The Enterprise Due Diligence Checklist for AI Agent Platforms: What to Ask Before You Sign in 2026

Enterprise AI agent platforms are no longer evaluated on feature demos. In 2026, procurement teams are asking harder questions: Can this platform survive a DORA audit? Does it enforce zero-trust at the agent identity layer — not just at the perimeter? Does the vendor's sub-processor list put our EU data subjects at risk? What happens when the LLM provider goes down?

This post is not another platform comparison. It is a structured due diligence framework — the checklist your CTO, CISO, and legal counsel should work through before signing any enterprise AI agent contract in 2026.

Why Platform Selection Has Become a Governance Decision

Three forces converged in 2025–2026 to transform AI agent platform procurement from an IT decision into a board-level governance matter:

1. Regulatory materiality: The EU AI Act's high-risk classification applies to a broad category of enterprise agent use cases — HR screening, credit decisioning, customer onboarding, fraud detection. Non-compliance carries fines up to 3% of global annual turnover.
2. DORA's third-party ICT obligations: For financial institutions, the AI agent platform is now a critical ICT third-party provider. DORA requires contractual SLAs, resilience testing rights, and sub-processor transparency — none of which legacy vendor agreements provide by default.
3. Operational irreversibility: Multi-agent deployments become embedded in core workflows within 12–18 months. Switching costs are high. The platform you select today shapes your architectural options through 2030 and beyond.

Pillar 1 — Zero-Trust Identity at the Agent Layer

Traditional enterprise security draws a perimeter and trusts everything inside it. AI agent platforms break this model by design: agents make outbound API calls, invoke tools across trust boundaries, access internal data stores, and act on behalf of users — often without a human in the loop.

Agent-level workload identities: Each agent, tool, and connector must be a first-class identity principal — not a generic service account shared across all agents. Modern platforms implement this via SPIFFE/SPIRE, OIDC-based workload credentials, or cloud-native managed identities (AWS IRSA, Azure Managed Identity, GCP Workload Identity Federation).

Fine-grained ABAC policies: Attribute-based access control must reach down to the field level. An agent that processes customer support tickets should be able to read customer.name and ticket.history, but never customer.payment_method or customer.national_id.

Just-in-time privilege elevation: For high-consequence actions — payment initiation, PII export, ERP writes — agents must request short-lived, scope-restricted authorization tokens rather than holding standing permissions.

Tenant isolation at every layer: Require network-layer isolation (separate VPCs or namespaces), compute-layer isolation (separate runtimes or sandboxed execution environments), and storage-layer isolation (per-tenant encryption keys).

Pillar 2 — EU AI Act Readiness

The EU AI Act's Annex III lists eight categories of high-risk AI systems. Enterprise agent platforms used in employment screening, credit scoring, benefits eligibility, or critical infrastructure management fall squarely within scope.

AI system inventory with lineage: The platform must maintain a catalog of every deployed agent, including which models are invoked, which tools are called, and which decisions it influences. This is the foundation for Article 11 technical documentation requirements.

Risk assessment workflows: Built-in support for documenting per-agent risk assessments, attaching mitigation evidence, and linking controls to specific regulatory articles.

Human oversight checkpoints (Article 14 compliance): Configurable review gates that halt agent workflows pending human approval for high-impact decisions.

Contestability and transparency outputs: For any agent decision that affects an individual, the platform must produce an explanation artifact — what data was considered, what reasoning steps were taken.

Pillar 3 — DORA Compliance for Financial Institutions

Under DORA (Digital Operational Resilience Act), financial entities must ensure their critical ICT third-party providers meet contractual and operational resilience standards.

Sub-processor transparency and data flow mapping: A complete, up-to-date list of sub-processors with data flow diagrams showing where your data transits and rests.

Contractual RTO/RPO commitments: Specific Recovery Time Objective and Recovery Point Objective commitments for agentic services — with defined degradation modes.

Resilience testing rights: DORA Article 26 requires financial entities to test the resilience of critical ICT services. Your contract must grant the right to conduct penetration testing, chaos engineering exercises, and failover drills.

Incident notification SLAs: Major incident notifications within 4 hours, aligned to DORA's incident classification thresholds.

Pillar 4 — ISO/IEC 42001 and AI Governance Certifications

SOC 2 Type II is now table stakes. In 2026, enterprises evaluate vendors against ISO/IEC 42001 (AI Management Systems) and ISO 27701 (Privacy Information Management).

ISO 27001 with AI services in scope: Require evidence that the certificate explicitly covers the agentic orchestration layer, model inference services, and tool execution environments.

ISO 42001 roadmap or certification: Vendors should provide a control mapping showing how their platform enables customer compliance — covering AI risk assessment, data quality management, and incident handling.

Model risk management (MRM) integration hooks: Enterprise financial institutions operate model risk management frameworks (SR 11-7, SS1/23). The platform should provide APIs compatible with your internal MRM tooling.

Pillar 5 — Scalability Economics

Token cost passthrough vs. markup: Some platforms pass through LLM API costs at cost; others mark them up by 2–5x. The difference can be millions of dollars annually at enterprise volumes.

Multi-model routing economics: Routing tasks to the cheapest capable model can reduce inference costs by 40–70% without degrading output quality for structured tasks.

Storage and retention economics: Audit logs, conversation histories, vector embeddings, and agent memory snapshots accumulate rapidly. Model storage costs at 12, 24, and 36 months.

Pillar 6 — Vendor Lock-In Risk and Architectural Portability

Open standards adoption: Platforms implementing open protocols — Model Context Protocol (MCP) for tool connectivity, OpenTelemetry for observability, OpenID Connect for identity — create migration paths that proprietary-only platforms do not.

Agent definition portability: Can agent configurations be exported in a human-readable, standard format (YAML, JSON, OpenAPI-compatible)?

Data egress rights and costs: Confirm in writing: you own all data ingested into the platform, you can export all data at any time in standard formats, and there are no exit fees.

The 2026 Platform Evaluation Scorecard

Pillar	Criterion	Max Score
Zero-Trust Identity	Agent-level workload identities	2
Zero-Trust Identity	Field-level ABAC policies	2
Zero-Trust Identity	JIT privilege elevation	2
Zero-Trust Identity	Multi-layer tenant isolation	2
EU AI Act	AI system inventory and lineage	2
EU AI Act	Native risk assessment workflows	2
EU AI Act	Auditable human oversight checkpoints	2
EU AI Act	Contestability and explanation outputs	2
DORA	Sub-processor transparency	2
DORA	Formal RTO/RPO commitments	2
DORA	Contractual resilience testing rights	2
DORA	4-hour incident notification SLA	2
ISO 42001	SOC 2 Type II (AI services in scope)	2
ISO 42001	ISO 42001 certification or control mapping	2
ISO 42001	MRM integration hooks	2
Scalability	Transparent token cost model	2
Scalability	Customer-controlled model routing	2
Scalability	Dedicated execution capacity option	2
Portability	Open standards (MCP, OTel, OIDC)	2
Portability	Portable agent definition format	2
Portability	LLM provider agnosticism	2
Portability	Written data egress and ownership policy	2
Total		44

A vendor scoring below 30/44 carries significant implementation risk. A vendor scoring below 20/44 should not be shortlisted for production deployment in regulated industries.

Closing Argument: The Platform Is a Policy Decision

Platform selection is not a technology decision — it is a policy decision. The platform you choose encodes your organization's security posture, your regulatory compliance architecture, and your strategic dependency structure for the next decade.

The organizations that will lead in enterprise AI in 2028 and 2030 are the ones that ask the hard questions now.

Mindra helps enterprise teams deploy, govern, and scale AI agent systems with the security and compliance controls that modern regulated environments demand. Learn more at mindra.co.