There is a newer version of the record available.

Published August 14, 2026 | Version v1

A 10-person SME can meet EU AI Act standards like a Fortune 500 — without needing a Fortune 500 legal team

Authors/Creators

Description

Abstract

The EU AI Act establishes an important framework for trustworthy, accountable, and human-centric artificial intelligence. A continuing practical challenge, however, is ensuring that the cost and complexity of implementing regulatory obligations do not fall disproportionately on smaller European companies.

Large organizations can distribute compliance responsibilities across dedicated legal, engineering, cybersecurity, privacy, governance, risk, and audit teams. A company employing only ten people may face many of the same underlying questions—what data may be used, for what purpose, under which authorization, in which jurisdiction, by which AI system, under what human oversight, and with what evidence—while relying on the same founders and engineers who must also build, test, sell, support, and improve the product.

This paper examines a conceptual alternative: shifting selected recurring compliance functions away from separately engineered application-level processes toward reusable infrastructure-level compliance capabilities.

At a high level, such infrastructure could allow relevant purpose, jurisdiction, authorization, privacy, accountability, and governance conditions to accompany a digital or AI operation in machine-verifiable form. Instead of requiring each SME to independently recreate substantially similar compliance mechanisms for every application, cloud provider, AI model, customer, partner, or Member State, common infrastructure could evaluate selected conditions consistently at the point where an AI-generated operation is about to produce an external consequence.

The conceptual framework considers four broad functions:

  1. Privacy-preserving identity and authority — establishing that an entity or system possesses appropriate authority without unnecessarily exposing reusable identity information.

  2. Jurisdiction-aware governance — allowing applicable territorial, organizational, contractual, or regulatory conditions to accompany an operation.

  3. Execution-time compliance evaluation — evaluating relevant governance conditions when an AI-generated action is about to become operationally effective, rather than relying exclusively on policy statements, pre-deployment documentation, or post-event auditing.

  4. Automatically generated governance evidence — producing machine-verifiable evidence that relevant conditions were evaluated before or at the point of execution.

These concepts are represented broadly through the VI, CJT, and related execution-time governance models. The purpose is not to require adoption of those particular labels or architectures, but to explore the broader possibility that compliance functions could increasingly become reusable properties of digital infrastructure.

A practical ten-person-company example

Consider a hypothetical European startup called MedAssist EU, employing ten people and developing an AI-assisted service for hospitals.

The company has:

  • two founders;

  • five software and AI engineers;

  • one product designer;

  • one sales and customer-support employee; and

  • one person handling administration, finance, and regulatory coordination.

MedAssist initially serves a hospital in one EU Member State. Its AI system assists authorized hospital personnel by analysing permitted information and producing recommendations or administrative actions.

As the company expands, it signs customers in several Member States and begins using multiple cloud environments, AI models, databases, external APIs, and healthcare software platforms.

Before an AI-generated operation is executed, several questions may become relevant:

  • Is this user or AI agent authorized to request the operation?

  • For what declared purpose is the information being processed?

  • Is the requested use compatible with that purpose?

  • Which jurisdiction and organizational rules apply?

  • Is the selected AI model permitted for this operation?

  • May the result be transmitted to this destination?

  • Has an authorization been revoked?

  • Is human approval required?

  • Has the required governance decision been recorded?

  • Can the company later demonstrate what conditions were evaluated before the operation occurred?

Without reusable infrastructure, MedAssist may need to implement parts of these controls separately across its application code, cloud policies, database permissions, workflow engines, customer-specific integrations, logging systems, contractual controls, and audit processes.

When a second AI model is introduced, another hospital platform is integrated, or a new jurisdiction is entered, portions of that work may need to be repeated.

For a ten-person company, this is not merely a legal issue. It becomes an engineering-capacity issue.

Now consider an infrastructure-level alternative.

An AI agent proposes an operation:

Candidate operation: send a patient-related AI recommendation to Hospital B's clinical workflow system.

Before that operation becomes externally effective, common infrastructure could evaluate a machine-readable governance context containing, for example:

Purpose: clinical decision support
Authorized entity: Hospital B clinician
Jurisdiction: applicable EU and national environment
Permitted AI model: Model X, approved configuration
Destination: Hospital B authorized clinical system
Human-oversight condition: approval required
Authorization status: active
Privacy condition: permitted for declared purpose
Governance epoch: current policy version

The infrastructure would not determine European law by itself. Nor would it replace legal interpretation.

Instead, once the relevant organizational and regulatory requirements had been translated into applicable machine-verifiable conditions, the infrastructure could help enforce those conditions consistently.

A simplified operational flow could therefore become:

AI proposes action → action remains non-effective → relevant identity, purpose, jurisdiction, authorization, destination, and governance conditions are evaluated → compliant action receives execution authority → permitted destination verifies that authority → action becomes effective → evidence of the governance decision is retained.

If a required condition is absent, expired, revoked, inconsistent, or unverifiable, the infrastructure could prevent the action from proceeding or require escalation.

The important economic distinction is that MedAssist would not necessarily need to build the entire mechanism itself.

Just as a ten-person company does not normally design its own CPU, TLS protocol, cloud networking stack, payment network, or database engine, it may eventually be unnecessary for every small AI company to independently construct every recurring compliance-enforcement mechanism.

A cloud provider, operating system, trusted execution environment, standards-based service, telecommunications platform, sectoral infrastructure provider, or interoperable European compliance layer could potentially provide portions of that capability.

The SME would still remain responsible for understanding its obligations, configuring appropriate policies, governing its AI system, and satisfying applicable law. Infrastructure would not eliminate responsibility.

It could, however, reduce the amount of duplicative technical implementation required to make those responsibilities operational.

Scaling the same infrastructure

The value becomes clearer when the same ten-person company expands.

Today:

France → Hospital A → AI Model X

Six months later:

Germany → Hospital B → AI Model X

Later:

Italy → Hospital C → AI Model Y

And eventually:

Multiple Member States → Multiple AI models → Multiple cloud providers → Multiple customers

Under a predominantly application-specific approach, each expansion can introduce new authorization logic, policy mappings, logging mechanisms, data-flow controls, integrations, testing requirements, and compliance evidence.

Under a reusable infrastructure model, substantial portions of the underlying mechanism could remain common while the applicable policies, permissions, jurisdictions, purposes, and destinations change.

This distinction may become especially important as AI systems become more agentic.

Traditional software frequently waits for explicit user instructions at individual stages. AI agents may instead perform multi-step operations involving retrieval, reasoning, tool invocation, API access, data transformation, inter-agent communication, and external execution.

Compliance therefore increasingly concerns not only what an AI model says, but also what an AI-enabled system is authorized to cause.

Infrastructure-level governance could provide a common enforcement point between machine-generated decision-making and externally effective consequences.

Implementation neutrality

The proposal is intentionally implementation-neutral.

It does not prescribe a particular:

  • token structure;

  • cryptographic construction;

  • identity scheme;

  • trusted hardware implementation;

  • validation algorithm;

  • policy language;

  • ledger;

  • audit-record structure;

  • cloud architecture;

  • operating system;

  • AI model;

  • communications protocol; or

  • enforcement technology.

Different industries and risk levels may require different technical implementations.

A low-risk SME productivity tool may need comparatively lightweight controls. A healthcare, financial, critical-infrastructure, robotics, defence, or high-risk AI deployment may require substantially stronger mechanisms.

The architectural question is therefore broader than any single implementation:

Can recurring governance requirements be represented and enforced through interoperable infrastructure rather than repeatedly reconstructed inside every individual AI application?

Nor is the framework intended to replace the EU AI Act, GDPR, the Digital Services Act, the Digital Markets Act, cybersecurity legislation, sector-specific regulation, or national law.

Technical infrastructure cannot determine regulatory obligations independently of law, organizational governance, and human decision-making.

Its purpose would instead be to help translate already-defined obligations and governance decisions into consistently enforceable technical conditions.

Implications for European SMEs

For SMEs, the potential benefits are straightforward:

Less repetitive compliance engineering.
Common mechanisms could reduce the need to recreate similar authorization, purpose, jurisdiction, logging, and evidence functions in each product.

More predictable implementation costs.
Infrastructure-level capabilities may make it easier to estimate the technical effort associated with regulatory expansion.

Easier cross-border growth.
A common technical framework could allow jurisdiction-specific conditions to change without rebuilding the entire underlying enforcement architecture.

Improved interoperability.
Standardized machine-verifiable governance conditions could potentially travel across applications, platforms, AI agents, cloud environments, and organizational boundaries.

Stronger compliance evidence.
Execution-time records could supplement traditional documentation by demonstrating what governance conditions were actually evaluated when an AI-enabled action occurred.

More engineering capacity for innovation.
Every engineering hour spent rebuilding a generic compliance mechanism is an hour unavailable for improving the underlying product.

The objective is therefore not deregulation.

It is compliance scalability.

A broader European policy question

Europe faces an important strategic question.

Should sophisticated compliance infrastructure remain something that primarily large organizations can afford to design internally?

Or can parts of that capability eventually become standardized infrastructure accessible to a ten-person company as readily as they are to a multinational corporation?

A large technology company may employ hundreds or thousands of specialists across legal, regulatory, cybersecurity, privacy, AI safety, governance, and compliance engineering.

A ten-person European startup cannot reproduce that organizational structure.

But it may not need to reproduce the organization if parts of the corresponding technical capability can be embedded into infrastructure.

This leads to a broader principle:

Large organizations can scale compliance by adding people.
Small organizations need a way to scale compliance through infrastructure.

The policy objective should therefore not be to weaken Europe's regulatory protections for smaller companies. Doing so could undermine the purpose of the regulatory framework and reduce public trust.

A more sustainable research direction is to investigate whether strong protections can be made less expensive to implement correctly.

Reusable infrastructure-level compliance could contribute to that objective.

It could allow sophisticated governance capabilities—purpose limitation, scoped authorization, jurisdiction awareness, revocation, execution-time validation, accountability evidence, and related controls—to become shared technical building blocks rather than bespoke systems repeatedly constructed by individual companies.

This could gradually create a more level competitive environment while preserving the underlying protections established by European law.

The principle is simple:

Ten people should not need the compliance engineering capacity of a ten-thousand-person company merely to compete responsibly in the European AI economy.

If Europe can transform selected recurring compliance functions from an organizational luxury into accessible digital infrastructure, regulation and innovation need not operate as opposing forces.

The same infrastructure that strengthens accountability could also reduce compliance asymmetry, lower barriers to responsible market entry, and give smaller European companies more capacity to build, compete, and innovate.

The objective is not less compliance. It is compliance that scales.

Files

Technical Disclosure 1 - EU AI ACT .zip

Files (509.3 kB)

Name Size
md5:03658ef1ffe4e639a09386e075b088a4
509.3 kB Preview Download