- Enterprise AI procurement is in the early stages of a structural shift. AI governance certification, previously a voluntary differentiator, is becoming a threshold qualification for vendor approval in regulated sectors. The first movers are financial services enterprises under DORA, followed by healthcare and public sector under EU AI Act Annex III obligations.
- The documentation that enterprise procurement teams require from AI vendors maps directly to the evidence that EU AI Act compliance, DORA third-party risk management, and AI insurance underwriting each require independently. A vendor who builds this evidence package once can reuse it across three separate commercial and regulatory processes.
- ISO/IEC 42001:2023 is becoming the organisation-level governance threshold. But ISO 42001 certifies a management system, not individual AI systems. Procurement teams that have moved past the ISO 42001 threshold are now asking for system-specific documentation: autonomy envelope, testing methodology, human oversight design, and incident response. Agent Certified addresses this second layer.
- Financial services procurement under DORA is the most advanced. DORA Article 28 requires institutions to maintain a register of all ICT third-party service providers, with a risk assessment for each. AI tools are ICT services. The documentation requirements for that risk assessment are generating specific vendor qualification criteria that banks and investment firms are now enforcing.
- The certification-to-coverage pathway creates a compounding advantage for pre-certified vendors. A vendor with a documented Agent Certified assessment is in a stronger position for AI insurance underwriting as well as procurement qualification, because both require the same underlying evidence of governance, testing, and oversight design.
The structural shift in enterprise AI procurement
Until mid-2025, AI governance certification in enterprise procurement was primarily a differentiator: a vendor who could produce evidence of structured AI governance stood out from competitors who could not. By mid-2026, the dynamic in regulated sectors has begun to shift. Threshold qualification rather than differentiation is the operative frame for financial services, healthcare, and public sector buyers who have been pushed by DORA, the EU AI Act, and the revised Product Liability Directive to formalise their AI third-party risk management.
The mechanism driving this shift is simple. An enterprise deploying an AI vendor's product in a regulated context becomes the deployer of that system under the EU AI Act and the ICT services customer under DORA. The enterprise's own regulatory obligations require it to verify the AI system's governance, testing, and oversight arrangements before deployment. If the vendor cannot provide that evidence, the enterprise faces a compliance gap it cannot fill from its own side. Refusing to approve vendors who cannot document their AI governance is not virtue signalling: it is the only rational response to an enterprise's own compliance obligations flowing back into vendor selection.
This dynamic is most mature in financial services. It is approximately twelve to eighteen months behind in healthcare and the public sector. It is emerging, but not yet formalised, in enterprise SaaS and technology procurement.
Financial services: DORA as the procurement driver
DORA, Regulation (EU) 2022/2554 on digital operational resilience for the financial sector, applied from 17 January 2025. It requires banks, investment firms, insurance companies, and other financial entities within its scope to implement a comprehensive ICT risk management framework that includes specific requirements for third-party ICT service providers.
Article 28 of DORA requires financial entities to maintain a register of all ICT third-party service providers with a risk assessment for each, distinguishing between providers of critical or important functions and other providers. AI tools used in credit decisioning, risk modelling, fraud detection, algorithmic trading, and customer service in financial services are ICT services under DORA. A bank that deploys a vendor's AI credit decisioning model must maintain a documented risk assessment of that vendor and that model as part of its DORA compliance.
The risk assessment content DORA requires maps to the documentation AI governance certification produces. The European Banking Authority's guidelines on ICT and security risk management under DORA specify that risk assessments should cover the provider's security arrangements, data governance, business continuity, change management, and exit procedures. For AI-specific tools, banks are extending this assessment to cover the AI system's testing methodology, known failure modes and mitigations, data provenance, and human oversight design.
The practical effect is a DORA-driven vendor qualification questionnaire that is, in substance, an AI governance assessment. Banks are sending these questionnaires to AI vendors before renewal and before new procurement decisions. Vendors who cannot answer them from pre-existing documentation face a disproportionate burden of ad hoc documentation work for each customer. Vendors who have completed a structured AI governance assessment against a recognised framework can provide a single reference document that answers the questionnaire with material reuse.
The European Central Bank's supervisory priorities for 2026 include ICT and operational resilience as a primary focus area, with specific attention to third-party risk and AI governance. ECB-supervised banks can therefore expect active supervisory inquiry into how they manage AI vendor risk, which reinforces the procurement pressure that DORA creates independently. For detailed analysis of how EU AI Act operator obligations interact with coverage, see the Article 26 deployer obligations guide at agentliability.eu.
Healthcare: EU AI Act Annex III as the procurement driver
Healthcare is the sector with the most clearly defined AI risk categories under the EU AI Act. Annex III point 5 of Regulation (EU) 2024/1689 classifies AI systems intended to be used as safety components of medical devices, AI intended for medical triage in emergency rooms, and AI intended to assist in determining treatment decisions as high-risk AI systems subject to the full Article 26 deployer obligation set and, for providers, the full Chapter III obligation set including conformity assessment.
Healthcare procurement of AI is therefore not just a commercial decision but a regulatory one. A hospital or health system deploying an AI clinical decision support tool that is classified as a high-risk medical device AI is the deployer of a high-risk AI system. Its Article 26 obligations, including use in accordance with provider instructions, human oversight implementation, and serious incident reporting, cannot be satisfied unless the AI vendor provides the provider-level technical documentation and instructions for use that Article 13 and Annex IV require.
Healthcare procurement teams are responding to this obligation by adding AI Act compliance evidence requests to their technology vendor assessments. The documentation they request includes technical documentation of the AI system's intended purpose and performance characteristics, validation studies and accuracy benchmarks on the relevant clinical population, human oversight design specifications, and incident response procedures. These requirements align with what the Medical Device Regulation and the In Vitro Diagnostic Medical Devices Regulation already require for software classified as medical devices, which creates a documentation overlap that experienced healthcare AI vendors are beginning to manage with reusable evidence packages.
The revised EU Product Liability Directive (Directive 2024/2853), which applies from 9 December 2026, adds a strict liability dimension to healthcare AI procurement. Under Article 8 of the Directive, multiple economic operators in the supply chain can be jointly and severally liable for damage caused by a defective product including AI software. A healthcare provider that procures and deploys a vendor's AI system that later causes harm to a patient is jointly liable with the vendor for that harm regardless of fault. Healthcare procurement teams are using this joint liability exposure as an additional justification for requiring vendor certification before procurement approval.
Public sector: EU AI Act high-risk categories and procurement obligations
Public sector procurement of AI faces the highest concentration of Annex III high-risk categories in the EU AI Act. AI used in critical infrastructure (point 2), AI for education and vocational training decisions (point 3), AI in employment and worker management (point 4), AI in access to essential public services and benefits (point 5), and AI in law enforcement, migration, and justice administration (points 6, 7, and 8) are all Annex III categories with a high public sector prevalence.
Public sector deployers of these categories face the full Article 26 obligation set. They also face the Article 26(9) fundamental rights impact assessment obligation, which requires a FRIA before deploying high-risk AI in the specific contexts listed in that provision, several of which are characteristic public sector deployments: employment decisions, access to essential public services, and biometric categorisation.
European public procurement law is beginning to formalise these requirements into tender specifications. Several member state public procurement bodies have issued guidance recommending or requiring AI governance certification evidence from AI vendors tendering for public contracts involving Annex III high-risk AI. The Netherlands' National AI Strategy and the German Federal Government's AI Action Plan both reference AI governance certification as a procurement quality criterion. In France, the ANSSI (Agence nationale de la securite des systemes d'information) has published guidance for public sector technology procurement that addresses AI security requirements aligned with NIST AI RMF controls.
What documentation enterprise procurement teams require from AI vendors
Based on procurement questionnaires currently in circulation in financial services and healthcare, the documentation typically required from AI vendors includes seven categories of evidence. Understanding these categories allows vendors to structure a reusable evidence package rather than responding to each procurement request from scratch.
The first category is system-level technical documentation. This describes what the AI system does, the data it was trained on and its provenance, the testing methodology used to validate performance, accuracy benchmarks achieved in relevant test populations, and known limitations and failure modes. This corresponds directly to Annex IV of the EU AI Act's technical documentation requirements for high-risk AI systems.
The second category is risk management documentation. This covers the risks the AI system presents, the probability and severity of those risks, and the mitigations in place. A risk management summary structured around NIST AI RMF categories (Govern, Map, Measure, Manage) or ISO/IEC 42001 controls is increasingly acceptable to procurement teams familiar with those frameworks.
The third category is governance and accountability. This covers who in the vendor organisation is responsible for the AI system's ongoing performance, what policies govern its use, what audit trail exists for decisions about the system's development and deployment, and how the vendor's board is informed of material AI risks.
The fourth category is human oversight design. This specifies what the AI system does autonomously without human review, what triggers human oversight, and who in the deploying organisation will be assigned oversight responsibility. This category maps directly to Article 14 and Article 26(2) of the EU AI Act and is the category procurement teams find most difficult to assess without a structured framework reference.
The fifth category is incident response procedures. This covers how the vendor detects failures in the AI system's operation, how failures are communicated to deployers, and how remediation is managed. Post-market monitoring obligations under Article 72 of the EU AI Act and serious incident reporting under Article 73 are the regulatory drivers for this category.
The sixth category is data protection and privacy. This covers how the AI system handles personal data, what data subjects can access, and how data minimisation is implemented. This category is driven by GDPR obligations and intersects with Article 10 of the EU AI Act on data governance.
The seventh category is third-party assessment evidence. Increasingly, procurement teams are requiring evidence of external validation rather than vendor self-attestation for the first six categories. ISO/IEC 42001 certification covers the organisation-level governance evidence. System-specific frameworks such as Agent Certified and AIUC-1 cover the system-level technical and oversight evidence. A vendor who can provide both organisation-level and system-level third-party assessment evidence is in the strongest procurement position.
How Agent Certified maps to procurement documentation requirements
Agent Certified's seven-dimension framework produces evidence that maps directly to the procurement documentation categories above. The mapping is not incidental: the framework was designed to produce the evidence that insurers, regulators, and enterprise counterparties each need, which overlap substantially because all three are addressing the same underlying question of whether an AI system is safe to rely on.
The Trust and Safety dimension generates evidence on guardrails, adversarial robustness, incident playbooks, and the kill switch, which feeds the risk management documentation, incident response, and technical documentation categories. The Governance dimension generates evidence on accountability structures, risk policy, and audit trails, which feeds the governance and accountability category and maps to ISO 42001 Clause 5 (Leadership) and Clause 9 (Performance evaluation). The Autonomy Envelope dimension generates evidence on human oversight design, which feeds the human oversight category and maps to EU AI Act Articles 14 and 26(2). The Context Integrity dimension generates evidence on data governance and training data provenance, which feeds the data protection category and maps to EU AI Act Article 10.
An Agent Certified assessment report structured around these seven dimensions provides a single reference document that addresses six of the seven procurement documentation categories with a consistent evidence base. The seventh category, third-party assessment evidence, is the assessment report itself. For a full description of the assessment process and what the report contains, see the request assessment page and the full methodology.
The connection to insurance coverage amplifies the commercial value of pre-certification. Munich Re's aiSure product and Armilla's Lloyd's-backed AI liability coverage both use pre-underwriting assessment processes that draw on similar evidence to what procurement teams require. A vendor who has completed Agent Certified certification has documentation that serves procurement qualification, EU AI Act compliance evidence, and insurance underwriting evidence simultaneously. For the link between certification evidence and insurance coverage, see the certification and insurance eligibility analysis at agentinsured.eu.