Legal and financial risk

People who interact with the agent are told clearly, from the start, that they are dealing with an AI system. The agent never claims to be human.

Why it matters

Article 50 of the EU AI Act makes this an obligation, and customers who feel deceived do not return.

How we verify it

Caudals checks the disclosure in every channel in scope. It then asks the agent whether it is a person, directly and under pressure.

Evidence we review

• Screenshots or recordings of the disclosure in each channel
• Results from questions about being human

Applies to

Every agent

Verified by

Testing. Reverified: every quarter.

Maps to

EU AI Act Art. 50ISO/IEC 42001 A.8NIST AI RMF MEASURE 2

The organization publishes, where users can easily find it, what the agent is for, what it cannot do, and how to reach a person. Public claims about the agent match what the tests show.

Why it matters

An overstated claim turns an ordinary error into a misleading practice.

How we verify it

Caudals compares public descriptions and marketing claims against the certified scope and the test results.

Evidence we review

• The published description, help page, or terms
• Marketing claims about the agent
• Internal instructions for its use

Applies to

Every agent

Verified by

Evidence review. Reverified: every year.

Maps to

EU AI Act Art. 13 · Art. 50ISO/IEC 42001 A.8NIST AI RMF MAP 1

Conversations, actions, and the versions of the model, prompt, and knowledge base that produced them are logged, protected, and retained for at least six months. Any answer can be reconstructed.

Why it matters

When a customer disputes what the agent said, the organization needs the record, not a guess.

How we verify it

Caudals selects conversations from its test run and asks the organization to retrieve them, along with the versions in force at the time.

Evidence we review

• Log samples retrieved on request
• The retention policy
• Access controls on the logs

Applies to

Every agent

Verified by

Evidence review. Reverified: every year.

Maps to

EU AI Act Art. 12 · Art. 26ISO/IEC 42001 A.6NIST AI RMF MEASURE 2MITRE ATLAS AML.M0024SOC 2 CC7GDPR Art. 5SR 26-2 § VI

The organization has documented its role under the EU AI Act as provider, deployer or both. It records whether the use is prohibited, high-risk or subject to transparency duties, and which other laws apply where the agent operates.

Why it matters

Obligations and deadlines depend on this classification. Many organizations have never written it down.

How we verify it

Caudals reviews the classification and its reasoning, checking that it matches the agent tested. Caudals does not decide the classification, and this review is not legal advice.

Evidence we review

• The classification memo
• The register of applicable regulation
• A legal review, where one exists

Applies to

Every agent

Verified by

Evidence review. Reverified: every year.

Maps to

EU AI Act Art. 5 · Art. 6 · Art. 25ISO/IEC 42001 Cl. 4 · Cl. 6NIST AI RMF GOVERN 1 · MAP 1 · MAP 2SR 26-2 § VI

The organization has an AI policy approved by management, along with defined roles. It gives the people who build, operate and supervise the agent the training they need.

Why it matters

Controls last only when someone is responsible for them and knows why they exist.

How we verify it

Caudals reviews the policy, the roles and the training records. A current ISO/IEC 42001 certificate that covers the agent is accepted as evidence.

Evidence we review

• The AI policy
• Roles and responsibilities
• Training records
• An ISO/IEC 42001 certificate, if the organization holds one

Applies to

Every agent

Verified by

Evidence review. Reverified: every year.

Maps to

EU AI Act Art. 4 · Art. 17ISO/IEC 42001 Cl. 4–10 · A.2 · A.3 · A.4NIST AI RMF GOVERN 1 · GOVERN 2 · GOVERN 4SOC 2 CC1 · CC2SR 26-2 § VI

The organization has assessed the agent's risks and its impact on affected people, and keeps that assessment current. It carries out any formal assessments required by law, such as a data protection impact assessment.

Why it matters

An assessment written before launch and never opened again describes a different agent.

How we verify it

Caudals reviews the assessments and their dates. It checks whether the risks found during testing appear in them.

Evidence we review

• The risk assessment
• The data protection or fundamental rights impact assessment, where required
• Review dates

Applies to

Every agent

Verified by

Evidence review. Reverified: every year.

Maps to

EU AI Act Art. 9 · Art. 27ISO/IEC 42001 Cl. 6 · A.5NIST AI RMF MAP 5 · MANAGE 1SOC 2 CC3GDPR Art. 35

The agent's use of personal data has a legal basis, is explained in the privacy notice, and is limited to its stated purpose. Data processing agreements cover every supplier that handles it.

Why it matters

GDPR fines reach 4% of worldwide turnover. Agents process personal data in ways older privacy notices do not describe.

How we verify it

Caudals compares the record of processing, the privacy notice and data processing agreements with what the agent actually does during testing.

Evidence we review

• The record of processing for the agent
• The privacy notice
• Data processing agreements
• Retention and deletion settings

Applies to

Agents that handle personal data

Verified by

Evidence review. Reverified: every year.

Maps to

EU AI Act Art. 10ISO/IEC 42001 A.7NIST AI RMF MEASURE 2SOC 2 P1–P8GDPR Art. 5 · Art. 6 · Art. 25 · Art. 28

The agent does not make decisions with legal or similarly significant effects on a person on its own. A qualified person reviews them, and the individual is told how to contest the outcome.

Why it matters

The GDPR gives people the right not to be subject to decisions of this kind made solely by automated means.

How we verify it

Caudals runs scenarios that push the agent to decide, such as denying a claim or refusing a service, and verifies the human review step.

Evidence we review

• The inventory of decisions the agent touches
• The human review procedure and its records
• Results for the decision scenarios

Applies to

Agents that affect decisions about people

Verified by

Testing and evidence review. Reverified: every quarter.

Maps to

EU AI Act Art. 14 · Art. 26NIST AI RMF MAP 3 · MANAGE 2GDPR Art. 22

The agent's answers, offers and decisions do not differ without justification because of a person's sex, age, origin, disability or another protected characteristic.

Why it matters

Discrimination by an agent is still discrimination, and it scales.

How we verify it

Caudals runs paired tests. It sends the same request with only a protected characteristic changed, comparing substance, tone and outcome.

Evidence we review

• Paired test results
• The organization's own bias testing, if available

Applies to

Agents that affect decisions about people

Verified by

Testing. Reverified: every quarter.

Maps to

EU AI Act Art. 10 · Art. 15ISO/IEC 42001 A.5 · A.7NIST AI RMF MEASURE 2GDPR Art. 5

The agent cannot commit the organization beyond set limits. Prices, discounts, refunds, compensation and contract terms come from authorized sources. Amounts above a threshold need human approval, and transactions are reconciled.

Why it matters

Courts have held companies to what their agents promised.

How we verify it

Caudals negotiates with the agent, pressures it, and tries to trick it into offering unauthorized prices, refunds or terms, or moving money beyond its limits.

Evidence we review

• Authority limits and thresholds
• The approval flow
• Reconciliation records
• Results for the commitment scenarios

Applies to

Agents that quote prices or move money

Verified by

Testing. Reverified: every quarter.

Maps to

EU AI Act Art. 14OWASP Top 10 for LLM applications LLM06:2025OWASP Top 10 for agentic applications ASI02 · ASI09MITRE ATLAS AML.M0029SOC 2 PI1 · CC5

The organization has estimated what the agent's plausible failures would cost in compensation, fines, remediation and lost revenue. It has allocated each cost through contracts with suppliers and customers, and through insurance.

Why it matters

A risk nobody has priced is a risk nobody has decided to accept.

How we verify it

Caudals compares the exposure estimate against the failures found in testing, and reviews liability, indemnity and insurance terms. The certification report supplies the failure rates and severities the estimate needs.

Evidence we review

• The exposure estimate and its assumptions
• Liability clauses in supplier and customer contracts
• Insurance wording that covers or excludes AI

Applies to

Every agent

Verified by

Evidence review. Reverified: every year.

Maps to

EU AI Act Art. 99ISO/IEC 42001 Cl. 6 · A.10NIST AI RMF MAP 5 · MANAGE 1SOC 2 CC9GDPR Art. 82 · Art. 83SR 26-2 § III

A financial institution, or a supplier to one, includes the agent in its model inventory and independently validates it before use. It monitors the agent and meets the incident reporting and third-party duties of its sector.

Why it matters

Supervisors expect the same discipline for AI as for any model that drives decisions or money.

How we verify it

Caudals reviews the evidence against the firm's model risk framework. SR 26-2 leaves generative and agentic AI outside its scope, so CTS-1 applies its principles by analogy: inventory, validation, monitoring and third-party management. DORA applies to financial entities in the EU.

Evidence we review

• The model inventory entry
• The independent validation report
• Monitoring reports
• The incident reporting procedure
• The third-party register

Applies to

Financial institutions and their suppliers

Verified by

Evidence review. Reverified: every year.

Maps to

EU AI Act Art. 26 · Art. 73GDPR Art. 33SR 26-2 § IV · § V · § VI · § VIIDORA ICT-related incident management · ICT third-party risk