How to Write an AI RFP: Template and Evaluation CriteriaHow to Write an AI RFP: Template and Evaluation Criteria

Writing an RFP (Request for Proposal) for an AI project is different from writing one for traditional software. AI projects involve more uncertainty, require different evaluation criteria, and have unique risk factors. This guide provides a practical template and scoring framework.

Why AI RFPs Are Different

Traditional software RFPs define exactly what you want built: specific features, screens, and workflows. AI RFPs need to account for the fact that AI performance is probabilistic - you cannot specify "the model will always give the correct answer." Instead, you need to define acceptable performance ranges, evaluation methods, and improvement processes.

Additionally, AI vendors vary much more widely in capability than traditional software vendors. Some are building genuine AI solutions with custom models and proprietary data pipelines. Others are thin wrappers over ChatGPT with slick marketing. Your RFP needs to distinguish between the two.

RFP Template: Essential Sections

Section 1: Project Overview

Describe the business problem, not the technical solution. Include:

• What process or capability you are trying to improve
• Current metrics and pain points (with data if possible)
• Desired outcomes with specific, measurable targets
• Timeline expectations and budget range (being transparent about budget range saves everyone time)

Section 2: Technical Requirements

• Data sources the AI will need to access
• Existing systems the AI must integrate with (list specific APIs and platforms)
• Performance requirements (latency, throughput, accuracy targets)
• Security and compliance requirements (SOC 2, HIPAA, GDPR, etc.)
• Deployment preferences (cloud, on-premise, hybrid)
• Data residency requirements

Section 3: Vendor Qualification Questions

These are the questions that separate real AI capabilities from marketing claims:

Technical depth:

• Describe your AI architecture for this use case (not generic - specific to what we are asking for)
• What models will you use and why? (Custom, fine-tuned, off-the-shelf API?)
• How will you evaluate AI quality? What metrics will you track?
• How does the system handle edge cases and errors?
• What is your approach to AI safety and guardrails?

Experience:

• Provide 3 case studies of similar AI projects with measurable outcomes
• Describe a project that did not go as planned and what you learned
• What is your team's experience with our industry?

Data handling:

• How will you handle our data? What security measures are in place?
• Will our data be used to train models for other clients?
• What happens to our data if we terminate the engagement?
• Do you sign BAAs / DPAs as needed?

Ongoing support:

• How do you handle model drift and degradation?
• What does ongoing maintenance include?
• How are model updates and improvements delivered?
• What is your SLA for production issues?

Section 4: Pricing Structure

Request pricing broken down by:

• Discovery and scoping phase
• Development and training phase
• Deployment and integration phase
• Ongoing maintenance and support (monthly/annual)
• Infrastructure costs (passed through or included?)
• Any usage-based pricing (per API call, per document processed, etc.)

Section 5: Evaluation Criteria

Share your scoring rubric with vendors so they know what you value:

Evaluation Scoring Framework

Technical Approach (30 points)

• Architecture quality and appropriateness for the use case (0-10)
• Evaluation and quality measurement plan (0-10)
• Scalability and performance design (0-10)

Experience and Team (25 points)

• Relevant case studies with measurable outcomes (0-10)
• Team qualifications and industry experience (0-10)
• References and reputation (0-5)

Data and Security (20 points)

• Data handling practices and security certifications (0-10)
• Compliance with applicable regulations (0-5)
• Data ownership and portability (0-5)

Cost and Value (15 points)

• Total cost of ownership including ongoing costs (0-10)
• Pricing transparency and predictability (0-5)

Partnership Fit (10 points)

• Communication quality during RFP process (0-5)
• Cultural alignment and collaboration approach (0-5)

Red Flags in Vendor Responses

Watch out for these warning signs:

• Guaranteed accuracy numbers: No honest AI vendor will guarantee "99% accuracy" before seeing your data. Accuracy depends entirely on data quality, use case complexity, and edge case frequency.
• No mention of evaluation: If a vendor does not explain how they will measure quality, they probably do not measure it rigorously.
• Vague architecture descriptions: "We use state-of-the-art AI" means nothing. Look for specific models, specific approaches, and specific reasons for those choices.
• No discussion of failure modes: Every AI system has failure modes. A good vendor discusses them proactively and explains mitigation strategies.
• Unwillingness to share references: If a vendor cannot provide references from similar projects, they may not have done similar work.
• All benefits, no tradeoffs: Every technical approach has tradeoffs. A vendor that only discusses benefits is either naive or dishonest.

Tips for the Evaluation Process

1. Include a technical evaluation: Ask finalists to do a small proof-of-concept with your actual data. Nothing reveals capability like working with real data.
2. Talk to references independently: Do not just call the references the vendor provides. Ask for introductions to recent clients and contact them directly.
3. Evaluate the team, not just the company: Ask to meet the actual engineers and data scientists who will work on your project, not just sales and leadership.
4. Assess communication quality: How the vendor communicates during the RFP process is a strong predictor of how they will communicate during the project.
5. Consider the long term: AI is not a one-time project. The vendor you choose will likely be a partner for years. Evaluate for long-term partnership fit, not just initial project delivery.