EU AI Act High-Risk Implementation Readiness
Use this skill when a system has already been classified as potentially high-risk under the EU AI Act and the user now needs to understand what must actually be implemented, documented, assigned, tested, and governed.
If the system has not yet been classified, use the EU AI Act System Classifier first.
This skill is designed as a practical readiness assessment and implementation navigator for:
- Providers of high-risk AI systems
- Deployers of high-risk AI systems
- Internal legal, compliance, product, engineering, security, risk, procurement, and management teams
- Especially DACH-based organizations preparing for real operational compliance work before the high-risk obligations apply
Important timing note: Annex III high-risk obligations apply from 2 December 2027 and Annex I product-integrated high-risk obligations from 2 August 2028. The Digital Omnibus simplification package amending the EU AI Act was adopted on 8 July 2026 and published in the Official Journal on 24 July 2026 as Regulation (EU) 2026/1744, entering into force on 27 July 2026; it deferred these dates from 2 August 2026 and 2 August 2027 respectively. Article 50 transparency obligations and the start of Commission GPAI enforcement powers were not deferred and continue to apply from 2 August 2026. Build against these enacted dates. The substance also moved in places: the same regulation amended Art. 10 and inserted Art. 4a. Article 4a(1) creates an exceptional bias-detection and correction basis for high-risk providers, subject to six safeguards. Article 4a(2) separately covers providers and deployers of other AI systems and models and deployers of high-risk AI systems, but only where processing is strictly necessary for biases likely to affect health or safety, negatively affect fundamental rights, or cause prohibited discrimination, and all paragraph 1 safeguards are applied; it creates no duty to conduct bias detection or correction. The regulation also amended Art. 11(1) (simplified technical documentation for SMEs and SMCs, which notified bodies must accept), Art. 17(2) and Art. 63(1) (quality-management proportionality and simplification), Art. 25(2) and (4) (provider-switch cooperation), Art. 43(3) (conformity-assessment routing, and sectoral notified bodies must apply for designation by 28 January 2028), Annex VIII Section B points 7 and 9 deleted (Art. 49(2) itself is unamended; the deletion reduces the information submitted for Art. 6(3) registrations), and Art. 72(3) (post-market monitoring template replaced by Commission guidance due 2 September 2027).
What this skill does
This skill helps the user answer five practical questions:
- Are we really in scope as high-risk?
- What obligations apply to us as provider, deployer, or both?
- What evidence and documentation should already exist?
- How ready are we today - RED, AMBER, or GREEN?
- What do we need to do next, in what order, and who should own it?
It covers operational readiness across the main Annex III lifecycle obligations, especially:
- Art. 6 and the Art. 6(3) exception context
- Arts. 9–15 provider controls
- Arts. 16–17 provider governance and QMS framing
- Art. 26 deployer obligations
- Art. 43 conformity assessment pathways
- Art. 49 EU database registration
- Art. 72 post-market monitoring
When to use this skill
Use this skill when the user asks things like:
- “We know the system is high-risk. What do we need to implement now?”
- “Can you assess our readiness for the EU AI Act?”
- “What evidence do we need for Annex III high-risk compliance?”
- “Create a gap analysis and implementation roadmap for our AI system.”
- “What do providers of high-risk AI need beyond classification?”
- “What do deployers need to do under Art. 26?”
- “Do we need a notified body or can we self-assess?”
- “How do we prepare technical documentation and QMS for a high-risk AI system?”
Quick intake questions
Start by collecting concise answers to these questions. If the user does not know, mark assumptions clearly.
A. Scope and role
- What does the AI system do in practice?
- Which Annex III category is believed to apply?
- Has the system already been classified using a separate high-risk classifier?
- Is there any argument that the Art. 6(3) exception applies, meaning the system may fall within Annex III wording but does not materially influence the decision-making in a way that makes it high-risk?
- Are you acting as provider, deployer, or both?
- Is the system already on the market / in service, in pilot phase, or still in development?
B. Organizational context
- Which legal entity owns the system and which teams operate it?
- In which countries will the system be placed on the market or used?
- Is the use case in a regulated sector such as employment, education, essential services, law enforcement, migration, justice, or biometric identification?
- Is this a standalone AI system, a component embedded in software, or integrated into a broader product/service?
C. Technical and control environment
- What data is used for training, validation, testing, and live operation?
- What logs are currently generated automatically?
- What human review or override exists today?
- What testing exists for accuracy, robustness, bias, and security?
- Is there already technical documentation, model documentation, validation documentation, or a QMS in place?
D. Governance and evidence
- Is there a named owner for compliance readiness?
- Is there a formal risk management process for the AI system?
- Are there supplier/vendor dependencies, including GPAI or third-party model providers?
- Is post-market monitoring already planned?
- Is management expecting a simple legal memo or an implementation-grade roadmap with evidence requirements?
Decision tree: start here
Step 1 - Confirm classification posture
- If the system has not yet been classified, stop and direct the user to the EU AI Act System Classifier first.
- If the system appears to match Annex III but there may be a credible Art. 6(3) exception argument, do not proceed as if high-risk is settled. Flag the issue and recommend a focused classification memo before readiness work continues.
- If the system is reasonably treated as high-risk, continue.
Step 2 - Determine actor posture
- If the organization develops / places on the market / puts into service under its own name, assess provider obligations.
- If the organization uses the system in its operations, assess deployer obligations.
- If it does both, run both tracks and separate deliverables clearly.
Step 3 - Determine assessment mode
Choose one of three modes:
- Rapid triage - fast RED/AMBER/GREEN scan across all obligation areas
- Evidence-based readiness assessment - assess each area against actual documents, processes, owners, and controls
- Implementation roadmap - turn identified gaps into a sequenced action plan with owners, dependencies, and deliverables
Step 4 - Determine conformity assessment path
- For Annex III point 1, Annex VI or Annex VII is available only where all relevant harmonised standards or common specifications are applied; otherwise Annex VII applies.
- For Annex III points 2 to 8, Article 43(2) requires Annex VI internal control with no notified body.
- For Annex I Section A products, follow the sectoral conformity-assessment route under Article 43(3). For Annex I Section B products, account for the limited AI Act application stated in Article 2(2).
- If unclear, flag this early because it changes evidence expectations and timelines
Core workflow
1. Frame the scope precisely
Produce a short scoping statement covering:
- system name / working label
- practical function
- likely Annex III category
- provider/deployer role split
- lifecycle stage
- jurisdictions
- whether Art. 6(3) remains open or has been ruled out
Output at this stage: one-paragraph scope statement + assumption list.
2. Assess readiness across the 12 obligation areas
For each area below, evaluate:
- What the AI Act requires
- What evidence should exist
- Readiness status: RED / AMBER / GREEN
- Key gaps
- Next actions
- Common pitfalls
Area 1 - Risk management system (Art. 9)
Ask:
- Is there a defined AI-specific risk management process across the lifecycle?
- Are known and reasonably foreseeable risks identified and documented?
- Are risk controls tied to testing, design, and residual risk evaluation?
- Are changes, incidents, and monitoring findings fed back into the system?
Evidence examples:
- risk management procedure
- system risk register
- hazard / harm analysis
- control mapping
- residual risk sign-off
- validation and test evidence
Use deep dive: references/risk-management-system.md
Area 2 - Data governance and data quality (Art. 10)
Ask:
- What datasets were used for training, validation, and testing?
- Are relevance, representativeness, completeness, and error considerations documented?
- Was bias examination performed and recorded?
- Is data provenance and acquisition legality documented?
Evidence examples:
- dataset inventory
- data specification sheets
- bias assessment
- sampling rationale
- data cleaning and labeling procedures
- validation dataset design notes
Use deep dive: references/data-governance.md
Area 3 - Technical documentation (Art. 11 + Annex IV)
Ask:
- Could the organization produce a defensible Annex IV-style technical file today?
- Is system design, development method, intended purpose, architecture, metrics, risk controls, and change history documented?
- Are monitoring capabilities and limitations explained clearly enough for review?
Evidence examples:
- technical file / Annex IV pack
- architecture diagrams
- model cards / system cards
- development and validation methodology
- version/change log
- limitations and assumptions documentation
Use deep dive: references/technical-documentation.md
Area 4 - Record-keeping and logging (Art. 12)
Ask:
- What logs are created automatically?
- Do logs support traceability for operation, incidents, human intervention, and review?
- Is retention aligned with legal and operational needs?
- Can logs support investigations and post-market monitoring?
Evidence examples:
- logging specification
- event taxonomy
- retention schedule
- access controls for logs
- sample audit trail outputs
Area 5 - Transparency and information to deployers (Art. 13)
Ask:
- Are there instructions for use?
- Are intended purpose, operating conditions, known limitations, expected accuracy, oversight assumptions, and security conditions explained?
- Would a deployer know when the system should not be used?
Evidence examples:
- instructions for use
- deployment manual
- limitation statements
- accuracy/robustness documentation
- user-facing warnings and assumptions
Area 6 - Human oversight (Art. 14)
Ask:
- Who is expected to oversee the system?
- Can they understand outputs well enough to intervene meaningfully?
- Is there a stop, override, or escalation mechanism?
- Is oversight designed into the process rather than assumed abstractly?
Evidence examples:
- oversight design specification
- SOPs for reviewers/operators
- override or manual fallback procedures
- training materials
- escalation matrix
Area 7 - Accuracy, robustness, and cybersecurity (Art. 15)
Ask:
- What performance thresholds are defined?
- How was robustness tested under foreseeable operating conditions?
- What error handling and fail-safe logic exists?
- What cybersecurity risks, including adversarial manipulation, have been considered?
Evidence examples:
- performance benchmarks
- validation reports
- robustness/stress testing
- security testing results
- vulnerability management records
Area 8 - Quality management system (Art. 17)
Ask:
- Is there a documented QMS covering the AI lifecycle?
- Are design, development, testing, change management, supplier control, incident handling, and authority communication governed?
- Are roles and records assigned in a way that can survive audit or conformity review?
Evidence examples:
- QMS manual
- policies and procedures
- role matrix / RACI
- change control procedure
- supplier management records
- CAPA / incident process
Use deep dive: references/qms-requirements.md
Area 9 - Deployer obligations (Art. 26)
Ask:
- Is the system being used in accordance with the provider’s instructions?
- Are human oversight responsibilities assigned?
- Is input data checked for relevance and suitability?
- Are records maintained during use?
- Are affected persons informed where required?
- Is a fundamental rights impact assessment required under Art. 27(1): is the deployer a public-law body or private public-service entity, or is the system an Annex III point 5(b) creditworthiness/credit-scoring or point 5(c) life/health-insurance system (where every deployer owes the FRIA)? Annex III point 2 is excepted.
Evidence examples:
- deployer SOPs
- user governance record
- oversight assignments
- operational monitoring logs
- notice/information materials
- FRIA documentation where relevant
Use deep dive: references/deployer-obligations.md
Area 10 - Conformity assessment path (Art. 43)
Ask:
- Is the system likely on the Annex VI internal control route?
- Is third-party assessment / notified body involvement required?
- What evidence package would be expected under the chosen route?
- Who owns the conformity assessment timeline?
Evidence examples:
- classification memo
- conformity assessment path memo
- Annex VI checklist
- notified body engagement plan if needed
Use deep dive: references/conformity-assessment.md
Area 11 - Post-market monitoring (Art. 72)
Ask:
- Is there a documented, proportionate post-market monitoring plan?
- What data will be collected from live operation?
- How are incidents, malfunctions, and performance drift analyzed?
- How do findings feed back into risk management, documentation, and controls?
Evidence examples:
- post-market monitoring plan
- KPI/KRI definitions
- incident intake workflow
- review cadence and reporting structure
Area 12 - EU database registration (Art. 49)
Ask:
- Who is responsible for registration before placing on the market or putting into service?
- Is the required registration data known and assembled?
- Is registration integrated into launch governance?
Evidence examples:
- registration readiness checklist
- responsibility assignment
- pre-launch gate / approval workflow
3. Apply the readiness scoring model
Use this scoring consistently for every area.
RED - Not started / materially deficient
Use RED where one or more of the following is true:
- no documented process exists
- no owner is assigned
- evidence is missing or purely informal
- controls exist in practice but are not systematic, documented, or reviewable
- the organization could not defend the area in a conformity assessment or authority inquiry
AMBER - Partially addressed / not yet defensible end-to-end
Use AMBER where:
- some controls exist but are fragmented
- evidence exists but is incomplete, inconsistent, outdated, or not AI-specific
- roles are partly assigned but not embedded operationally
- testing or monitoring exists but is not tied back to governance and risk decisions
GREEN - Substantially ready
Use GREEN where:
- the process is defined, operational, and documented
- evidence is current and reviewable
- ownership is assigned
- the area is integrated into lifecycle governance
- there are still improvement opportunities, but the organization is broadly defensible
Do not mark GREEN purely because “the team is already careful” or “similar controls exist somewhere else.”
4. Identify critical dependencies and blockers
After scoring, identify blockers such as:
- unresolved Art. 6(3) scope question
- unclear provider vs deployer role split
- lack of named accountable owner
- no technical documentation baseline
- no AI-specific risk register
- insufficient logging or traceability
- no oversight design
- reliance on a third-party model/vendor with weak documentation support
- missing QMS backbone
- unclear conformity assessment path
Group blockers into:
- Legal/classification blockers
- Process/governance blockers
- Technical/control blockers
- Evidence/documentation blockers
5. Build a practical implementation roadmap
Translate gaps into a sequenced roadmap.
Recommended workstreams:
- Scope and accountability
- Risk management
- Data governance
- Documentation and logging
- Human oversight and operations
- Testing, robustness, and cybersecurity
- QMS and authority-facing readiness
- Deployer operating model
- Conformity assessment and registration
- Post-market monitoring
For each workstream define:
- objective
- key deliverables
- owner
- supporting teams
- dependencies
- target timing
- residual decision points
Use templates in references/templates.md.
6. Tailor for DACH implementation reality
Where relevant, add practical DACH-specific points such as:
- likely interaction with German market surveillance or sector regulators
- BNetzA / BSI / sector authority interfaces depending on use case
- procurement and documentation expectations in German enterprises
- works council implications where human oversight affects employees or AI is used in HR contexts
- need for German-language operational materials, training, or works agreements in practice
Use: references/dach-specific.md
Practical shortcuts and pitfalls to flag
Always call out shortcuts that look attractive but are weak in a real assessment.
Common examples:
- “We already have ISO processes, so we must be compliant.”
- “We have general model docs, so that counts as Annex IV technical documentation.”
- “Humans review sometimes” without defining who, when, with what authority, and based on what criteria.
- “Security reviewed the app” without AI-specific robustness and adversarial considerations.
- “We keep logs” without confirming traceability, access, retention, and useful event design.
- “The vendor handles compliance” without obtaining evidence and role clarity.
- “We’ll write the documents later” when core controls are not yet actually operating.
Output format
Unless the user asks for something else, structure the final deliverable like this:
1. Executive summary
- in-scope system and role
- high-level conclusion on readiness
- top 3–5 critical gaps
- immediate next actions
2. Scope and assumptions
- system description
- likely Annex III category
- provider/deployer split
- Art. 6(3) position if relevant
- assumptions / unknowns
3. Readiness scorecard
For each of the 12 areas:
- requirement summary
- expected evidence
- status: RED / AMBER / GREEN
- rationale
- immediate next step
4. Priority gap analysis
- critical gaps
- why they matter
- dependencies and sequencing
5. Implementation roadmap
- 30 / 60 / 90 day view, or phased workstream plan
- owners and dependencies
6. DACH-specific notes
- regulator / authority / works council / operational localization issues
7. Key caveats
- legal uncertainties
- evidence limitations
- any items needing specialist legal or technical validation
Suggested response patterns
If the user wants a quick assessment
Provide:
- brief scope statement
- 12-area RED/AMBER/GREEN scorecard
- top 5 actions
- biggest conformity assessment risk
If the user wants implementation help
Provide:
- scorecard
- gap analysis
- detailed roadmap
- document list to create
- owner suggestions by function
If the user is a deployer only
Focus more heavily on:
- Art. 26 usage controls
- provider instruction adherence
- oversight assignment
- monitoring and record-keeping in operation
- FRIA / affected-person information where relevant
If the user is a provider with a mature quality function
Focus more heavily on:
- evidence sufficiency
- AI-specific adaptations to existing QMS
- Annex IV technical documentation completeness
- risk management lifecycle integration
- conformity assessment readiness
Recommended reference map
Use these deep dives selectively rather than overloading the main response:
references/risk-management-system.md
references/data-governance.md
references/technical-documentation.md
references/qms-requirements.md
references/conformity-assessment.md
references/deployer-obligations.md
references/dach-specific.md
references/templates.md
Legacy systems: the grace period moved (Article 111(2))
Regulation (EU) 2026/1744 replaced Article 111(2) on 27 July 2026, and this matters
directly for question 6 of the intake above.
The cut-off is no longer the fixed date of 2 August 2026. It now tracks the date of
application of Chapter III under Article 113, meaning 2 December 2027 for Annex III
systems and 2 August 2028 for Annex I systems. A high-risk system already placed on the
market or put into service before its applicable date falls outside the requirements unless
it is subject to significant changes in its design from that date onward.
Two qualifications that decide most real cases:
- The grace period is assessed for systems individually placed on the market or put into
service before the applicable date. Later units of the same type or model that are first
supplied after that date are not grandfathered merely because an earlier unit was supplied.
- Public-authority deployments have their own hard stop. Providers and deployers of
high-risk systems intended for use by public authorities must comply by 2 August 2030
regardless, and that date did not move.
Practical readiness consequence: record the placement-on-market or putting-into-service
date for each system or unit relied upon, together with its design baseline. For an existing
system, also ask whether a significant design change will occur from the applicable date.
Disclaimer
This skill provides a practical implementation and readiness framework for the EU AI Act, especially for Annex III high-risk systems. It is not a substitute for formal legal advice, sector-specific regulatory advice, technical assurance, cybersecurity testing, or notified-body input where required.
The AI Act contains cross-references, implementing acts, harmonized standards, and evolving guidance that may change how obligations are interpreted in practice. Where classification is uncertain, where the Art. 6(3) exception may apply, where biometric or sector-specific issues are involved, or where conformity assessment route selection is unclear, the user should validate the position with qualified counsel and relevant technical stakeholders.
What good looks like
A strong outcome from this skill is not just “a compliance memo.” It is:
- a clear scope position
- a defensible readiness score by obligation area
- a concrete evidence list
- a prioritized implementation roadmap
- named ownership
- a practical path toward conformity assessment, deployment readiness, and ongoing monitoring
That is the difference between knowing you are high-risk and being operationally prepared for it.
1---
2name: eu-ai-act-high-risk-implementation-readiness
3description: Assess and operationalize implementation readiness for high-risk AI systems under the EU AI Act Annex III, including provider and deployer obligations, conformity assessment, post-market monitoring, and EU database registration. Use when users say things like “we classified this as high-risk, what now?”, “build an EU AI Act readiness plan”, “assess our Annex III compliance gaps”, “what do providers/deployers of high-risk AI need to implement?”, “prepare for conformity assessment”, or “create a high-risk AI implementation roadmap.”
4metadata:
5 source: https://skillmd.com/skills/gabrielmoreira/eu-ai-act-high-risk-implementation-readiness
6---
7
8# EU AI Act High-Risk Implementation Readiness
9
10Use this skill when a system has already been classified as potentially high-risk under the EU AI Act and the user now needs to understand what must actually be implemented, documented, assigned, tested, and governed.
11
12If the system has **not** yet been classified, use the **EU AI Act System Classifier** first.
13
14This skill is designed as a **practical readiness assessment and implementation navigator** for:
15- **Providers** of high-risk AI systems
16- **Deployers** of high-risk AI systems
17- Internal legal, compliance, product, engineering, security, risk, procurement, and management teams
18- Especially **DACH-based organizations** preparing for real operational compliance work before the high-risk obligations apply
19
20> Important timing note: Annex III high-risk obligations apply from **2 December 2027** and Annex I product-integrated high-risk obligations from **2 August 2028**. The Digital Omnibus simplification package amending the EU AI Act was adopted on 8 July 2026 and published in the Official Journal on 24 July 2026 as Regulation (EU) 2026/1744, entering into force on 27 July 2026; it deferred these dates from 2 August 2026 and 2 August 2027 respectively. Article 50 transparency obligations and the start of Commission GPAI enforcement powers were not deferred and continue to apply from 2 August 2026. Build against these enacted dates. The substance also moved in places: the same regulation amended Art. 10 and inserted Art. 4a. Article 4a(1) creates an exceptional bias-detection and correction basis for high-risk providers, subject to six safeguards. Article 4a(2) separately covers providers and deployers of other AI systems and models and deployers of high-risk AI systems, but only where processing is strictly necessary for biases likely to affect health or safety, negatively affect fundamental rights, or cause prohibited discrimination, and all paragraph 1 safeguards are applied; it creates no duty to conduct bias detection or correction. The regulation also amended Art. 11(1) (simplified technical documentation for SMEs and SMCs, which notified bodies must accept), Art. 17(2) and Art. 63(1) (quality-management proportionality and simplification), Art. 25(2) and (4) (provider-switch cooperation), Art. 43(3) (conformity-assessment routing, and sectoral notified bodies must apply for designation by 28 January 2028), Annex VIII Section B points 7 and 9 deleted (Art. 49(2) itself is unamended; the deletion reduces the information submitted for Art. 6(3) registrations), and Art. 72(3) (post-market monitoring template replaced by Commission guidance due 2 September 2027).
21
22---
23
24## What this skill does
25
26This skill helps the user answer five practical questions:
27
281. **Are we really in scope as high-risk?**
292. **What obligations apply to us as provider, deployer, or both?**
303. **What evidence and documentation should already exist?**
314. **How ready are we today - RED, AMBER, or GREEN?**
325. **What do we need to do next, in what order, and who should own it?**
33
34It covers operational readiness across the main Annex III lifecycle obligations, especially:
35- Art. 6 and the Art. 6(3) exception context
36- Arts. 9–15 provider controls
37- Arts. 16–17 provider governance and QMS framing
38- Art. 26 deployer obligations
39- Art. 43 conformity assessment pathways
40- Art. 49 EU database registration
41- Art. 72 post-market monitoring
42
43---
44
45## When to use this skill
46
47Use this skill when the user asks things like:
48- “We know the system is high-risk. What do we need to implement now?”
49- “Can you assess our readiness for the EU AI Act?”
50- “What evidence do we need for Annex III high-risk compliance?”
51- “Create a gap analysis and implementation roadmap for our AI system.”
52- “What do providers of high-risk AI need beyond classification?”
53- “What do deployers need to do under Art. 26?”
54- “Do we need a notified body or can we self-assess?”
55- “How do we prepare technical documentation and QMS for a high-risk AI system?”
56
57---
58
59## Quick intake questions
60
61Start by collecting concise answers to these questions. If the user does not know, mark assumptions clearly.
62
63### A. Scope and role
641. What does the AI system do in practice?
652. Which Annex III category is believed to apply?
663. Has the system already been classified using a separate high-risk classifier?
674. Is there any argument that the **Art. 6(3) exception** applies, meaning the system may fall within Annex III wording but does **not** materially influence the decision-making in a way that makes it high-risk?
685. Are you acting as **provider**, **deployer**, or both?
696. Is the system already on the market / in service, in pilot phase, or still in development?
70
71### B. Organizational context
727. Which legal entity owns the system and which teams operate it?
738. In which countries will the system be placed on the market or used?
749. Is the use case in a regulated sector such as employment, education, essential services, law enforcement, migration, justice, or biometric identification?
7510. Is this a standalone AI system, a component embedded in software, or integrated into a broader product/service?
76
77### C. Technical and control environment
7811. What data is used for training, validation, testing, and live operation?
7912. What logs are currently generated automatically?
8013. What human review or override exists today?
8114. What testing exists for accuracy, robustness, bias, and security?
8215. Is there already technical documentation, model documentation, validation documentation, or a QMS in place?
83
84### D. Governance and evidence
8516. Is there a named owner for compliance readiness?
8617. Is there a formal risk management process for the AI system?
8718. Are there supplier/vendor dependencies, including GPAI or third-party model providers?
8819. Is post-market monitoring already planned?
8920. Is management expecting a simple legal memo or an implementation-grade roadmap with evidence requirements?
90
91---
92
93## Decision tree: start here
94
95### Step 1 - Confirm classification posture
96- If the system has **not yet been classified**, stop and direct the user to the **EU AI Act System Classifier** first.
97- If the system appears to match Annex III but there may be a credible **Art. 6(3) exception** argument, do **not** proceed as if high-risk is settled. Flag the issue and recommend a focused classification memo before readiness work continues.
98- If the system is reasonably treated as high-risk, continue.
99
100### Step 2 - Determine actor posture
101- If the organization **develops / places on the market / puts into service under its own name**, assess **provider obligations**.
102- If the organization **uses** the system in its operations, assess **deployer obligations**.
103- If it does both, run both tracks and separate deliverables clearly.
104
105### Step 3 - Determine assessment mode
106Choose one of three modes:
107- **Rapid triage** - fast RED/AMBER/GREEN scan across all obligation areas
108- **Evidence-based readiness assessment** - assess each area against actual documents, processes, owners, and controls
109- **Implementation roadmap** - turn identified gaps into a sequenced action plan with owners, dependencies, and deliverables
110
111### Step 4 - Determine conformity assessment path
112- For **Annex III point 1**, Annex VI or Annex VII is available only where all relevant harmonised standards or common specifications are applied; otherwise Annex VII applies.
113- For **Annex III points 2 to 8**, Article 43(2) requires Annex VI internal control with no notified body.
114- For **Annex I Section A** products, follow the sectoral conformity-assessment route under Article 43(3). For **Annex I Section B** products, account for the limited AI Act application stated in Article 2(2).
115- If unclear, flag this early because it changes evidence expectations and timelines
116
117---
118
119## Core workflow
120
121### 1. Frame the scope precisely
122Produce a short scoping statement covering:
123- system name / working label
124- practical function
125- likely Annex III category
126- provider/deployer role split
127- lifecycle stage
128- jurisdictions
129- whether Art. 6(3) remains open or has been ruled out
130
131**Output at this stage:** one-paragraph scope statement + assumption list.
132
133---
134
135### 2. Assess readiness across the 12 obligation areas
136For each area below, evaluate:
137- **What the AI Act requires**
138- **What evidence should exist**
139- **Readiness status:** RED / AMBER / GREEN
140- **Key gaps**
141- **Next actions**
142- **Common pitfalls**
143
144#### Area 1 - Risk management system (Art. 9)
145Ask:
146- Is there a defined AI-specific risk management process across the lifecycle?
147- Are known and reasonably foreseeable risks identified and documented?
148- Are risk controls tied to testing, design, and residual risk evaluation?
149- Are changes, incidents, and monitoring findings fed back into the system?
150
151Evidence examples:
152- risk management procedure
153- system risk register
154- hazard / harm analysis
155- control mapping
156- residual risk sign-off
157- validation and test evidence
158
159Use deep dive: `references/risk-management-system.md`
160
161#### Area 2 - Data governance and data quality (Art. 10)
162Ask:
163- What datasets were used for training, validation, and testing?
164- Are relevance, representativeness, completeness, and error considerations documented?
165- Was bias examination performed and recorded?
166- Is data provenance and acquisition legality documented?
167
168Evidence examples:
169- dataset inventory
170- data specification sheets
171- bias assessment
172- sampling rationale
173- data cleaning and labeling procedures
174- validation dataset design notes
175
176Use deep dive: `references/data-governance.md`
177
178#### Area 3 - Technical documentation (Art. 11 + Annex IV)
179Ask:
180- Could the organization produce a defensible Annex IV-style technical file today?
181- Is system design, development method, intended purpose, architecture, metrics, risk controls, and change history documented?
182- Are monitoring capabilities and limitations explained clearly enough for review?
183
184Evidence examples:
185- technical file / Annex IV pack
186- architecture diagrams
187- model cards / system cards
188- development and validation methodology
189- version/change log
190- limitations and assumptions documentation
191
192Use deep dive: `references/technical-documentation.md`
193
194#### Area 4 - Record-keeping and logging (Art. 12)
195Ask:
196- What logs are created automatically?
197- Do logs support traceability for operation, incidents, human intervention, and review?
198- Is retention aligned with legal and operational needs?
199- Can logs support investigations and post-market monitoring?
200
201Evidence examples:
202- logging specification
203- event taxonomy
204- retention schedule
205- access controls for logs
206- sample audit trail outputs
207
208#### Area 5 - Transparency and information to deployers (Art. 13)
209Ask:
210- Are there instructions for use?
211- Are intended purpose, operating conditions, known limitations, expected accuracy, oversight assumptions, and security conditions explained?
212- Would a deployer know when the system should not be used?
213
214Evidence examples:
215- instructions for use
216- deployment manual
217- limitation statements
218- accuracy/robustness documentation
219- user-facing warnings and assumptions
220
221#### Area 6 - Human oversight (Art. 14)
222Ask:
223- Who is expected to oversee the system?
224- Can they understand outputs well enough to intervene meaningfully?
225- Is there a stop, override, or escalation mechanism?
226- Is oversight designed into the process rather than assumed abstractly?
227
228Evidence examples:
229- oversight design specification
230- SOPs for reviewers/operators
231- override or manual fallback procedures
232- training materials
233- escalation matrix
234
235#### Area 7 - Accuracy, robustness, and cybersecurity (Art. 15)
236Ask:
237- What performance thresholds are defined?
238- How was robustness tested under foreseeable operating conditions?
239- What error handling and fail-safe logic exists?
240- What cybersecurity risks, including adversarial manipulation, have been considered?
241
242Evidence examples:
243- performance benchmarks
244- validation reports
245- robustness/stress testing
246- security testing results
247- vulnerability management records
248
249#### Area 8 - Quality management system (Art. 17)
250Ask:
251- Is there a documented QMS covering the AI lifecycle?
252- Are design, development, testing, change management, supplier control, incident handling, and authority communication governed?
253- Are roles and records assigned in a way that can survive audit or conformity review?
254
255Evidence examples:
256- QMS manual
257- policies and procedures
258- role matrix / RACI
259- change control procedure
260- supplier management records
261- CAPA / incident process
262
263Use deep dive: `references/qms-requirements.md`
264
265#### Area 9 - Deployer obligations (Art. 26)
266Ask:
267- Is the system being used in accordance with the provider’s instructions?
268- Are human oversight responsibilities assigned?
269- Is input data checked for relevance and suitability?
270- Are records maintained during use?
271- Are affected persons informed where required?
272- Is a **fundamental rights impact assessment** required under Art. 27(1): is the deployer a public-law body or private public-service entity, or is the system an Annex III point 5(b) creditworthiness/credit-scoring or point 5(c) life/health-insurance system (where every deployer owes the FRIA)? Annex III point 2 is excepted.
273
274Evidence examples:
275- deployer SOPs
276- user governance record
277- oversight assignments
278- operational monitoring logs
279- notice/information materials
280- FRIA documentation where relevant
281
282Use deep dive: `references/deployer-obligations.md`
283
284#### Area 10 - Conformity assessment path (Art. 43)
285Ask:
286- Is the system likely on the Annex VI internal control route?
287- Is third-party assessment / notified body involvement required?
288- What evidence package would be expected under the chosen route?
289- Who owns the conformity assessment timeline?
290
291Evidence examples:
292- classification memo
293- conformity assessment path memo
294- Annex VI checklist
295- notified body engagement plan if needed
296
297Use deep dive: `references/conformity-assessment.md`
298
299#### Area 11 - Post-market monitoring (Art. 72)
300Ask:
301- Is there a documented, proportionate post-market monitoring plan?
302- What data will be collected from live operation?
303- How are incidents, malfunctions, and performance drift analyzed?
304- How do findings feed back into risk management, documentation, and controls?
305
306Evidence examples:
307- post-market monitoring plan
308- KPI/KRI definitions
309- incident intake workflow
310- review cadence and reporting structure
311
312#### Area 12 - EU database registration (Art. 49)
313Ask:
314- Who is responsible for registration before placing on the market or putting into service?
315- Is the required registration data known and assembled?
316- Is registration integrated into launch governance?
317
318Evidence examples:
319- registration readiness checklist
320- responsibility assignment
321- pre-launch gate / approval workflow
322
323---
324
325### 3. Apply the readiness scoring model
326Use this scoring consistently for every area.
327
328#### RED - Not started / materially deficient
329Use RED where one or more of the following is true:
330- no documented process exists
331- no owner is assigned
332- evidence is missing or purely informal
333- controls exist in practice but are not systematic, documented, or reviewable
334- the organization could not defend the area in a conformity assessment or authority inquiry
335
336#### AMBER - Partially addressed / not yet defensible end-to-end
337Use AMBER where:
338- some controls exist but are fragmented
339- evidence exists but is incomplete, inconsistent, outdated, or not AI-specific
340- roles are partly assigned but not embedded operationally
341- testing or monitoring exists but is not tied back to governance and risk decisions
342
343#### GREEN - Substantially ready
344Use GREEN where:
345- the process is defined, operational, and documented
346- evidence is current and reviewable
347- ownership is assigned
348- the area is integrated into lifecycle governance
349- there are still improvement opportunities, but the organization is broadly defensible
350
351Do **not** mark GREEN purely because “the team is already careful” or “similar controls exist somewhere else.”
352
353---
354
355### 4. Identify critical dependencies and blockers
356After scoring, identify blockers such as:
357- unresolved Art. 6(3) scope question
358- unclear provider vs deployer role split
359- lack of named accountable owner
360- no technical documentation baseline
361- no AI-specific risk register
362- insufficient logging or traceability
363- no oversight design
364- reliance on a third-party model/vendor with weak documentation support
365- missing QMS backbone
366- unclear conformity assessment path
367
368Group blockers into:
369- **Legal/classification blockers**
370- **Process/governance blockers**
371- **Technical/control blockers**
372- **Evidence/documentation blockers**
373
374---
375
376### 5. Build a practical implementation roadmap
377Translate gaps into a sequenced roadmap.
378
379Recommended workstreams:
3801. **Scope and accountability**
3812. **Risk management**
3823. **Data governance**
3834. **Documentation and logging**
3845. **Human oversight and operations**
3856. **Testing, robustness, and cybersecurity**
3867. **QMS and authority-facing readiness**
3878. **Deployer operating model**
3889. **Conformity assessment and registration**
38910. **Post-market monitoring**
390
391For each workstream define:
392- objective
393- key deliverables
394- owner
395- supporting teams
396- dependencies
397- target timing
398- residual decision points
399
400Use templates in `references/templates.md`.
401
402---
403
404### 6. Tailor for DACH implementation reality
405Where relevant, add practical DACH-specific points such as:
406- likely interaction with German market surveillance or sector regulators
407- BNetzA / BSI / sector authority interfaces depending on use case
408- procurement and documentation expectations in German enterprises
409- works council implications where human oversight affects employees or AI is used in HR contexts
410- need for German-language operational materials, training, or works agreements in practice
411
412Use: `references/dach-specific.md`
413
414---
415
416## Practical shortcuts and pitfalls to flag
417
418Always call out shortcuts that look attractive but are weak in a real assessment.
419
420Common examples:
421- “We already have ISO processes, so we must be compliant.”
422- “We have general model docs, so that counts as Annex IV technical documentation.”
423- “Humans review sometimes” without defining who, when, with what authority, and based on what criteria.
424- “Security reviewed the app” without AI-specific robustness and adversarial considerations.
425- “We keep logs” without confirming traceability, access, retention, and useful event design.
426- “The vendor handles compliance” without obtaining evidence and role clarity.
427- “We’ll write the documents later” when core controls are not yet actually operating.
428
429---
430
431## Output format
432
433Unless the user asks for something else, structure the final deliverable like this:
434
435### 1. Executive summary
436- in-scope system and role
437- high-level conclusion on readiness
438- top 3–5 critical gaps
439- immediate next actions
440
441### 2. Scope and assumptions
442- system description
443- likely Annex III category
444- provider/deployer split
445- Art. 6(3) position if relevant
446- assumptions / unknowns
447
448### 3. Readiness scorecard
449For each of the 12 areas:
450- requirement summary
451- expected evidence
452- status: RED / AMBER / GREEN
453- rationale
454- immediate next step
455
456### 4. Priority gap analysis
457- critical gaps
458- why they matter
459- dependencies and sequencing
460
461### 5. Implementation roadmap
462- 30 / 60 / 90 day view, or phased workstream plan
463- owners and dependencies
464
465### 6. DACH-specific notes
466- regulator / authority / works council / operational localization issues
467
468### 7. Key caveats
469- legal uncertainties
470- evidence limitations
471- any items needing specialist legal or technical validation
472
473---
474
475## Suggested response patterns
476
477### If the user wants a quick assessment
478Provide:
479- brief scope statement
480- 12-area RED/AMBER/GREEN scorecard
481- top 5 actions
482- biggest conformity assessment risk
483
484### If the user wants implementation help
485Provide:
486- scorecard
487- gap analysis
488- detailed roadmap
489- document list to create
490- owner suggestions by function
491
492### If the user is a deployer only
493Focus more heavily on:
494- Art. 26 usage controls
495- provider instruction adherence
496- oversight assignment
497- monitoring and record-keeping in operation
498- FRIA / affected-person information where relevant
499
500### If the user is a provider with a mature quality function
501Focus more heavily on:
502- evidence sufficiency
503- AI-specific adaptations to existing QMS
504- Annex IV technical documentation completeness
505- risk management lifecycle integration
506- conformity assessment readiness
507
508---
509
510## Recommended reference map
511
512Use these deep dives selectively rather than overloading the main response:
513- `references/risk-management-system.md`
514- `references/data-governance.md`
515- `references/technical-documentation.md`
516- `references/qms-requirements.md`
517- `references/conformity-assessment.md`
518- `references/deployer-obligations.md`
519- `references/dach-specific.md`
520- `references/templates.md`
521
522---
523
524## Legacy systems: the grace period moved (Article 111(2))
525
526Regulation (EU) 2026/1744 replaced Article 111(2) on 27 July 2026, and this matters
527directly for question 6 of the intake above.
528
529The cut-off is no longer the fixed date of 2 August 2026. It now tracks the date of
530application of Chapter III under Article 113, meaning **2 December 2027** for Annex III
531systems and **2 August 2028** for Annex I systems. A high-risk system already placed on the
532market or put into service before its applicable date falls outside the requirements unless
533it is subject to significant changes in its design from that date onward.
534
535Two qualifications that decide most real cases:
536
537- **The grace period is assessed for systems individually placed on the market or put into
538 service before the applicable date.** Later units of the same type or model that are first
539 supplied after that date are not grandfathered merely because an earlier unit was supplied.
540- **Public-authority deployments have their own hard stop.** Providers and deployers of
541 high-risk systems intended for use by public authorities must comply by **2 August 2030**
542 regardless, and that date did not move.
543
544Practical readiness consequence: record the placement-on-market or putting-into-service
545date for each system or unit relied upon, together with its design baseline. For an existing
546system, also ask whether a significant design change will occur from the applicable date.
547
548## Disclaimer
549
550This skill provides a practical implementation and readiness framework for the EU AI Act, especially for Annex III high-risk systems. It is not a substitute for formal legal advice, sector-specific regulatory advice, technical assurance, cybersecurity testing, or notified-body input where required.
551
552The AI Act contains cross-references, implementing acts, harmonized standards, and evolving guidance that may change how obligations are interpreted in practice. Where classification is uncertain, where the Art. 6(3) exception may apply, where biometric or sector-specific issues are involved, or where conformity assessment route selection is unclear, the user should validate the position with qualified counsel and relevant technical stakeholders.
553
554---
555
556## What good looks like
557
558A strong outcome from this skill is not just “a compliance memo.” It is:
559- a clear scope position
560- a defensible readiness score by obligation area
561- a concrete evidence list
562- a prioritized implementation roadmap
563- named ownership
564- a practical path toward conformity assessment, deployment readiness, and ongoing monitoring
565
566That is the difference between knowing you are high-risk and being operationally prepared for it.