The 2026 EU AI Act technical audit for CTOs

Topic

Sovereign AI

Kind

Perspective

Author

Kjeld Oostra

Read time

7 min read

/

/

The 2026 EU AI Act technical audit for CTOs

/

/

The 2026 EU AI Act technical audit for CTOs

A practical engineering checklist for organisations building and deploying AI in Europe

The EU AI Act is now an engineering concern.

On 2 August 2026, the European Commission and national authorities began enforcing several parts of the regulation. Transparency requirements for certain AI systems are now applicable, obligations concerning general-purpose AI are in force, and prohibited AI practices are already restricted.

Following the 2026 AI Omnibus, some high-risk requirements have later application dates.

For CTOs, the sensible response is not to wait for every deadline.

It is to make AI systems governable now.

This checklist provides a practical starting point.

It is an engineering framework, not legal advice.

1. Create an AI system inventory

Start with visibility.

You cannot govern systems you do not know exist.

Create an inventory covering:

  • internally developed AI applications
  • third-party AI products
  • embedded AI functionality
  • employee-facing copilots
  • customer-facing AI
  • automated decision systems
  • models accessed through APIs
  • locally hosted models
  • experimental AI applications
  • agentic workflows

For each system, record its owner, purpose, users, model providers, data sources, integrations and deployment environment.

Do this before buying another governance platform.

A spreadsheet with accurate information is more useful than sophisticated governance software containing incomplete information.

2. Determine your role

The AI Act distinguishes between different actors in the AI value chain.

Your obligations can depend on whether you are acting as a provider, deployer, importer, distributor or provider of a general-purpose AI model.

Do not assume that purchasing a model API automatically determines your role for the complete system.

An organisation can integrate third-party models into an AI application and have responsibilities relating to the resulting system.

Document your role for each material AI implementation and have the classification reviewed where necessary.

3. Classify the use case

The AI Act follows a risk-based approach.

For each AI system, determine whether the use case may fall into:

  • prohibited practices
  • high-risk categories
  • systems subject to transparency obligations
  • minimal or lower-risk applications
  • general-purpose AI requirements, where relevant

Classification should happen at the use-case level.

The same underlying model can support a low-risk marketing application and a much more sensitive employment application.

Govern the system according to what it does.

4. Map every data boundary

For each AI system, draw the actual information flow.

Start with the user.

Then map:

  • User
  • → application
  • → orchestration layer
  • → retrieval system
  • → model
  • → tools
  • → databases
  • → logs
  • → monitoring systems

For every boundary, answer:

What data crosses it?

Is personal data included?

Is confidential information included?

Where is processing performed?

Is information retained?

For how long?

Who can access it?

Is it used for model training or improvement?

Which subprocessors are involved?

Can the organisation delete it?

A vendor's "enterprise" label is not a data-flow diagram.

5. Document every model

Maintain a model registry.

For every production model, record:

  • provider
  • model and version
  • deployment location
  • intended use
  • known limitations
  • input data categories
  • output types
  • retention settings
  • fallback model
  • evaluation results
  • date introduced
  • responsible owner

When a provider silently changes a model underneath your application, behaviour can change.

Version awareness therefore matters operationally as well as for governance.

6. Build traceability into execution

Agentic systems introduce another problem.

An AI system can now perform a sequence of actions rather than simply produce an answer.

Logs should make it possible to reconstruct important executions.

Depending on the use case, record:

  • user request
  • system instructions
  • retrieved context
  • model used
  • tools called
  • relevant decisions
  • human approvals
  • resulting actions
  • errors and retries
  • final output

Be deliberate about what you log.

Logging everything can itself create privacy and security problems.

The objective is meaningful traceability, not indiscriminate data collection.

7. Define human oversight technically

"Human oversight" needs an implementation.

Create explicit action classes.

For example:

Read

The AI can retrieve information.

Recommend

The AI can propose an action but cannot execute it.

Prepare

The AI can prepare an artefact or transaction for approval.

Execute with approval

The AI acts after a human confirms.

Execute autonomously

The AI can perform a predefined action within strict permissions.

Prohibited

The AI cannot perform the action.

Then map tools and workflows to those classes.

This turns an abstract governance principle into access control.

8. Apply least privilege to agents

Do not give an AI agent access to an entire enterprise system because it occasionally needs one function.

Create narrow tools.

Instead of:

database_access()

provide operations such as:

get_customer_orders(customer_id)

or:

create_draft_quote(account_id)

The narrower the tool, the easier it is to control what an agent can do.

The same principle applies to file systems, email, CRM, ERP and code execution.

Treat AI agents like software identities with permissions.

9. Separate reasoning from execution

An LLM generating an instruction and software executing that instruction should be separate operations.

This creates a control point.

Before execution, the system can validate:

  • permissions
  • parameters
  • data types
  • business rules
  • transaction limits
  • approval requirements

Never assume that because a model produced syntactically valid output, the requested action is safe.

Models reason probabilistically.

Business systems execute deterministically.

Keep the boundary clear.

10. Test the system, not only the model

Model benchmarks tell you surprisingly little about the safety of an enterprise application.

Test the complete system.

Include:

  • incorrect inputs
  • prompt injection
  • missing data
  • conflicting instructions
  • excessive permissions
  • malformed tool calls
  • unavailable models
  • unavailable databases
  • hallucinated identifiers
  • unusual user behaviour
  • attempts to retrieve unauthorised information

A strong model inside a poorly designed system remains a poorly designed system.

11. Address transparency in the interface

As of August 2026, EU AI Act transparency requirements apply to certain AI systems and AI-generated or manipulated content.

Review customer-facing and employee-facing AI interfaces.

Users may need to understand when they are interacting with AI.

The technical implementation can include interface notices, metadata, machine-readable markings or other mechanisms depending on the application and applicable requirements.

Do not bury this exclusively in terms and conditions.

Design transparency into the product.

12. Establish AI literacy measures

AI literacy requirements have applied since February 2025, with supervision and enforcement powers applying from August 2026.

Training should match people's actual responsibilities.

A developer building an agent needs different knowledge from an employee using an AI writing assistant.

Useful training areas include:

  • capabilities and limitations
  • confidential information
  • personal data
  • hallucinations
  • prompt injection
  • human verification
  • approved tools
  • escalation procedures

Keep evidence of what measures were taken.

13. Review your general-purpose AI dependencies

General-purpose AI rules became applicable to relevant providers in August 2025, with Commission enforcement powers applying from August 2026.

Even if your organisation is primarily a downstream user, understand the models underneath your products.

Ask providers for appropriate information about capabilities, limitations, documentation and relevant compliance measures.

Also consider concentration risk.

If one model provider disappearing tomorrow would break every AI application in the company, you have an architectural dependency worth addressing.

14. Prepare for high-risk requirements before their deadline

Following the 2026 AI Omnibus, requirements for high-risk AI systems listed under Annex III apply from 2 December 2027.

Rules for high-risk systems embedded into regulated products apply from 2 August 2028.

Those dates should not become implementation start dates.

Systems in areas such as employment, education, critical infrastructure, biometrics and certain essential services may require considerably more governance than ordinary enterprise AI applications.

Architecture decisions made today can determine how difficult compliance becomes later.

15. Create an AI change process

AI systems change unusually quickly.

Models are updated.

Prompts change.

Knowledge bases grow.

New tools are connected.

Agents receive additional permissions.

A system that was reviewed six months ago may therefore be materially different today.

Define which changes trigger reassessment.

For example:

  • new model
  • new data source
  • new user group
  • new automated action
  • new geography
  • new sensitive data category
  • significant prompt or workflow change

Governance should follow the system throughout its lifecycle.

The CTO checklist

If you want a shorter version, start here.

Can you answer these questions for every production AI system?

What does it do?

Who owns it?

Which models does it use?

What data does it process?

Where does that data go?

What information is retained?

Which systems can it access?

Which actions can it perform?

Where is human approval required?

Can you reconstruct important actions?

How was the system tested?

How are changes controlled?

What happens when the model fails?

Can the model provider be replaced?

Who is responsible when something goes wrong?

If several answers are unclear, start there.

Compliance and good architecture are converging

The interesting consequence of the EU AI Act is that many governance requirements point towards practices strong engineering teams should want anyway.

Clear ownership.

Controlled permissions.

Traceable execution.

Documented systems.

Model evaluation.

Human oversight.

Security boundaries.

Change management.

These are useful properties whether a regulator ever asks to see them or not.

At WeAreBrain, we increasingly design enterprise AI around this assumption.

AI should be powerful enough to change how an organisation works.

Its architecture should still make it possible to understand, control and replace the technology underneath it.

That is the standard enterprise AI now needs.

Avatar-1
Avatar-2
Avatar-4
Avatar-4
Avatar-3

Fresh insights, straight to your inbox.

Engineering breakdowns, sovereign-AI perspectives and practical guides from the people building enterprise AI. One email when there is something worth your time.