If a regulator asks how an AI system meets its requirements, a vendor contract or an ISO certification alone is not enough. Compliance evidence is documented information that supports that assessment. For a provider, an evidence pack brings the relevant artifacts together in a form a reviewer can read in order. section 7.4.2 section 7.7
Build the pack alongside the system. Record design decisions when they are made, and collect and label artifacts as risk assessment, data governance, logging, human oversight, and monitoring take place. Otherwise, a team may have to reconstruct decisions from memory after people have rotated away. section 7.4.2 section 7.7
A provider’s evidence file can include:
- Risk assessments covering privacy, security, accuracy, bias, and overreliance. section 7.4.2
- Training and test data summaries describing sources, fitness for intended use, and bias testing. section 7.4.2
- System and decision logs, with retention and export capability checked. section 7.4.2
- Evidence of system behavior and controls, including robustness, fairness, safety, and red-teaming. section 7.4.2
Table 7.7 What goes in a provider’s evidence file
|
Topic
|
What the
book covers
|
What’s
still needed
|
|
Risk
assessment
|
Privacy and
data risks (Chapter 5); security threats (Chapter 4); ethical harms including
accuracy failures, bias or overreliance (Chapter 6); impact assessments
(Chapter 3)
|
Add it to
the evidence folder
|
|
Training and
test data: what you used, why, and how you checked it
|
Lawful
basis, data minimization, and provenance (Chapter 5); bias analysis, fairness
testing and Intellectual Property (Chapter 6)
|
A brief
summary of sources, whether it is fit for the intended use, and bias test
data
|
|
System logs
|
Audit
logging (Chapter 4); decision logging (Chapter 6)
|
Confirm logs
run for at least six months and that you can export them
|
|
How the
system behaves
|
Output
integrity and hallucination controls (Chapters 5 and 6); robustness, bias and
fairness, safety and red-teaming (Chapter 6)
|
File all
test results together, with the method used for each
|
|
Security
posture
|
Threat
model, adversarial testing, supply chain controls (Chapter 4)
|
File
alongside the other test results
|
|
Privacy
controls
|
Data
minimization, PII leakage prevention, deletion pipeline, DPIA (Chapter 5)
|
File the
DPIA with the risk assessment
|
|
Who oversees
the system and what they can do
|
Overreliance
controls and oversight design (Chapter 6); meaningful review standard
(Chapter 5)
|
Put names to
roles; write down who can override, halt, or escalate
|
|
IP and data
provenance
|
Data
register, copyright clearance, training-use-output IP (Section 6.7)
|
Add the data
register to the evidence file
|
|
Agentic
controls (if the system plans its own actions)
|
Scope
manifests, intervention points, cascading failure mitigations (Chapters 4, 5
and 6)
|
File
alongside the human oversight design
|
|
Instructions
for the team deploying your system
|
Purpose,
limits, and failure modes (Chapters 3-6)
|
Write them
as a standalone document another team can follow without calling you
|
|
How your
organization manages AI
|
The 6L-G
governance framework (Chapter 3)
|
A short
statement mapping your process to what the Act expects
|
|
Risk
classification decision
|
Classification
work from Section 7.2
|
Documentation
with the reasoning, categories considered, and dated sign-off from legal and
the technical lead
|
|
Documentation
from your model provider
|
Section
7.3.5
|
Add what the
provider delivered
|
|
Post-deployment
monitoring plan
|
Monitoring
infrastructure (Chapter 3)
|
Write a plan
with specific triggers and a response plan (Section 7.5.1)
|
|
EU
authorized representative (non-EU providers only)
|
Not in
earlier chapters
|
Sign a
mandate with someone in the EU who can speak to authorities for you
|
Use the provider evidence-file table to map governance practices to documentation and remaining work. section 7.4.2
The evidence supports a conformity assessment and the declaration that the system meets applicable requirements. Organizational frameworks such as ISO 42001 and the NIST AI RMF provide governance scaffolding, but do not replace product-level documentation for each AI system. section 7.4.2 section 7.7
Keep evidence current as the system changes. A new decision context or a vendor model update can change system behavior and potentially shift applicable obligations. Start with the system that would do the most damage if it failed. section 7.7
