# Responsible AI in 2026: EU AI Act, NIST and ISO 42001

URL: https://sdclabs.com/responsible-ai-governance/

> Less and less. The EU AI Act has applied since 2 August 2026, with exceptions, and only 11% of firms lack responsible AI policies. What applies when, and how to start.

# Is responsible AI still optional in 2026?

[Mykhailo Poliarush](https://sdclabs.com/author/poliarush/) 2026-08-06Updated 2026-11-176 min read

In 2026, responsible AI is no longer a set of principles on a slide. It is turning into obligations, standards and evidence: the EU AI Act became applicable on 2 August 2026, with some exceptions, and the share of organizations with no responsible AI policies fell from 24% in 2024 to 11% in 2025, according to the [Stanford AI Index 2026](https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf) . The hard part now is practice: knowing what you run, testing it and proving it.

This article summarizes public sources for engineering and product teams. It is not legal advice.

## What does the EU AI Act require, and when?

The Act applies in stages. According to the [European Commission](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai) , it entered into force on 1 August 2024 and became applicable on 2 August 2026, with exceptions:

- Prohibited AI practices and AI literacy obligations have applied since 2 February 2025.
- Governance rules and obligations for general-purpose AI models have applied since 2 August 2025.
- Rules for high-risk use cases in sensitive areas (Annex III), such as employment, education and critical infrastructure, apply from 2 December 2027.
- Rules for high-risk AI embedded in regulated products (Annex I), such as lifts or toys, apply from 2 August 2028.

The later high-risk dates come from the AI Omnibus amendment, which entered into force on 27 July 2026 after a political agreement on 7 May 2026 ([European Commission](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai) ). From 2 August 2026, the EU AI Office and Member State authorities are responsible for supervising and enforcing the Act. Which category your system falls into is a legal question: classify with your counsel, not with a blog post. For the wider engineering picture, see our [state of AI engineering in 2026](https://sdclabs.com/state-of-ai-engineering-2026/) .

## Which frameworks are teams using: NIST AI RMF or ISO/IEC 42001?

Both, and they do different jobs. The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) , released on January 26, 2023, is voluntary guidance; its core has four functions, Govern, Map, Measure and Manage, with Govern cutting across the other three ([NIST AI 100-1](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf) ). NIST added a Generative AI Profile (NIST-AI-600-1) on July 26, 2024. [ISO/IEC 42001](https://www.iso.org/standard/42001) , published in December 2023, specifies requirements for an AI management system that an organization establishes, maintains and improves; ISO calls it the world's first AI management system standard.

Both now shape real decisions. In the AI Index and McKinsey survey of business leaders, 36% cited ISO/IEC 42001 and 33% cited the NIST AI RMF as influences on their responsible AI practices in 2025, the first year these options were offered ([AI Index 2026](https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf) ). GDPR is still the most cited at 60%, down from 65%, and the EU AI Act is at 43% ([AI Index 2026](https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf) ).

Regulations and standards influencing responsible AI practices, 2025

**EU GDPR**

60%

**EU AI Act**

43%

**ISO/IEC 42001**

36%

**NIST AI RMF**

33%

**US Executive Order on AI**

21%

**OECD AI Principles**

16%

**None**

12%

Source: [Stanford HAI, AI Index 2026, Figure 3.3.11 (AI Index and McKinsey survey of business leaders, 2025)](https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf)

A practical reading: NIST gives you a shared vocabulary and a structure for managing AI risk, ISO/IEC 42001 gives you a management system to establish, maintain and continually improve, and the EU AI Act sets legal obligations for AI systems on the EU market. One set of engineering controls and evidence can serve all three, if you design it that way.

## Are AI incidents actually rising?

Yes, by every count available. The AI Incident Database recorded 362 AI incidents in 2025, up from 233 in 2024, and the annual number stayed under 100 until 2022 ([AI Index 2026](https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf) ). The database is curated by human editors and skews toward English-language media, so it undercounts. The OECD's automated monitor, which uses a wider net, recorded a monthly peak of 435 incidents in January 2026 ([AI Index 2026](https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf) ).

Model behavior is part of the problem. On a new accuracy benchmark reported in the AI Index, hallucination rates across 26 top models ranged from 22% to 94% ([AI Index 2026](https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf) ). Transparency moved the wrong way too: the average Foundation Model Transparency Index score fell from 58 in 2024 to 40 in 2025 ([AI Index 2026](https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf) ), with gaps in disclosure about training data and post-deployment impact. If you build on third-party models, you inherit risks you cannot fully inspect, which makes your own testing more important.

## What are the main challenges in putting responsible AI into practice?

Knowledge, budget and uncertainty, in that order. In the 2025 AI Index and McKinsey survey, 59% of organizations named knowledge and training gaps as a main obstacle to implementing responsible AI, up from 51% in 2024, followed by resource or budget constraints at 48% and regulatory uncertainty at 41% ([AI Index 2026](https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf) ). Technical limitations rose to 38%, from 32% ([AI Index 2026](https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf) ).

Main obstacles to implementing responsible AI, 2025

**Knowledge and training gaps**

59%

**Resource or budget constraints**

48%

**Regulatory uncertainty**

41%

**Technical limitations**

38%

**Organizational resistance**

26%

**Lack of executive support**

14%

Source: [Stanford HAI, AI Index 2026, Figure 3.3.9 (AI Index and McKinsey survey of business leaders, 2025)](https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf)

Agents raise the stakes. When the same survey asked about scaling agentic AI, 62% named security and risk concerns as the primary obstacle ([AI Index 2026](https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf) ). An agent that can act on systems turns a bad output into a bad action.

Four challenges show up in most programs:

1. **No inventory.** Teams cannot assess risk for AI they do not know they run, including AI features inside SaaS tools.
1. **Policies without tests.** A fairness or safety policy that no test suite checks does not change what ships.
1. **Evidence that goes stale.** Documentation written once falls out of date with the next model or prompt change.
1. **Unclear ownership.** Someone must be able to approve, override and switch off each system.

## How mature are responsible AI programs today?

Early, but moving. The AI Index and McKinsey survey rates maturity on a four-point scale, and the 2025 global average was 2.3, which means most organizations are still integrating responsible AI practices rather than running them fully ([AI Index 2026](https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf) ). Ownership is getting clearer: organizations naming a dedicated AI governance role as the primary owner rose from 14% to 17%, and those with no designated owner fell from 9% to 5% ([AI Index 2026](https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf) ). These are self-reported survey answers from business leaders, and the survey does not include China.

## What does good look like, and how do you start?

Responsible AI that works looks like ordinary engineering discipline applied to AI:

1. **Build an inventory.** Every AI system you build or buy, with purpose, owner, data, users and a preliminary risk level for legal review.
1. **Assess risk per system.** Use the NIST functions (govern, map, measure, manage) to structure the work, and record decisions.
1. **Turn policy into tests.** Write repeatable tests for bias, harmful output, prompt injection and data leakage, and run them before each release.
1. **Keep humans able to intervene.** Define approval points, override and shutdown controls, and log who decided what.
1. **Monitor in production.** Track incidents, drift and user complaints, and feed them back into tests.
1. **Keep documentation alive.** Generate system cards and test reports from the pipeline, so evidence updates with each change.

Start with the inventory. It is cheap, it surfaces surprises, and every later step depends on it.

## How can SDCLabs help with responsible AI?

SDCLabs approaches responsible AI as engineers: AI agents help draft inventories, risk assessments and test cases, senior engineers review and decide, and automated tests check every change. Our founders created the open-source testing frameworks Codeception and CodeceptJS, so turning a policy into a test suite that runs on every release is familiar ground. We are not lawyers, and we work alongside your legal counsel on classification and obligations.

If you need to know what AI you run and whether you can prove it is tested, see our [responsible AI services](https://sdclabs.com/services/responsible-ai/) or [contact us](https://sdclabs.com/contact/) . Related: our [AI testing services](https://sdclabs.com/services/ai-testing/) and [AI agent benchmarking](https://sdclabs.com/services/ai-agent-benchmarking/) .

Sources checked on 2026-11-17.

## Frequently asked questions

### When do the EU AI Act high-risk rules apply?

According to the [European Commission](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai) , rules for Annex III high-risk use cases apply from 2 December 2027 and for high-risk AI in regulated products (Annex I) from 2 August 2028. This is a summary, not legal advice.

### What is the difference between the NIST AI RMF and ISO/IEC 42001?

The [NIST AI RMF](https://www.nist.gov/itl/ai-risk-management-framework) is voluntary guidance built on four functions: govern, map, measure and manage. [ISO/IEC 42001](https://www.iso.org/standard/42001) specifies requirements for an AI management system that an organization establishes, maintains and improves.

### Are AI incidents increasing?

Yes: the AI Incident Database recorded 362 incidents in 2025, up from 233 in 2024, as reported in the [AI Index 2026](https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf) , and the human-curated database likely undercounts.

### What is the biggest obstacle to responsible AI?

Knowledge and training gaps, named by 59% of organizations in the 2025 AI Index and McKinsey survey, followed by budget constraints (48%) and regulatory uncertainty (41%), per the [AI Index 2026](https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf) .

[#responsible-ai](https://sdclabs.com/tag/responsible-ai/) [#ai-governance](https://sdclabs.com/tag/ai-governance/) [#eu-ai-act](https://sdclabs.com/tag/eu-ai-act/) [#ai-risk](https://sdclabs.com/tag/ai-risk/)

[Mykhailo PoliarushFounder and CEO](https://sdclabs.com/author/poliarush/)

## Related posts

[
### Generative AI fails after the demo, not before it

88% of surveyed organizations use AI, but only 13% of CIOs report significant value. Generative AI stalls between pilot and production, on accuracy, cost and fit.

AI services · 2026-11-09

](https://sdclabs.com/generative-ai-production/) [
### AI agents can take the first page, but humans still own the outage

AI agents are entering IT operations fast, but supervision lags. High-impact outages cost a median $2 million per hour. Here are the challenges and how to start.

AI services · 2026-10-13

](https://sdclabs.com/ai-agents-it-operations/) [
### Only 18% of engineering orgs measure AI's impact: a rollout playbook

66% of engineering orgs have adopted AI tools, but only 18% measure the impact. A data-backed guide to rolling out AI with training, governance and metrics.

AI services · 2026-08-05

](https://sdclabs.com/ai-enabled-engineering-rollout/)

## Get an estimate for your product.

[Get an estimate](https://sdclabs.com/contact/) [Book a call](https://sdclabs.com/contact/#call)
