A vendor-neutral checklist to evaluate on-prem AI agent platforms, air-gapped AI deployment, private RAG, sovereign AI infrastructure, BYOM enterprise AI, governance, security, compliance evidence, and predictable pricing. Tick items as you go — your progress is saved in your browser.
Checked items are saved locally. Nothing is uploaded.
These answer-first prompts are written for security, architecture, procurement, and compliance teams comparing cloud-only, hybrid, on-premises, and air-gapped AI agent platforms.
Start with deployment architecture, data residency, model governance, private RAG, auditability, and cost predictability. If a vendor cannot explain where prompts, documents, embeddings, model outputs, and logs go, pause the evaluation.
The strongest answer is an architecture where the AI data plane runs on-premises, in your private cloud tenancy, or air-gapped, with no required external inference API and customer-managed identity, keys, models, logs, and storage.
Score vendors by evidence, not promises: diagrams, control mappings, sample audit logs, subprocessor clarity, offline update process, private RAG design, model inventory, and support for EU AI Act, GDPR, DORA, HIPAA, and NIS2 programs.
Document whether the platform must run on-premises, in a VPC, in sovereign cloud, or fully air-gapped. This decides which vendors are viable before feature demos begin.
Ask for reference architectures, data-flow diagrams, audit samples, model-governance examples, subprocessor notes, and patch/update procedures.
Use a workflow with real access-control, retrieval, approval, and logging needs. A generic chatbot demo will not reveal production risk.
Estimate inference, storage, vector indexing, seats, connectors, environments, professional services, and support at production volume.
Full platform runs without outbound internet; updates arrive as inspected offline artifacts.
Embeddings, indexes, citations, and permission checks stay inside customer infrastructure.
Customer can approve, run, route, evaluate, and retire models under one governance layer.
External inference APIs are optional, policy-controlled, and can be fully disabled.
Runtime produces inventory, logs, approval records, dependency registers, and technical documentation support.
Use each item as a yes/no question during RFPs, security reviews, architecture reviews, and vendor demos. Require written evidence for any answer that affects regulated data or production access.
Where the platform runs, and where your data goes when it does.
How the platform supports your regulatory obligations and audits.
How the platform fits your existing security architecture.
How retrieval works without leaking internal knowledge.
How agents are built, routed, and governed across models.
How the platform behaves after the pilot becomes production.
What it actually costs, and how you get to production.
We will complete this checklist against your specific deployment, security, and compliance requirements — and back it with a reference architecture and security questionnaire response. Start at the Trust Center.
An enterprise AI agent RFP should cover deployment model, data residency, air-gapped support, private RAG, identity and RBAC, audit logging, model governance, BYOM support, compliance evidence, subprocessors, operating model, and pricing predictability.
Evaluate whether orchestration, retrieval, embeddings, model inference, logs, and governance can run inside your environment. Then ask for diagrams, evidence samples, SIEM integration details, offline update process, model lifecycle controls, and production support commitments.
A major red flag is any architecture where sensitive prompts, source documents, embeddings, or audit logs must leave your perimeter without a clear technical reason, customer approval path, and compensating control.
Require air-gapped deployment if your environment has no outbound internet, supports defense or government workloads, touches OT or critical infrastructure, or has policy restrictions that prohibit external model APIs.
Model the cost at production volume: prompts, tool calls, private RAG retrieval, batch jobs, retries, agent loops, users, environments, connectors, support, and infrastructure. Token-only pricing often looks simple in pilots but can be hard to forecast at enterprise scale.
Useful evidence includes AI inventory, risk classification, human-oversight controls, model and dependency registers, audit logs, incident-review data, operational resilience test support, technical documentation, and controlled change history.