Skip to content
SimpleAct Logo

Guide · IT & engineering

As of 6 October 2026 · after the Digital Omnibus

EU AI Act for IT Teams and CTOs: Duties, Shadow AI and Integrations

Sooner or later the AI Act lands on IT’s desk: who knows which AI services run in the tenant? Who keeps the logs, who runs the interfaces, who sets access rights? This guide shows which duties need a technical implementation, where IT only contributes, and what is due in the next 90 days.

  • 12 duties with article references
  • Find and assess shadow AI
  • RACI excerpt for IT
  • Checklist with progress
app.simpleact.de

AI Discovery

Demo Organisation GmbH · last scan today

A

14

findings in total

5

new, not reviewed

6

promoted to inventory

3

ignored

Findings from Entra ID, Intune and network logs

  • ChatGPT · OpenAIEntra ID (OAuth) · 23 usersNew
  • Translation service · DeepLNetwork log (CSV) · 41 usersPromoted
  • Transcription appIntune · 6 devicesNew
  • Copilot · MicrosoftEntra ID (OAuth) · 112 usersReviewed

Connectivity

  • SSO: OIDC · SAML · LDAPPro
  • SCIMEnt.
  • REST APIEnt.
  • Jira
  • Teams
  • ServiceNow

Sample data · AI Discovery from Professional

In short

For IT teams the EU AI Act mostly means: know which AI runs in the organisation, operate it as the provider instructs, and be able to prove how it is run. For high-risk AI, logs (at least six months), monitoring and reporting paths follow from 2 December 2027. IT does not make the legal classification alone – it supplies facts, technology and evidence.

Basics

Why the EU AI Act is an IT topic – even though IT is not the “addressee”

The AI Act (Regulation (EU) 2024/1689) assigns duties to roles, not departments. An organisation is a deployer when it uses an AI system under its own authority (Art. 3(4)), and a provider when it develops an AI system, or has one developed, and places it on the market or puts it into service under its own name (Art. 3(3)). The IT department is not mentioned anywhere. Even so, much of the implementation lands with IT – because almost every duty has a technical side.

It starts with visibility. Without a complete AI inventory no duty can be assigned, and most AI tools do not arrive through an IT project but through a credit card, an OAuth consent or a new feature in software you already use. Whoever manages the tenant, the device fleet and the proxy sees those traces first. In most organisations shadow AI is therefore an IT finding before it becomes a compliance topic.

It continues with operations. For high-risk AI, Art. 26 requires deployers to take technical and organisational measures so the system is used in accordance with its instructions for use (para. 1), to ensure that input data under their control is relevant and sufficiently representative for the intended purpose (para. 4), to monitor operation (para. 5) and to keep automatically generated logs under their control for at least six months (para. 6). Those are jobs for platform, operations and security teams.

And it ends with the question of when a deployer becomes a provider. Anyone who puts their own name on a high-risk system, substantially modifies it, or changes the purpose of an AI system so that it becomes high-risk takes on the provider obligations under Art. 25. For engineering teams that fine-tune models, build them into their own products or point workflows at a new audience, that line is an architecture question – not a purely legal one.

The Digital Omnibus (Regulation (EU) 2026/1744, in force since 27 July 2026) moved the high-risk deadlines: Annex III applies from 2 December 2027, Annex I from 2 August 2028. That buys time for clean engineering but does not change the duties that already apply: measures on AI literacy (Art. 4) and the ban on certain practices (Art. 5) since 2 February 2025, and the transparency duties under Art. 50 since 2 August 2026.

minimum log retention for deployers of high-risk AI (Art. 26(6))
6 mo
minimum log retention for deployers of high-risk AI (Art. 26(6))
high-risk duties for Annex III systems after the Digital Omnibus
2 Dec 2027
high-risk duties for Annex III systems after the Digital Omnibus
to a solid IT baseline (recommendation)
90 days
to a solid IT baseline (recommendation)

Duties

What the EU AI Act means for IT teams and CTOs

The table sorts the main duties under the AI Act and the GDPR by who they address and what IT contributes. The duty always falls on the organisation in its role as deployer or provider; the right-hand column describes the part that typically sits with IT.

Scroll the table sideways →

What the EU AI Act means for IT teams and CTOs
DutyBasisAddresseeApplies since / fromIT’s contribution
Promote AI literacyArt. 4 AI ActProviders and deployers2 Feb 2025Technical briefings for admins and power users, usage rules in the tools, attendance tracked.
Avoid prohibited practicesArt. 5 AI ActAll operators2 Feb 2025Spot features such as workplace emotion recognition or biometric categorisation in tools in use and have them switched off.
Know and map AI systemsPrerequisite for Art. 4, 5, 26, 50DeployersongoingFind shadow AI in the tenant, devices and proxy logs; turn findings into inventory records.
Transparency for chatbots and AI contentArt. 50(1), (2), (4) AI ActProviders (1, 2), deployers (4)2 Aug 2026Implement notices in the chat widget; check that generators mark output machine-readably; deepfake labelling in the publishing workflow.
Marking for legacy systemsArt. 50(2) AI Act (Omnibus)Providers2 Dec 2026Own generative systems placed on the market before 2 Aug 2026: retrofit marking by 2 Dec 2026.
Use according to instructionsArt. 26(1) AI ActDeployers of high-risk AI2 Dec 2027Set up configuration, rights and interfaces as the provider’s instructions for use (Art. 13) describe.
Control input dataArt. 26(4) AI ActDeployers of high-risk AI2 Dec 2027Define data sources and interfaces so that input data under your control fits the purpose and is representative.
Monitor operation, inform the providerArt. 26(5) AI ActDeployers of high-risk AI2 Dec 2027Monitoring, alerting and a reporting path to the provider; be able to suspend use on risk or a serious incident.
Keep logsArt. 26(6) AI ActDeployers of high-risk AI2 Dec 2027Keep automatically generated logs under your control for at least six months – aligned with the retention policy.
Spot a switch to the provider roleArt. 25(1) AI ActDeployers, distributors, third parties2 Dec 2027 (Annex III)Have own branding, substantial modifications and purpose changes on high-risk systems reviewed before release.
Security of processingArt. 32 GDPRControllers, processorssince 25 May 2018Access control, encryption, logging and recoverability for AI services and their interfaces too.
Report personal data breachesArt. 33 GDPRControllerssince 25 May 2018Detect, contain and document – notification to the authority where feasible within 72 hours.

Simplified overview. High-risk duties apply only to Annex III systems (from 2 Dec 2027) or Annex I systems (from 2 Aug 2028). Whether a system is high-risk follows from its purpose, not its technology.

In detail

Six topics IT teams cannot avoid

Each topic covers what the Regulation requires, what that means technically and where the usual pitfalls are. Recommendations are marked as such.

01

Finding shadow AI – without a surveillance project

Basis for Art. 4, 5, 26, 50 AI Act · German works council law (§ 87(1) no. 6 BetrVG)

The AI Act contains no explicit duty to keep an inventory. In practice it is still the prerequisite for everything else: a system nobody knows about cannot be classified, labelled or monitored. The hardest records are not the big projects but browser tools, plug-ins and AI features that arrive with an update to software you already use.

Data you already produce is usually enough for the search: OAuth consents granted in the directory, the app list of managed devices, proxy or DNS logs, and licence and card statements. No agent on every endpoint is required. The second step is what matters: decide on each finding individually – promote, ignore or mark as reviewed – instead of dumping every hit into the inventory.

Note: discovery itself processes employee data. It belongs in the record of processing activities, and because analyses can be suitable for monitoring behaviour or performance, involve the works council early where one exists (in Germany § 87(1) no. 6 BetrVG). Recommendation: analyse at service level rather than person level, and define purpose and retention period up front.

  • Sources: directory, device management, proxy/DNS, purchasing
  • Decide on findings one by one, never blindly
  • Record discovery itself in the RoPA, involve the works council early
02

From finding to inventory record

Art. 3(3) and (4), Art. 6, Annex III AI Act

A finding is not yet an inventory record. For compliance and the business to work with it, each record needs a few mandatory fields: name, provider, purpose, run internally or bought as SaaS, role of the AI (decision-critical or supporting), affected areas and a responsible person. IT usually knows the provider, user count and integrations, but rarely the purpose – that comes from the business unit.

Important for classification later: one record per purpose, not per product. The same language model can be minimal-risk in marketing and high-risk in candidate selection (Annex III point 4). Anyone who just enters “ChatGPT” hides exactly the cases that matter.

  • One record per purpose
  • IT supplies provider, users, interfaces – the business supplies the purpose
  • Name a responsible person per system
03

Logs: six months, but not forever

Art. 12, Art. 26(6) AI Act · Art. 5(1)(e) GDPR

High-risk AI must technically allow automatic recording of events – that is a requirement on the provider (Art. 12). The deployer must keep the logs the system generates for at least six months, to the extent they are under its control and unless other law provides otherwise (Art. 26(6)). With SaaS the logs often sit with the provider; then access and retention belong in the contract.

At the same time AI logs often contain personal data, such as prompts with customer names. The GDPR requires storage limitation. Six months is therefore a floor, not an invitation to keep everything. Recommendation: a fixed period per system in the retention policy, log access only for named roles, and export into an existing SIEM instead of a separate log silo.

  • Floor of six months for high-risk AI (deployer)
  • SaaS: settle access and retention in the contract
  • Record period and access roles in the retention policy
04

Monitoring, incidents and the path to the provider

Art. 26(5), Art. 73 AI Act · Art. 33 GDPR

Deployers of high-risk AI monitor operation on the basis of the instructions for use. If they have reason to consider that use presents a risk, they inform the provider and the market surveillance authority without undue delay and suspend use; where they identify a serious incident, they first inform the provider immediately (Art. 26(5)). Reporting serious incidents to the authorities is the provider’s job (Art. 73).

For IT that means: the AI system needs monitoring like any other critical service, a documented way to switch it off must exist, and the incident process needs an extra “AI incident” branch next to the GDPR breach path. One incident can be both – for example when a model discloses personal data to unauthorised people.

  • Monitoring and alerting for AI services
  • Be able to suspend use technically
  • Incident process with AI and GDPR branches
05

When engineering becomes the provider

Art. 25(1) AI Act

Most organisations are deployers. For teams that build, three triggers turn a deployer into the provider of a high-risk system: putting their own name or trademark on a high-risk system already on the market, a substantial modification after which it remains high-risk, and a change of purpose that turns a non-high-risk AI system into a high-risk one (Art. 25(1)).

The third case is the tricky one: a general chat assistant is rebuilt internally to pre-sort job applications. Technically that is a new prompt and an interface to the applicant tracking system; legally it may be a new high-risk system with provider duties. Recommendation: make purpose changes a mandatory field in the change process before anything goes live.

  • Own branding on a high-risk system
  • Substantial modification
  • Purpose change towards high-risk
06

Integrations, access and co-determination

Art. 32 GDPR · § 90(1) no. 3 BetrVG

A compliance tool that sits beside the system landscape does not get maintained. Tasks belong where teams already work: the ticketing system, chat, the repository. Evidence often already lives in document libraries. And users should come from the identity provider, not a second user directory.

Selection therefore turns on the same questions as for any SaaS service: single sign-on, provisioning, role model, an API with documented limits, audit log, hosting and sub-processors. In Germany, an employer planning to use AI in work processes and workflows must inform the works council in good time (§ 90(1) no. 3 BetrVG) – that covers tooling around AI as well.

  • SSO and provisioning from the identity provider
  • Route tasks into ticketing and chat systems
  • Check hosting, sub-processors, audit log

RACI excerpt

Who does what? IT in AI governance

Excerpt from an AI governance RACI matrix with the tasks IT takes part in. The highlighted column shows the role of IT leadership. Each task has exactly one A. The assignment is a suggestion for a mid-sized organisation.

Scroll the table sideways →

Who does what? IT in AI governance
TaskManagementAI governance leadBusiness unit / system ownerIT lead / CTOData protection (DPO)Legal
Detect shadow AI and assess findingsCompleteness of the inventoryIACRC–
Keep the AI inventory technically currentBasis for all dutiesIARRC–
Classify risk and give reasonsArt. 5, 6, 50 AI ActIA/RCCCC
Configure per instructions for useArt. 26(1) AI Act–IAR––
Keep and delete logsArt. 26(6) AI Act · Art. 5 GDPR–IARC–
Monitor operation, suspend useArt. 26(5) AI ActIIARI–
Assess and report incidentsArt. 26(5) AI Act · Art. 33 GDPRARRRCC
Review changes of purposeArt. 25(1) AI ActIARRCC
Manage SSO, roles and accessArt. 32 GDPR–ICA/RI–
  • RResponsible – does the work
  • AAccountable – owns and decides
  • CConsulted – asked beforehand
  • IInformed – kept informed
  • IT is rarely accountable for an AI system: the business unit owns the purpose and therefore the risk. IT owns the technical platform – access, operations, logs.
  • The data protection column shows only C or I on decisions: DPOs inform, advise and monitor (Art. 39 GDPR) and should not take on tasks that create a conflict of interests (Art. 38(6) GDPR).
  • The full matrix with twelve tasks including the works council is in the AI governance framework.

Interactive

IT’s top tasks for the next 90 days

Twelve tasks in three phases. Tick off what is done – progress shows at the top. Expand any task to see why it matters.

0 of 12 tasks done

0%

Your progress is stored in this browser only, never transmitted.

Days 1–30

0/4

Visibility

  • Review OAuth consents in the directory for AI services
  • Analyse proxy or DNS logs for known AI domains
  • Turn findings into inventory records with the governance lead
  • Interim rule: new AI tools only via request

Days 31–60

0/4

Technical ground rules

  • Define a log concept for critical AI systems
  • Add an AI branch to the incident process
  • Make purpose change a mandatory change field
  • Check chatbot notices and content marking

Days 61–90

0/4

Operations and evidence

  • Rule out prohibited features in tools in use
  • Brief admins and power users technically
  • Set up single sign-on and roles for the governance tool
  • Schedule repeat discovery and inventory reconciliation

Order and time frame are a recommendation, not a legal requirement. Organisations with their own high-risk systems or in-house development usually need longer.

From IT’s point of view

Questions that reach IT – and good answers

Management, business units and purchasing typically put these questions to IT leads and CTOs. The answers are suggestions to adapt to your organisation.

“Can we use Copilot or ChatGPT in the company?”

The AI Act does not prohibit it. The purpose decides: for drafting and research the risk is usually minimal; for candidate selection or performance assessment a high-risk case can arise. GDPR, confidentiality and your own AI policy apply as well. IT’s answer: approve via request, with data rules and an inventory record.

AI policy template

“How many AI tools do we actually have?”

Honest answer: more than were reported. A reliable number comes from three sources – directory, device management and network logs – plus purchasing data. Count per purpose, not per product.

Build an AI register

“Do we have to store every prompt?”

No. The duty to keep logs for at least six months applies to deployers of high-risk AI and covers logs under their control. For all other systems you decide by purpose, security and GDPR – often storing less is better.

“Is our home-built assistant a high-risk system?”

That depends on its purpose, not its architecture. If it falls under an Annex III use case it is high-risk in principle unless an exception under Art. 6(3) applies. Whoever puts it into service under their own name is the provider.

Check Annex III

“What must the compliance tool do technically?”

At minimum: single sign-on, a role model, a tamper-evident audit log, export, a documented API and clear statements on hosting and sub-processors. Plus connections to ticketing and chat so tasks do not land in yet another inbox.

Selection criteria

“Who is responsible when an AI system makes mistakes?”

The business unit owns the system as system owner. IT monitors operation, can suspend use technically and informs the provider together with the governance lead. A process agreed in advance saves the argument when it counts.

Incident management

Deadlines

The dates IT should have in the calendar

These dates come from the AI Act and the Digital Omnibus. Next to each is what it means for IT.

  1. 2 Feb 2025

    Prohibited practices (Art. 5) and AI literacy (Art. 4)

    Rule out prohibited features in tools, brief technical user groups.

  2. 2 Aug 2025

    Duties for general-purpose AI models, governance, penalties

    Relevant if you provide your own foundation models; otherwise obtain provider information.

  3. 2 Aug 2026

    General application, transparency duties under Art. 50

    Notices in chatbots, labelling AI content in publishing.

  4. 2 Dec 2026

    Marking under Art. 50(2) for legacy systems

    Retrofit own generative systems placed on the market before 2 Aug 2026.

  5. 2 Dec 2027

    High-risk duties for Annex III systems

    Logs, monitoring, instructions for use, input data – the technology must be in place by then.

  6. 2 Aug 2028

    High-risk duties for Annex I systems

    AI as a safety component of regulated products, such as machinery or medical devices.

Status after the Digital Omnibus, Regulation (EU) 2026/1744, in force since 27 July 2026. All AI Act deadlines →

With SimpleAct

What SimpleAct covers for IT teams

SimpleAct keeps the AI inventory, risk classification, approvals and evidence in one place and connects to the systems you already run. The view on the right shows the AI inventory with sample data from the demo tenant.

  • AI Discovery

    Professional

    Analyses OAuth consents in Microsoft Entra ID, optionally the Intune app inventory and uploaded proxy or DNS logs (CSV), and matches findings against a vendor catalogue. No endpoint agent.

  • Single sign-on

    Professional

    Sign-in via OIDC, SAML or LDAP. SCIM provisioning from Microsoft Entra ID or Okta is part of the Enterprise plan.

  • REST API

    Enterprise

    API keys per application (X-API-Key header), OpenAPI specification, 1,000 requests per hour per key. Webhooks with a signed test delivery; business events are added step by step.

  • Integrations

    by plan

    Jira and ServiceNow for tickets, Microsoft Teams and Slack for notices, GitHub and GitLab for issues, SharePoint and Google Drive as evidence sources, MLflow for model versions.

  • Audit trail and roles

    all plans

    Audit trail as a hash chain with integrity check, roles Admin, Editor and Viewer, 2FA via TOTP. Hosted at Hetzner in Nuremberg, backups in Falkenstein.

Plan assignment per the pricing page. Starter: up to 5 AI systems and 3 users; from Professional unlimited AI systems and users. The ISO 27001 certificate covers the data centre, not SimpleAct. Sub-processors are Hetzner, Cloudflare (file storage), Stripe (payments), Brevo (email) and Anthropic (AI features only after opt-in; support chat see privacy policy).

app.simpleact.de

AI Inventory

3 systems registered

A
Search systems …
AllHighLimitedMinimalUnassessed
1 system not yet classifiedClassify now →
OAIHigh

AI-based Application Screening

OpenAI · HR & Recruiting

Dr. Anna Schmidt2026-07-12
MSLimited

Service Chatbot "Lia"

Microsoft · Customer Support

Thomas Müller2026-06-28
OAIMinimal

Marketing Copy Generator

OpenAI · Marketing

Lisa Weber2026-08-03
DLUnassessed

Website Translation Plugin

DeepL

Classify now →

FAQ

Frequently asked questions from IT teams and CTOs

Mainly three things: make all AI systems visible (including shadow AI), operate them as the provider instructs, and be able to prove how they are run. For high-risk AI the deployer duties in Art. 26 follow from 2 December 2027, such as monitoring, controlling input data and keeping logs for at least six months.
Not on its own. The duties fall on the organisation as deployer or provider; overall responsibility lies with management. The business unit usually owns the purpose and therefore the risk of a system. IT owns the technical implementation: platform, access, operations, logs and interfaces.
Deployers of high-risk AI must keep automatically generated logs for at least six months, to the extent the logs are under their control and unless other law provides otherwise (Art. 26(6) AI Act). For other systems the AI Act sets no period; purpose, security and GDPR storage limitation decide.
Shadow AI means AI services used in the organisation without being reported or approved – such as browser tools or apps connected via OAuth. Good sources are app permissions granted in the directory, the app list of managed devices, proxy or DNS logs and purchasing data. The analysis itself processes employee data and should be agreed with data protection and the works council.
Under Art. 25(1) AI Act, when it puts its own name or trademark on a high-risk system, substantially modifies one so that it remains high-risk, or changes the purpose of an AI system so that it becomes high-risk. The provider obligations then apply. Engineering teams should therefore review purpose changes in the change process.
Yes, as a deployer. Many technical requirements such as logging capability or robustness sit with the provider, but the deployer must use the system per its instructions, monitor it and keep logs under its control. With SaaS, log access, reporting paths and information duties therefore belong in the contract.
In Germany an employer planning to use AI in work processes and workflows must inform the works council in good time (§ 90(1) no. 3 BetrVG). A co-determination right can exist where a system is suitable for monitoring behaviour or performance (§ 87(1) no. 6 BetrVG). For high-risk AI at work, Art. 26(7) AI Act additionally requires informing workers’ representatives and affected workers beforehand.
Yes, from the Professional plan via OIDC, SAML and LDAP. SCIM provisioning from Microsoft Entra ID or Okta, plus the REST API, are part of the Enterprise plan. Starter covers up to 5 AI systems and 3 users, Professional unlimited AI systems and users.
SimpleAct runs at Hetzner in Nuremberg with backups in Falkenstein; application and database stay in Germany. The ISO 27001 certificate covers the data centre. Sub-processors are Hetzner, Cloudflare (file storage), Stripe (payments), Brevo (email) and Anthropic (AI features only after opt-in; support chat see privacy policy). The SLA states a target availability of 99.5% per calendar month.
Yes, 7 days free on the Starter scope (credit card required, cancel during the trial). For integrations, single sign-on or AI Discovery a demo is also worthwhile, in which we walk through your system landscape.

More questions? We're happy to help. Send email · Try 7 days (credit card required)

Sources and status

Last reviewed on · SimpleAct editorial team

This page is for orientation and is not legal advice. The text published in the Official Journal of the EU is authoritative. RACI assignment, checklist and time frames are recommendations, not legal requirements.

Changes to this page

  • 6 October 2026 – Rebuilt as a guide: duties table with articles, six topics (shadow AI, logs, incidents, provider role, integrations), RACI excerpt, interactive 90-day checklist, deadlines after the Digital Omnibus, ten FAQs.
  • 18 May 2026 – First version as a role page for IT teams.

Make shadow AI visible, keep evidence in one place

SimpleAct connects the AI inventory, AI Discovery, approvals and the audit trail with your identity provider and ticketing systems.