Assessing AI Agent Providers: A Questionnaire for NIS2, DORA and GDPR Article 28

What the law requires

Regulation	Requirement	Article
GDPR	Use only processors that give sufficient guarantees of appropriate technical and organisational measures, and sign a contract with the mandatory terms, including sub-processor approval, assistance and deletion at the end	Art. 28(1)-(4)
GDPR	A transfer tool for any processing outside the EEA	Arts. 44-49
NIS2	Supply chain security as a minimum measure; take into account each supplier's specific vulnerabilities, the quality of its products and cybersecurity practices, including secure development	Art. 21(2)(d), 21(3)
DORA	Before contracting: assess whether the service supports a critical or important function, identify all relevant risks including concentration risk, and carry out due diligence on the provider	Art. 28(4)
DORA	Record the arrangement in the register of information; contract only providers that comply with appropriate information security standards; exit strategies for critical or important functions	Art. 28(3), (5), (8)
DORA	Mandatory contract clauses, with extra clauses when the service supports critical or important functions	Art. 30(2), (3)

Agents are different: a normal SaaS tool stores your data. An agent also takes actions with your credentials. Add questions about what it can do on its own, not only about where the data is.

The questionnaire

Send these 24 questions before signing, and keep the answers as evidence of your due diligence (GDPR Art. 5(2) accountability, DORA Art. 28(4)(d)). The Data use and Model questions also cover the check EDPB Opinion 28/2024 expects from a company deploying someone else's model: that it was not developed by unlawfully processing personal data.

Area	Question
Data use	Is our data (prompts, files, outputs) used to train or improve your models? Can we switch that off by contract?
Data use	Which personal data categories does the service process, and for which purposes?
Data use	How long are prompts, outputs, tool-call logs and agent memory kept, and can we set or shorten that?
Location	Where is data processed and stored, including by the model provider behind your service?
Location	Which transfer tool covers processing outside the EEA (adequacy decision, SCCs, BCRs)?
Sub-processors	Which sub-processors, including model providers, do you use, and how do you notify us of changes?
Security	Which certifications or reports cover the service (ISO/IEC 27001, SOC 2), and what is in their scope?
Security	How do you test the agent against prompt injection and data leakage, and how often?
Security	How are credentials for our systems stored, and are tokens scoped to each tool and tenant?
Agent behaviour	Which actions can the agent take without human approval, and can we require approval per action type?
Agent behaviour	Can we restrict the tools, data sources and outbound destinations the agent uses?
Agent behaviour	Do you log every tool call, and can we export those logs to our monitoring?
Agent behaviour	Is there a way for us to stop the agent and revoke its access at once?
Model	How was the model's training data obtained, and has any authority or court found it unlawful?
Model	Which model versions do you use, and how do you notify us before changing them?
Incidents	Within how many hours do you notify us of a security incident or personal data breach, and with what information?
Incidents	Will you support our own reporting deadlines (GDPR 72 hours, NIS2 24 hours, DORA 4 hours after classification)?
Continuity	What are the service levels, and what happens to our agents if the model provider is unavailable?
Continuity	How do we export our configuration and data, and how is data deleted at the end of the contract?
Audit	What audit and information rights do we have, including for our regulators?
AI Act	Are you the provider of the AI system under the AI Act? Which instructions for use and documentation do you supply?
AI Act	Is any part of the service intended for a high-risk purpose listed in Annex III (for example recruitment or credit scoring)?
AI Act	How do you meet the transparency obligations of Article 50 (AI interaction notices, marking of generated content)?
Legal	Will you sign a data processing agreement with all Article 28(3) terms, and, for financial entities, the DORA Article 30 clauses?

Red flags

Training on your data by default, with no contractual opt-out.
No list of sub-processors, or a model provider outside the EEA without a transfer tool.
The agent can send email, payments or data out without an approval step you control.
No incident notification time in the contract, or one longer than your own reporting deadline.
No way to export logs or to switch the agent off.
For financial entities: refusal to accept the DORA Article 30 clauses or audit rights.

After the contract

Record the provider in your vendor register; for DORA entities, in the register of information.
Configure the agent with the least tools and data it needs, and switch on human approval for outbound actions.
Review the provider at least yearly, and whenever the model, sub-processors or data location change.

Ready-made documents. The Vendor Security Assessment Pack includes a vendor questionnaire, a risk register and a data processing agreement checklist; our free Vendor Risk Assessment scores a provider and suggests contract clauses.

Assessing AI Agent Providers: A Questionnaire for NIS2, DORA and GDPR Article 28

What the law requires

The questionnaire

Red flags

After the contract

Sources

Related tools

Vendor Risk Assessment

DORA Compliance Check