How we assess, how we stay independent, what the seal means
Everything a buyer needs to trust a SoAITrust result: the method, the scoring, the evidence rules, the conflicts we avoid and the rules of the seal. Public, versioned, and the same for every vendor.
What we assess
The assessment covers one AI system at a time, in the vendor's real environment, across four blocks.
- Environment security: 26 controls in 6 domains (security governance, identity and access, data protection and privacy, application and model security, infrastructure and operations, incidents and continuity). References: ISO/IEC 27001:2022, NIST CSF 2.0, OWASP Top 10 for LLM, MITRE ATLAS, LGPD/GDPR.
- AI governance: 21 controls in 6 domains (leadership and policy, inventory and classification, risk and impact, data and model lifecycle, transparency and human oversight, monitoring and improvement). References: ISO/IEC 42001:2023, NIST AI RMF 1.0, EU AI Act, ISO/IEC 23894.
- Risks and criticality: 14 AI risks scored for likelihood and impact on a 5×5 matrix, each linked to the gaps that support it, and a criticality score from 0 to 100.
- Due diligence report: verdict, findings P1 to P3 and an action plan with deadlines.
The full list of controls, with the crosswalk to the EU AI Act, SIG and CSA CCM, is public on the crosswalk page.
How we score
- Each control is answered Yes (2 points), Partial (1), No (0) or Not applicable (excluded). The score of a domain and of a block is the percentage of points obtained over the maximum.
- Levels: Initial (0 to 20), Basic (21 to 40), Defined (41 to 60), Managed (61 to 80), Optimized (81 to 100).
- A "Yes" only counts after the evidence is reviewed by a SoAITrust analyst. Without evidence the answer is recorded as Partial at most.
- ISO certifications are verified at the source in the IAF CertSearch database. A certificate that cannot be verified does not count.
- Risk likelihood is suggested from the gaps and adjusted by the analyst; impact depends on the system's data and decisions. The criticality band drives the recommended tier and the validity conditions.
Evidence rules
- Accepted: approved policies, procedures, screenshots of configurations, audit and pentest reports, contracts and DPAs, tickets, logs and dashboards. Dated and attributable.
- Not accepted alone: verbal statements, marketing pages, drafts, evidence older than 12 months for controls that require recurrence.
- AI triage flags whether an evidence item appears to cover the control. The decision is always human.
- Evidence is stored encrypted, accessible only to the SoAITrust analysts assigned to the assessment, deleted 24 months after the seal expires, or earlier on request.
Tiers, proportional to risk
Self-service
Low-risk systems. Questionnaire on the platform, automated triage, sample human review.
Assisted diagnostic
Personal data or regulated customers. Mandatory evidence, full human review, source verification, gap report.
Full due diligence
Credit, fraud, automated decisions, sensitive data. 15-to-20-page report, interviews, contract and privacy analysis, action plan.
Fees are fixed per tier and agreed before kickoff. Request a proposal at soaitrust.com/contact.
Independence statement
- SoAITrust does not implement, sell or resell the controls it assesses, and does not provide remediation consulting to a vendor during the validity of its seal.
- The result does not depend on the commercial relationship. Fees are fixed per tier, agreed before kickoff, and are never conditional on the outcome.
- Analysts declare conflicts of interest before each assessment. A vendor in which an analyst holds an interest is assessed by someone else or not at all.
- The methodology is public and versioned. Changes are announced with 60 days' notice and never applied retroactively to a valid seal.
- The vendor sees the full result before the seal is issued and can contest findings with evidence within 10 business days.
Seal policy
- Meaning: the seal states that the named AI system was assessed by SoAITrust with this methodology, with evidence reviewed. It is not a certification and does not guarantee the absence of incidents.
- Wording: "Assessed by SoAITrust". Never "certified", "approved" or "compliant".
- Validity: 12 months from the assessment date, or until a relevant change in the system (new model provider, new data categories, new automated decisions).
- Verification: every seal has a unique code checked at soaitrust.com/verify, showing vendor, system, scope, criticality band, date and validity.
- Use: on the vendor's website, proposals and questionnaires, always with the code and next to the assessed system's name. Not on systems outside the scope.
- Suspension and revocation: the seal is suspended when a relevant change is not reported, when evidence turns out to be false or when a P1 finding is not treated within the agreed deadline; it is revoked after 30 days without resolution. The verification page shows the status.
- Renewal: a simplified reassessment focused on what changed, for vendors that renew before expiry.
Limitations
The assessment is documentary and technical, performed remotely on the evidence provided and on verifiable public sources. It is not a penetration test, not an on-site audit and not a legal opinion. It reflects the system at the assessment date. Buyers remain responsible for their own risk decisions.
Methodology v1.0 · September 2026 · Questions: [email protected]
Frequently asked questions
Is the SoAITrust seal a certification?
No. The seal states that the named AI system was assessed by SoAITrust with a public methodology and reviewed evidence. It is not a certification and does not guarantee the absence of incidents. The correct wording is 'assessed by SoAITrust'.
How long does an assessment take?
Tier 1 takes about 2 weeks, Tier 2 about 4 weeks and Tier 3 about 6 weeks, counted from the kickoff. The vendor's effort is 6 to 8 hours of one technical person over the period.
Which frameworks is the methodology based on?
The methodology is aligned with NIST AI RMF 1.0 and based on ISO/IEC 27001 and ISO/IEC 42001, with references to OWASP Top 10 for LLM, MITRE ATLAS, ISO/IEC 23894, the EU AI Act, LGPD and GDPR. A public crosswalk maps every control to the EU AI Act, SIG and CSA CCM/CAIQ.
Does SoAITrust run penetration tests?
No. The assessment is documentary and technical, performed remotely on evidence and verifiable public sources. It does not include penetration testing, on-site audits or access to production environments.
How is a certification verified at the source?
ISO certifications are checked in the IAF CertSearch database, the registry of accredited certification bodies, instead of accepting the PDF sent by the vendor. A certificate that cannot be found there does not count.
How long is the seal valid and how is it verified?
The seal is valid for 12 months or until a relevant change in the system. Anyone can verify it by code at soaitrust.com/verify, which shows vendor, system, scope, criticality band, date and status.