dpia_status,gdpr_record_ref,human_oversight,literacy_module,approved_by,approved_on,vendor_terms_checked_on,last_reviewed,next_review,notes
EXAMPLE-001,ChatGPT,OpenAI,ChatGPT Business (company workspace),standalone_tool,"Drafting, summarising and translating internal text",[Name] (Head of Operations),Sales;Marketing;Operations,18,approved,deployer,minimal,,yes 2026-09-28,none,no,internal,yes,no_by_default,[per contract],yes,not_required,,Author checks facts before sending,AI basics v1,[Name] (CEO),2026-09-28,2026-09-28,2026-09-28,2027-03-28,"EXAMPLE ROW - replace. OpenAI states it does not train on ChatGPT Business data by default."
EXAMPLE-002,Microsoft 365 Copilot Chat,Microsoft,Work account (Entra ID) with enterprise data protection,standalone_tool,Questions and drafting with company documents,[Name] (IT lead),All staff,25,approved,deployer,minimal,,yes 2026-09-28,none,yes,confidential,yes,no_by_default,[per contract],yes,done,ROPA-07,Author checks output,AI basics v1,[Name] (CEO),2026-09-28,2026-09-28,2026-09-28,2027-03-28,"EXAMPLE ROW - replace. Microsoft states prompts and responses are not used to train foundation models."
EXAMPLE-003,ChatGPT,OpenAI,Personal or free account,standalone_tool,None for company work,[Name] (Policy owner),,0,prohibited,deployer,not_assessed,,,none,no,public,no,user_controlled,,no,not_assessed,,,,[Name] (CEO),2026-09-28,2026-09-28,2026-09-28,2027-03-28,"EXAMPLE ROW - replace. Company has no admin control over personal accounts."
EXAMPLE-004,CV screening feature,[ATS vendor],Company licence,embedded_feature,Ranking incoming job applications,[Name] (HR manager),HR,2,under_review,deployer,high_risk_annex_iii,4(a),yes 2026-09-28,none,yes,personal,no,unknown,[ask vendor],no,required,ROPA-03,Recruiter makes and records every decision,AI basics v1 + HR module,,,2026-09-28,2026-09-28,2026-10-31,"EXAMPLE ROW - replace. Annex III rules apply from 2 Dec 2027; legal review before any use."
EXAMPLE-005,Website chat assistant,[Chatbot vendor],Company deployment under own brand,built_in_house,Answering customer questions on the website,[Name] (Customer service lead),Customer service,3,restricted,unclear,transparency_art50,,yes 2026-09-28,chatbot_notice,yes,personal,no,unknown,[ask vendor],yes,not_assessed,ROPA-11,Escalation to a person on request,AI basics v1 + chatbot module,[Name] (CEO),2026-09-28,2026-09-28,2026-09-28,2026-12-28,"EXAMPLE ROW - replace. Check whether branding it makes you a provider (Art 3(3)); tell users they are talking to AI."
EXAMPLE-006,Meeting transcription,[Transcription vendor],Company licence,embedded_feature,Notes for internal meetings,[Name] (Office manager),All staff,12,restricted,deployer,minimal,,yes 2026-09-28,none,yes,internal,no,unknown,[ask vendor],no,not_assessed,,Organiser checks notes before sharing,AI basics v1,[Name] (CEO),2026-09-28,2026-09-28,2026-09-28,2026-12-28,"EXAMPLE ROW - replace. Internal meetings only; tell participants before recording."

How do you build the register in a week?

A first version takes a few days if you start from real usage rather than a survey.

1. 1Collect what is actually used. Check SaaS invoices and expense claims for AI subscriptions, ask each team lead, and look at browser or network data if you have it. Surveys and firewall logs miss personal accounts and remote work; discovery in the company browser shows which AI sites people actually open.
2. 2Find the AI inside existing software. Go through your main systems (CRM, HR, finance, support, office suite) and note which AI features are switched on.
3. 3Create one row per tool per use. Fill in the twelve required fields first.
4. 4Decide a status for each row. Approved, restricted, prohibited or under review. Give people an approved company option for common tasks.
5. 5Check role and risk. For each row, record whether you are deployer or provider, check the use against Article 5 and Annex III, and note any Article 50 disclosure. Send anything marked high-risk or unclear to legal counsel.
6. 6Link to privacy work. For rows with personal data, add the GDPR record reference, check the data processing agreement and decide whether a DPIA is needed.
7. 7Publish and connect. Make the approved list visible to staff through your AI acceptable use policy, and plan AI literacy training around the tools and risks in the register.
8. 8Set a review rhythm. Review the register at least every six months, whenever a new tool appears, and when the rules change.

How does the register connect to your GDPR records?

Keep them separate but linked. Under GDPR Article 30, controllers must keep a record of processing activities, including purposes, categories of data subjects and personal data, recipients, transfers, retention and security measures, and make it available to the supervisory authority on request. Article 30(5) exempts organisations with fewer than 250 employees, unless the processing is likely to result in a risk to people's rights and freedoms, is not occasional, or includes special categories of data or criminal-offence data. Regular use of AI tools with customer or employee data is usually not occasional, so check this with your privacy contact.

A practical split: the AI register records the tool, the decision and the AI Act view; the GDPR record describes the processing. The gdpr_record_ref column links them.

When a planned AI use is likely to result in a high risk to people's rights and freedoms, for example new technology applied to employee data, a data protection impact assessment is required before processing starts, as the Finnish Data Protection Ombudsman explains. If you later deploy a high-risk AI system, Article 26(9) says to use the provider's Article 13 information in that assessment.

What are the most common mistakes?

• Calling it an "AI tool inventory" and stopping at tool names. A list without owners, statuses and data classes does not support any decision.
• Missing embedded AI. The riskiest uses are often features switched on inside HR or finance software, not chat assistants.
• One row per tool, not per use. The same tool can be minimal risk in marketing and high-risk in recruitment.
• Assuming "we only use AI, so the AI Act does not apply". Deployers have duties now: Article 4 literacy measures and Article 5 prohibitions have applied since 2 February 2025.
• Assuming you must register everything in the EU database. That duty sits with providers of Annex III high-risk systems and with public-authority deployers.
• Never recording vendor terms. Whether a vendor trains on your inputs depends on the plan. Record what the terms said and when you read them.
• No review date. A register that was accurate in March is wrong by September.

Checklist

• Every AI tool and embedded AI feature found, including personal-account use
• One row per tool per use, twelve required fields complete
• Status decided for every row; approved options available for common tasks
• AI Act role and risk category recorded; Article 5 check done
• High-risk and unclear rows sent to legal counsel
• Article 50 disclosures noted where needed
• Personal-data rows linked to GDPR records, DPA and DPIA decisions
• Register matches the AI acceptable use policy's tool list and data classes
• Literacy training mapped to the tools in the register
• Owner and next review date set

How does VAHTOR help keep the register current?

VAHTOR is an AI governance platform from Vahtor Oy (Helsinki, Finland). It keeps the browser-based part of an AI register current. It does not decide your AI Act classifications, is not legal advice, and does not by itself make a company compliant.

• Discovery: people install the VAHTOR extension in Chrome or Edge and join with a personal device code; the AI sites they use appear in one company inventory with the tool name and time.
• Status decisions: approve, restrict or block each tool once for the company; devices pick up the decision on their next sync, and changes are logged in the audit trail.
• High-risk flagging and reporting: flag tools or uses that may fall into a high-risk category for review, and export basic EU AI Act reporting; classification decisions stay with you and your advisers. High-risk flagging is in the Business and Enterprise plans; Starter includes EU AI Act basic reporting.
• No chat content: VAHTOR does not collect prompts or conversations.

What VAHTOR does not cover: AI features inside other software that do not run as AI sites in the browser, and AI you build yourself. Record those rows by hand. See EU AI Act records for companies that use AI and shadow AI discovery.