AI Resource Library

Claude Prompting Best Practices↗(opens in a new tab)

Anthropic/Documentation/Continuously updated/45 min

Anthropic's living guide to prompt construction — clarity and directness, worked examples, XML structuring, role assignment, long-context placement, and the prompting patterns specific to tool use and agentic loops. Unlike most prompting content it is maintained against current models rather than frozen at publication, and the agentic sections are genuinely hard to find elsewhere. The practical companion to the Google whitepaper's theory.

• Prompt engineering
• System prompts
• Agentic prompting

Deep Dive into LLMs like ChatGPT↗(opens in a new tab)

Andrej Karpathy/Video/2025-02/3h 31m

The full pipeline — tokenisation, pretraining, supervised fine-tuning, RLHF, reasoning models — explained end to end without hand-waving. The payoff is diagnostic: once you understand where a model's knowledge comes from and how it was shaped, hallucination, sycophancy and brittleness stop being mysteries and become predictable failure modes you can design around. Still the best long-form technical grounding available for non-researchers.

• Pretraining
• RLHF
• Hallucination
• Model behaviour

How I Use LLMs↗(opens in a new tab)

Andrej Karpathy/Video/2025-02/2h 11m

A founding OpenAI researcher walks through how he personally uses LLMs day to day: model selection, thinking modes, tool use, file uploads, voice, code interpreters, and where each breaks down. It is the best available answer to "what does fluent, everyday AI use actually look like?" — the gap that separates organisations reporting productivity gains from those reporting none. Chaptered, so it skims well.

• Practical use
• Tooling
• Workflows

LLM Visualization↗(opens in a new tab)

Brendan Bycroft/Interactive tool/Ongoing/15-30 min

A 3D, step-by-step walkthrough of a GPT-style model executing a single forward pass — every embedding, attention head and matrix multiply rendered and animated. Nothing else makes the transformer architecture as tangible this quickly, and it works for audiences who bounce off equations. Use it to end the "black box" conversation in a workshop.

• Transformers
• Attention
• Visualisation

Large Language Models Explained Briefly↗(opens in a new tab)

3Blue1Brown (Grant Sanderson)/Video/2024-11/8 min

The single most efficient explanation of what a large language model actually is, assuming zero prior knowledge. Sanderson builds the intuition visually: an LLM is a next-word predictor whose behaviour emerges from hundreds of billions of parameters tuned on enormous text corpora, not a database of facts or a reasoning engine in the human sense. Eight minutes here prevents a year of category errors in strategy conversations.

• LLM basics
• Transformers
• Explainer

Prompt Engineering (Whitepaper)↗(opens in a new tab)

Google (Lee Boonstra) via Kaggle/Whitepaper/2024-11, last updated 2026-06/~2 hours

The most complete and most recently maintained vendor-neutral treatment of prompting technique: zero- and few-shot, chain-of-thought, ReAct, step-back prompting, structured output, and the sampling parameters (temperature, top-k, top-p) most people set by superstition. Written by a Google engineer but almost entirely model-agnostic. Treat it as the reference manual rather than a read-through.

• Prompt engineering
• Chain-of-thought
• Structured output

AI Engineering: Building Applications with Foundation Models↗(opens in a new tab)

Chip Huyen (O'Reilly)/Book/2024-12/~20 hours

The one book-length treatment that connects evaluation, RAG, fine-tuning, agents, and inference cost/latency engineering into a single coherent discipline, written by a practitioner rather than a vendor. Its chapters on evaluation methodology and the build-versus-buy-versus-fine-tune decision are the most rigorous available anywhere. The link goes to the author's open companion repository — chapter notes, case studies and resources are free; the book itself is paid.

Book is a paid purchase

• Evaluation
• RAG
• Fine-tuning
• Inference optimisation

Agentic AI↗(opens in a new tab)

DeepLearning.AI (Andrew Ng)/Course/2025-09/~10 hours

Ng's framework-agnostic treatment of the four core agentic patterns — reflection, tool use, planning and multi-agent collaboration — with evaluation and error analysis woven through rather than bolted on. It supersedes his earlier framework-specific agent courses and is the best structured on-ramp for architects who need vocabulary and judgement rather than a particular SDK. Free to audit.

• Agent patterns
• Reflection
• Tool use
• Planning

Elements of AI↗(opens in a new tab)

University of Helsinki & MinnaLearn/Course/Ongoing since 2018/~30 hours

A free, no-code AI course from a European public university, available in every EU language and taken by over a million people. Its real value in 2026 is institutional: it is the most defensible vendor-neutral option for organisations building evidence of workforce AI literacy under Article 4 of the EU AI Act. Note that no course confers automatic presumption of compliance — the Commission is explicit about this.

• AI literacy
• EU AI Act Article 4
• Non-technical

Hugging Face AI Agents Course↗(opens in a new tab)

Hugging Face/Course/Ongoing, actively maintained/~15-20 hours

The strongest free hands-on agent course: builds the same agent across smolagents, LangGraph and LlamaIndex so you learn the concepts rather than one vendor's abstractions, then finishes with a capstone evaluated against the GAIA benchmark. Bonus units cover fine-tuning and observability. For engineers who learn by shipping rather than reading.

• Hands-on
• smolagents
• LangGraph
• LlamaIndex
• Benchmarks

AI Index Report 2026↗(opens in a new tab)

Stanford Institute for Human-Centered AI/Report/2026-04/Chapter-by-chapter; ~45 min per chapter

The closest thing the field has to an agreed factbase: nine chapters of methodologically transparent data on model performance, investment, adoption, policy, public opinion and responsible AI, updated annually. Everyone else's slide deck ultimately cites this. Read the Economy and Policy chapters if nothing else — they are the ones that settle arguments in a boardroom.

• Benchmark data
• Investment
• Policy
• Adoption

Anthropic Economic Index↗(opens in a new tab)

Anthropic/Research index/Ongoing, updated 2026/40 min

Ongoing analysis of how AI is actually used at work, built on anonymised usage telemetry rather than self-report — which makes it a rare corrective to survey-based claims. Tracks the automation-versus-augmentation split by occupation and task, and now includes geographic and enterprise API breakdowns. The underlying dataset is published openly, so the claims are checkable.

• Labour market
• Usage data
• Automation vs augmentation

BCG AI Radar 2026: As AI Investments Surge, CEOs Take the Lead↗(opens in a new tab)

Boston Consulting Group/Survey/2026-01/25 min

A CEO-level rather than CIO-level read, based on roughly 2,400 executives including several hundred CEOs across sixteen markets. Its central finding is about ownership: AI has moved from a delegated technology programme to something CEOs claim personally, with investment commitments that hold regardless of near-term returns. The segmentation of CEOs by posture is a useful diagnostic to run on your own leadership team.

• CEO agenda
• Investment
• Governance

EIB Investment Survey 2025: EU Overview↗(opens in a new tab)

European Investment Bank/Survey/2025-10/1 hour

Roughly 13,000 EU firms and 800 US firms surveyed on the same instrument — the only large-sample, like-for-like transatlantic comparison of enterprise AI adoption. The headline is uncomfortable and specific: European firms are close to parity on whether they use generative AI, and well behind on how deeply. That distinction between adoption and depth is the actual European competitiveness problem, and it is a consulting thesis in one chart.

• Europe
• Competitiveness
• EU vs US
• Investment

Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity↗(opens in a new tab)

METR/Research paper/2025-07/30 min

A pre-registered randomised trial in which experienced developers working in codebases they knew well were measurably slower with AI tools — while believing they had been faster. The finding is narrow and the authors say so, but the mechanism is not: perceived productivity and actual productivity diverge, which invalidates the self-reported surveys most AI business cases rest on. Essential calibration before promising a client a percentage.

• Productivity
• RCT
• Developer tools
• Measurement

Navigating the Jagged Technological Frontier↗(opens in a new tab)

Harvard Business School AI Institute with BCG/Research paper/2023-09; peer-reviewed 2025/45 min

The gold-standard randomised field experiment on knowledge work: 758 BCG consultants, large gains in speed and quality on tasks inside the model's capability frontier, and degraded performance on tasks just outside it. The "jagged frontier" framing it introduced — capability boundaries that are invisible and non-intuitive — is now the standard vocabulary for why naive deployment backfires. Task selection, not model selection, is the variable that matters.

• Field experiment
• Knowledge work
• Task selection

The State of AI 2026↗(opens in a new tab)

McKinsey & Company (QuantumBlack)/Survey/2026-08/30 min

The most-cited annual enterprise AI survey, and the clearest quantification of the gap that defines this market phase: near-universal adoption alongside a small minority of firms able to trace AI to EBIT. The 2026 edition also tracks how far large enterprises have got with agents specifically. Useful as a benchmark to position a client against peers, and as evidence that the constraint is organisational, not technological.

• Adoption
• ROI
• Agents
• Benchmarking

Use of Artificial Intelligence in Enterprises↗(opens in a new tab)

Eurostat/Official statistics/Annual; 2025 wave published 2025-12/20 min

The official, methodologically standardised AI adoption statistic across all 27 member states, broken down by firm size, sector and country — the numbers a European board will already have seen quoted in national press. Adoption is starkly size-dependent and the Nordic-to-Southeast spread is wide enough to reframe any pan-European rollout plan. Free, downloadable, and citable without licensing friction.

• Europe
• Adoption statistics
• Member state comparison

AI in the Enterprise: Lessons from Seven Frontier Companies↗(opens in a new tab)

OpenAI/Guide/2025-05/45 min

Seven named enterprise deployments — Morgan Stanley, BBVA, Klarna, Lowe's and others — distilled into seven operating lessons, the first of which is to start with evaluations rather than pilots. Vendor-authored, and worth reading anyway: the case detail is specific enough to argue with, and the emphasis on embedding AI in products rather than bolting it onto workflows is correct. A good artefact to hand a sceptical executive sponsor.

Vendor-authored (OpenAI)

• Case studies
• Deployment
• Evals
• Change management

Apply AI Strategy (COM(2025) 723)↗(opens in a new tab)

European Commission/Strategy document/2025-10/1 hour

The Commission's answer to the Draghi competitiveness diagnosis, and the policy document that shapes where EU AI money and procurement preference actually flow: sectoral adoption targets, an Apply AI Alliance, an AI Observatory, an "AI-first" stance for public administration and an explicit European-preference posture. If your client sells to or buys from the European public sector, this determines the weather.

• EU policy
• Industrial strategy
• Funding
• Sovereignty

Making AI Work: Leadership, Lab, and Crowd↗(opens in a new tab)

Ethan Mollick (Wharton)/Article/2025-05/12 min

The best twelve minutes in the entire adoption literature. Mollick names the mechanism behind the adoption-without-ROI paradox — employees quietly achieving large personal speedups and hiding them, because disclosing efficiency has historically been punished — and proposes a three-part fix: leadership that sets direction and safety, a lab that industrialises what works, and a crowd that discovers it. If you read one thing on why AI programmes stall, read this.

• Change management
• Adoption
• Incentives

Real-World Generative AI Use Cases from Industry Leaders↗(opens in a new tab)

Google Cloud/Living catalogue/2024-04, continuously expanded/Browse

A continuously growing catalogue of named, attributed generative AI deployments — it began at 101 use cases in 2024 and stood at 1,302 as of the April 2026 update — organised so you can filter to your client's industry and function. Every entry names the company, which makes it usable as evidence rather than inspiration. The single best antidote to a use-case workshop that stalls.

Vendor-authored (Google Cloud); use-case count grows over time

• Use cases
• Industry examples
• Ideation

The Agentic Organization: Contours of the Next Paradigm for the AI Era↗(opens in a new tab)

McKinsey & Company/Article/2025-09/20 min

The most substantive attempt yet to answer what an organisation looks like when agents do a meaningful share of the work: spans of control, decision rights, and the shift from managing people to supervising fleets of agents. It is deliberately speculative in places, but it asks the right question — almost every enterprise is running twenty-first-century AI inside a twentieth-century functional hierarchy, and that mismatch is where value leaks.

• Operating model
• Org design
• Decision rights

12-Factor Agents↗(opens in a new tab)

HumanLayer (Dex Horthy)/Repository / guide/2025, actively maintained/45 min

Twelve principles for taking an agent from impressive demo to something you can operate: own your prompts, own your context window, own your control flow, unify execution and business state, keep agents small and focused, compact errors back into context. Written by practitioners rather than a lab, and it converges independently on the same conclusions Anthropic reached — which is the strongest evidence it is right.

• Production readiness
• Control flow
• State management

A Practical Guide to Building Agents↗(opens in a new tab)

OpenAI/Guide/2025-04/1 hour

Thirty-three pages on when a problem actually warrants an agent, how to choose between single-agent and multi-agent orchestration, and — the part worth the download — a layered guardrail design combining relevance and safety classifiers, tool-risk ratings and human escalation. The most enterprise-legible agent guide from either major lab, and the one to hand an architect who needs a design template rather than a philosophy.

Vendor-authored (OpenAI)

• Agent design
• Guardrails
• Orchestration
• Human escalation

Building Effective Agents↗(opens in a new tab)

Anthropic/Article/2024-12/15 min

The article that gave the industry its shared vocabulary: the workflow-versus-agent distinction, and the five composable patterns (prompt chaining, routing, parallelisation, orchestrator-workers, evaluator-optimiser) that almost every production system still decomposes into. Its most valuable advice remains the least followed — most problems do not need an agent, and the simplest composition that works is the right one. Anthropic now flags the tooling detail as dated; the concepts are not.

Tooling detail superseded; read for the patterns

• Agent patterns
• Workflows
• Orchestration

Context Engineering for AI Agents: Lessons from Building Manus↗(opens in a new tab)

Manus (Yichao Ji)/Article/2025-07/20 min

A production war story from an engineer who rebuilt his agent framework four times, and the rare piece that gives you mechanisms rather than principles: treat KV-cache hit rate as your primary performance metric, mask tools instead of removing them to preserve the cache, use the filesystem as externalised memory, and deliberately leave failure traces in context so the model learns from them. Directly actionable on cost and latency.

• Context engineering
• KV cache
• Cost
• Failure handling

Don't Build Multi-Agents↗(opens in a new tab)

Cognition (Walden Yan)/Article/2025-06; revisited 2026-04/10 min

The sharpest counter-argument to fashionable multi-agent design: parallel sub-agents cannot see each other's implicit decisions, so their outputs conflict in ways that compound silently. The 2026 follow-up softens the position usefully — multiple agents contributing intelligence is fine, provided writes stay single-threaded and one agent owns the outcome. Essential paired reading with Anthropic's multi-agent case study.

• Multi-agent
• Architecture
• Reliability

Effective Context Engineering for AI Agents↗(opens in a new tab)

Anthropic/Article/2025-09/20 min

The reframe that replaced prompt engineering as the central discipline: your job is to curate the smallest set of high-signal tokens the model needs at each step, across a whole trajectory rather than a single turn. Covers just-in-time retrieval versus upfront loading, compaction, structured note-taking and sub-agent context isolation. This is the current best entry point to Anthropic's agent thinking and the piece their later work builds on.

• Context engineering
• Memory
• Sub-agents
• Compaction

How We Built Our Multi-Agent Research System↗(opens in a new tab)

Anthropic/Case study/2025-06/20 min

The most honest public account of a production multi-agent system: what the lead-agent/sub-agent architecture bought in quality, what it cost in tokens (roughly an order of magnitude more than a single chat turn), and how they debugged non-deterministic failures with full-trace observability and staged rollout. Read it alongside Cognition's argument against multi-agent systems — the disagreement between the two is where the real design judgement lives.

• Multi-agent
• Orchestration
• Observability
• Cost

Scaling Managed Agents: Decoupling the Brain from the Hands↗(opens in a new tab)

Anthropic/Article/2026-04/15 min

Anthropic's 2026 answer to the problem every enterprise hits about eighteen months in: the agent harness you built encodes assumptions about model capability that go stale with each model release, and rewriting it each time is untenable. The proposal is to virtualise reasoning, execution and session state behind swappable interfaces so the model can improve without the scaffolding being rebuilt. Read it before you commit architecture you will regret.

• Architecture
• Harness design
• Future-proofing

Agent2Agent (A2A) Protocol↗(opens in a new tab)

Agentic AI Foundation (Linux Foundation), originated by Google/Specification/2025, foundation-governed since 2026-08/30 min

The complement to MCP: where MCP connects an agent to tools, A2A lets agents built by different vendors discover each other, negotiate capability and delegate work. Backed by a steering committee spanning AWS, Cisco, Google, IBM, Microsoft, Salesforce, SAP and ServiceNow, and now under the same foundation as MCP without merging into it. Matters most for enterprises that will inevitably run agents from several vendors at once.

• A2A
• Interoperability
• Multi-vendor
• Agent discovery

Model Context Protocol Specification↗(opens in a new tab)

Agentic AI Foundation (Linux Foundation)/Specification/Revision 2026-07-28/45 min

The protocol that standardises how agents connect to tools and data, now governed by the Linux Foundation's Agentic AI Foundation with AWS, Anthropic, Google, Microsoft and OpenAI among its backers — which is what turned it from a vendor convention into an integration standard you can architect against. The current revision adds asynchronous long-running tasks, elicitation and opt-in extensions. Read the authorisation section closely; it is where enterprise risk concentrates.

• MCP
• Tool integration
• Standards
• Authorisation

AI Evals: Everything You Need to Know↗(opens in a new tab)

Hamel Husain & Shreya Shankar/Guide/Continuously updated/1-2 hours

The free companion to the most respected practitioner evals curriculum in the field, structured as answers to the questions teams actually get stuck on: how to do error analysis, when LLM-as-judge is trustworthy, how many human annotations you need, how to evaluate RAG and agentic systems. Opinionated, empirical, and refreshingly hostile to vanity metrics. Start here before buying any evaluation tooling.

• Error analysis
• LLM-as-judge
• Annotation
• RAG evaluation

Demystifying Evals for AI Agents↗(opens in a new tab)

Anthropic/Article/2026-01/20 min

Why agents cannot be evaluated like single-turn models — they take multiple steps, call tools and change state, so the thing under test is a trajectory, not an answer. Sets out when to use code-based graders, model-based graders and human review, and how the mix differs for coding, research, conversational and computer-use agents. The most practical primary-source guidance on the discipline that separates demos from deployments.

• Evals
• Graders
• Testing
• Reliability

MITRE ATLAS↗(opens in a new tab)

MITRE/Knowledge base/Living resource, updated 2026/Reference

ATT&CK for AI systems: a structured knowledge base of real adversary tactics, techniques and mitigations against machine learning and, increasingly, agentic systems, backed by documented case studies. It is the common reference every other AI security framework maps back to, which makes it the right substrate for red-teaming and for talking to a CISO in language they already use.

• Adversarial ML
• TTPs
• Red teaming
• Case studies

OWASP GenAI LLM Top 10 (2026)↗(opens in a new tab)

OWASP GenAI Security Project/Guideline/2026-08/1 hour

The de facto standard risk taxonomy for LLM applications — prompt injection, insecure output handling, supply chain, data leakage and the rest — refreshed for 2026 from real incident data and cross-mapped to NIST, MITRE ATLAS and CWE. Every serious AI security assessment is scoped against this list. If your reference is still the 2025 edition, it has been superseded.

Supersedes the 2025 edition

• Prompt injection
• Application security
• Threat modelling

OWASP Top 10 for Agentic Applications (2026)↗(opens in a new tab)

OWASP GenAI Security Project — Agentic Security Initiative/Guideline/2025-12/1 hour

The agent-specific companion to the LLM Top 10, covering the risks that only appear once a system takes autonomous multi-step action: memory poisoning, tool misuse, privilege compromise, cascading failures across agent chains. Built with over a hundred contributors. If you are deploying agents against real systems and have not threat-modelled against this list, you have not threat-modelled.

• Agent security
• Memory poisoning
• Tool misuse
• Privilege

Secure AI Framework (SAIF)↗(opens in a new tab)

Google/Framework/2025-10/45 min

Google's security framework, extended in its 2.0 revision to agents specifically, with three principles that make good design constraints: agents must have well-defined human controllers, their powers must be bounded by purpose, and their actions and reasoning must be observable. The accompanying agent risk self-assessment turns those principles into something you can review a design against. Vendor-authored but unusually free of product placement.

Vendor-authored (Google)

• Agent security
• Design principles
• Risk mapping

AI Controls Matrix (AICM) & STAR for AI↗(opens in a new tab)

Cloud Security Alliance/Control framework/v1.0 2025-07; updated 2026/Reference

The granular control catalogue that sits underneath the frameworks: roughly 240 control objectives across eighteen domains, mapped simultaneously to ISO 42001, ISO 27001, BSI AIC4 and the EU AI Act. Where NIST and ISO tell you what to govern, this tells you what to implement and how to evidence it. The STAR for AI programme turns it into a public assurance registry, which is useful for vendor due diligence.

• Controls
• Assurance
• ISO 42001 mapping
• EU AI Act mapping

ISO/IEC 42001 — AI Management Systems↗(opens in a new tab)

ISO/IEC JTC 1/SC 42/Standard/2023-12/Implementation programme

The only certifiable AI management system standard, and increasingly a line item in European enterprise RFPs — which is what makes it commercially relevant regardless of what you think of management-system standards. Its 2025 companions complete the toolkit: 42005 for AI system impact assessment and 42006 setting requirements for the bodies that certify you. Certification is not a presumption of conformity with the EU AI Act, but it is the most credible bridge available.

Standard text is paywalled

• Certification
• AIMS
• Procurement
• Audit

International AI Safety Report 2026↗(opens in a new tab)

Independent expert panel chaired by Yoshua Bengio, backed by 30+ countries/Report/2026-02/Executive summary 20 min; full report 4+ hours

The nearest thing to an IPCC-style consensus assessment of AI capability and risk: over a hundred independent experts, thirty-plus governments and international bodies, and no vendor or single national sponsor. That independence is precisely why it works for board-level risk conversations where a lab's own safety framework would not. Read the twenty-page policymaker summary; the full report is a reference.

• Frontier risk
• Scientific consensus
• Board briefing

NIST AI Risk Management Framework & Generative AI Profile↗(opens in a new tab)

NIST (US Department of Commerce)/Framework/AI RMF 1.0 2023-01; GenAI Profile 2024-07/3 hours

The most widely cross-referenced AI risk framework in the world, organised around four functions — Govern, Map, Measure, Manage — with a companion Generative AI Profile enumerating the risks specific to foundation models. Voluntary and US-origin, but it is the lingua franca: ISO 42001, the CSA control matrix and most vendor frameworks publish crosswalks to it. The searchable Playbook turns it from a document into actionable controls.

NIST has signalled AI RMF 1.0 is under revision

• Risk management
• Govern-Map-Measure-Manage
• Crosswalks

CEN-CENELEC JTC 21 — AI Standardisation↗(opens in a new tab)

CEN-CENELEC Joint Technical Committee 21/Standards programme/Ongoing; EN 18286 published 2026/Reference

Where the harmonised standards that give presumption of conformity under the AI Act are actually being written — and the honest picture is that most of them are not finished. The quality management standard EN 18286 published in 2026 as the first of the set; risk management, data and bias, logging and cybersecurity standards remain in development. Anyone planning high-risk conformity work needs to track this, because the gap is a live compliance risk, not a technicality.

Most standards still in development

• Harmonised standards
• Presumption of conformity
• High-risk

EDPB Opinion 28/2024 on Data Protection and AI Models↗(opens in a new tab)

European Data Protection Board/Official guidance/2024-12/1.5 hours

The EU-wide regulatory position on how GDPR applies to AI models, and the document that decides three questions every European enterprise eventually faces: when a model can be considered anonymous, when legitimate interest supports training on personal data, and what happens downstream when training data was unlawfully obtained. Answering these badly is where AI programmes in banking and insurance actually get stopped.

• GDPR
• Legitimate interest
• Anonymity
• Training data

EIOPA Opinion on AI Governance and Risk Management↗(opens in a new tab)

European Insurance and Occupational Pensions Authority/Supervisory guidance/2025-08/1.5 hours

The only EU-level supervisory opinion interpreting existing insurance governance duties for AI use across pricing, underwriting, claims and fraud — deliberately covering the AI systems the AI Act does not classify as high-risk, so the two read as complements rather than duplicates. Proportionate and risk-based in structure. If you advise European insurers, this is the document their board risk committee will be working from.

• Insurance
• Solvency II
• IDD
• Fairness
• Sector supervision

EU AI Act Explorer↗(opens in a new tab)

Future of Life Institute/Tool/Continuously updated/Reference

The best free way to actually navigate the AI Act: article-by-article browsing with recitals, cross-references and annexes linked in place, plus a maintained implementation timeline and a compliance checker. Independent of the Commission and of industry, and kept current against amendments. This is the working tool; the EUR-Lex text is the citation.

Independent, not an official EU source

• AI Act
• Navigation
• Timeline
• Article-by-article

EU AI Act Service Desk & Compliance Checker↗(opens in a new tab)

European Commission — AI Office/Tool/Launched 2025, ongoing/Interactive

The Commission's own front door for AI Act questions: an interactive compliance checker that scopes whether and how the Act applies to a given system, plus a direct channel to AI Office experts. The compliance checker is explicitly in beta and is not a legal determination, but as an official first-pass scoping tool it carries weight that no consultancy's questionnaire does.

Compliance checker is in beta; not a legal determination

• AI Act
• Official guidance
• Scoping
• SME support

EU AI Act — Consolidated Text↗(opens in a new tab)

EUR-Lex, Publications Office of the EU/Regulation/Consolidated 2026-07-27/Reference

The actual law, consolidated to include the 2026 AI Omnibus amendments — which matters more than it sounds, because a great deal of commentary still cites the pre-amendment timeline. The Omnibus deferred the high-risk obligations (Annex III systems to December 2027, product-embedded systems to August 2028) while leaving the GPAI obligations in force from August 2025. When a date is disputed, this is the source that settles it.

Consolidated texts are unofficial; the Official Journal version is authoritative

• AI Act
• Primary source
• High-risk
• GPAI

General-Purpose AI Code of Practice↗(opens in a new tab)

European Commission — AI Office/Code of practice/2025-07/2-3 hours

The Commission-endorsed route to demonstrating compliance with the AI Act's general-purpose model obligations, in three chapters — transparency (with a model documentation form), copyright, and safety and security for systemic-risk models. Signed by most major model providers, which makes it the de facto industry baseline. Read it even if you only deploy models: it tells you exactly what documentation you are entitled to demand from your suppliers.

• GPAI
• Transparency
• Copyright
• Systemic risk