This is the RFP template I send when buyers ask "give me the question list". Sixty questions across six sections, with a 1-to-5 scoring rubric and walk-away criteria. The template assumes enterprise procurement; small-team buyers should consider the lighter structured evaluation instead. Free to copy. Companion to the PoC checklist, vendor evaluation, and procurement checklist.
RFP structure
The six-section structure tracks the six evaluation criteria. Each section starts with two open-ended questions ("describe your approach to X") and then has 8 to 10 specific evidence-based questions.
Vendors should respond inline; do not let them attach a separate document. Inline responses are scoreable; document attachments are work avoidance.
Section 1: Capability and use cases (10 questions)
- Describe how your platform handles agent reasoning. Workflow with LLM steps, true agent reasoning (plan-and-execute, ReAct), or a hybrid? Cite the reference architecture or paper.
- Provide a 5-minute video of an agent on your platform running a use case similar to: [insert your top use case]. Real run, not a recording of a demo.
- What is the maximum context window the platform supports today, and what is the latency profile at the top of that window?
- How does the platform handle multi-step agent runs that pause for human approval? Provide the API for resume.
- How do you measure and report agent run quality? What are the headline metrics for your customers similar to us?
- Describe how a customer adds a custom tool to an agent. Provide a sample tool definition and the call surface the agent uses.
- How does the platform handle agent runs that exceed token, time, or cost budgets? What controls does the customer set?
- What models does the platform support today? Pinned snapshots, latest aliases, or both? How do customers control which model their agents use?
- Provide three reference customers running use cases similar to ours, with contact information. We will conduct independent reference calls.
- What does your roadmap look like for the next 6 months in capabilities relevant to our use cases?
Section 2: Security and compliance (12 questions)
- Provide the most recent SOC 2 Type II report (or ISO 27001, depending on jurisdiction). Date of issue and date of next audit.
- List the data centers and regions where customer data is processed. Confirm where our data would specifically reside.
- Describe tenant isolation at the data, retrieval, and prompt layers. Provide an architecture diagram.
- How are customer credentials and secrets stored? Provide the KMS or vault solution and the rotation policy.
- Does the platform train its own models on customer data? If yes, how does opt-out work? If no, is this contractually enforced?
- Provide the audit log API specification, including: events captured, retention, access controls, and a sample response for a single agent run.
- How does the platform handle GDPR data subject access requests, deletion requests, and data export requests? Provide the API and response SLA.
- Describe the incident response process. RTO and RPO commitments? Mean time to notify customers of a security incident?
- What is the platform's policy and tooling for prompt injection defense? Cite OWASP LLM Top 10 controls you address (OWASP, 2025).
- How does the platform handle PII detection and redaction in agent inputs and outputs?
- What encryption is used at rest and in transit? Specify cipher suites and key management.
- Provide your most recent penetration test summary. Date, scope, summary of findings, remediation status.
Section 3: Integrations and data (10 questions)
- List native integrations to: [your top 10 systems]. For each, specify auth model (OAuth, API key, service account) and whether reads, writes, or both are supported.
- How does a customer build a custom integration? Provide the SDK or framework.
- What is the latency and throughput profile of your integration layer? Per-integration if it varies.
- How are integration credentials scoped per tenant? Confirm one tenant cannot use another's credentials.
- How does the platform handle vendor-side rate limits on integrations? Backoff, queueing, error reporting to the agent.
- Does the platform support webhook ingress for event-driven agent triggers? Auth model, replay handling, delivery guarantees.
- How is data passed from integrations to the agent context? Inline, retrieval, or both?
- What knowledge-source connectors are native? (e.g., Confluence, Notion, Google Drive, SharePoint, S3.) Document refresh cadence.
- Describe the platform's retrieval architecture. Embedding model, vector store, retrieval algorithm, chunking strategy.
- How does the platform handle data residency at the integration layer? Can data from EU sources be processed only in EU regions?
Section 4: Pricing and contract (10 questions)
- Provide the complete pricing structure: per-seat, per-run, per-token, credits, flat-rate. Include all overage charges and how usage is metered.
- Project our cost at: [Year 1 volume estimate], [Year 2 volume estimate], [Year 3 volume estimate]. Include all costs you can foresee.
- What is the model usage cost passthrough? Are model costs included in your pricing or billed separately?
- What is the volume discount schedule? At what point does enterprise pricing kick in?
- What are the standard contract terms? Term length, payment terms, auto-renewal, termination for convenience.
- What is your standard MSA and DPA? Provide both for legal review.
- What are your standard SLAs and SLA credits structure?
- Are there fees for: implementation, training, integration build, additional environments (dev, staging), API access, premium support?
- How does pricing scale if we add a second product team or department? Same agreement or separate?
- What is your policy on price increases at renewal? Cap on annual increase?
Section 5: Support and SLAs (8 questions)
- Describe support tiers offered, response time commitments per tier, and what tier comes with our contract size.
- Publish the platform uptime SLA and provide actual uptime for the last 12 months by month.
- Is there a public status page? Provide the URL and the notification subscription options.
- What is the support process for a customer-impacting incident? Channels, escalation path, executive engagement criteria.
- Provide named customer success or technical account management commitments at our contract size.
- Describe the documentation and self-service knowledge base. Are runbooks for common issues publicly available?
- What is the training and enablement provided to our team during onboarding?
- How does support escalate to engineering? Provide a recent example timeline.
Section 6: Vendor stability (10 questions)
- Provide the company's funding history, current investors, and approximate runway.
- Provide revenue scale (range), customer count (range), and year-over-year growth.
- Has the company pivoted product focus in the last 24 months? Describe.
- Describe customer concentration. What percent of revenue comes from the top 1, top 5, top 10 customers?
- Provide the leadership team and engineering org chart. Who owns this product specifically?
- What is the company's policy on acquisition and the customer impact of one? Reference the data portability and contract continuity terms.
- How does the platform support customer data export? Provide the export API and document format.
- What is your customer churn rate over the past 12 months?
- Has the platform deprecated any major feature in the last 24 months? How was the deprecation handled?
- What happens to our deployment and data if the company is acquired, restructured, or winds down?
Scoring framework
The scoring rubric applied per question.
- 1 - Missing or evasive. No answer, unrelated answer, or refusal to provide requested evidence. Walk-away signal in capability or security sections.
- 2 - Below standard. Acknowledges the question but the answer reveals immaturity, gaps, or genuine inability.
- 3 - Standard with evidence. Solid, defensible answer with the requested evidence (logs, references, documents, sample outputs).
- 4 - Above standard. Goes beyond the request with relevant context. Multiple references at our scale, multiple evidence types.
- 5 - Best in class. Public-by-default transparency. Documented patterns. Reference architecture the vendor lets us share with stakeholders.
The PMI procurement guidance treats RFP scoring as a multi-stage process: technical evaluation first, commercial evaluation second, with explicit conflict-of-interest separation (PMI procurement management). Section scoring follows the same separation: capability and security drive technical scoring; pricing drives commercial.
Process and timeline
A workable RFP timeline.
- Week 0: Long-list to short-list. Send abbreviated version to 6 to 10 vendors. Receive responses in 1 week.
- Week 1: Score short-list responses, select 2 to 4 finalists for the full RFP.
- Week 2-3: Full RFP response window. Vendor questions answered via a shared FAQ document.
- Week 4: Score responses, conduct reference calls.
- Week 5: Live demos with each finalist using the same use case. Same evaluators, same scoring.
- Week 6: Pricing negotiation, contract review, security review.
- Week 7-8: Final selection, contract signed, kickoff.
FAQ
- What should an AI agent platform RFP cover?
- Six sections: capability and use case fit, security and compliance, integrations and data, pricing and contract terms, support and SLAs, and vendor stability.
- How long should an AI agent platform RFP be?
- Sixty questions across six sections. Shorter and you miss material risk; longer and vendors will copy-paste rather than answer specifically.
- How do you score AI agent platform RFP responses?
- Score each answer 1 to 5: 1 missing or evasive, 3 standard with evidence, 5 best-in-class with documentation. Weight section totals by importance. Walk-away criteria override the total.
- What questions catch vendors who oversell?
- Ask for specifics: a reference customer running our use case, a live demo of audit log retrieval, a published SLA with credits, an architecture diagram showing tenant isolation.
- Should I share my RFP template publicly?
- Yes if you can. Public templates raise the floor for the whole market.
- How long should vendors get to respond?
- Two to three weeks is standard. Less signals copy-paste; more signals you are not yet committed.
Sources
- OWASP, "Top 10 for Large Language Model Applications", 2025, owasp.org
- PMI, "Procurement Management Body of Knowledge", pmi.org
- NIST, "AI Risk Management Framework", 2023, nist.gov
- AICPA, "SOC 2 Trust Services Criteria", 2025, aicpa-cima.com
- Gartner, "How to Run a Successful RFP", 2024, gartner.com
