Issuing an RFP for an AI project is an exercise in translating business problems into technical requirements. Most organizations get this wrong in one of two directions: they're either too vague (asking vendors to propose a solution without enough context to price it accurately) or too prescriptive (specifying the technology stack in detail when they'd be better served by letting vendors propose the right architecture). The result is either incomparable proposals that are impossible to evaluate fairly, or constrained proposals that miss better solutions.
Here's how to write an AI RFP that attracts capable vendors, generates comparable proposals, and sets your project up for success.
Start With the Business Problem, Not the Technology
The most common mistake in AI RFPs is leading with technology: "We need a machine learning model to predict X." Good vendors can handle a technical brief. But the most capable vendors want to understand your business problem first, because the right technical solution is often not what the client initially envisioned.
Your RFP should open with a clear description of the business context: what your organization does, what workflow or problem you're trying to address, and what success looks like in business terms. How many hours per week does the current manual process take? What's the cost of false negatives or false positives? What does a successful outcome look like in terms you could measure on your P&L?
This context lets vendors propose solutions that actually fit your problem rather than the solution they were most recently asked about. It also reveals whether a vendor genuinely understands your industry or is pattern-matching to previous projects.
Define Your Data Situation Honestly
AI vendors need to understand your data situation before they can give you an accurate proposal. Projects often run over time and budget because vendors discover data quality issues mid-engagement that weren't disclosed upfront. Being honest about your data situation saves everyone time.
Document what data you have: Where is it stored? In what format? How complete is it? How far back does it go? Does it include the labels or outcomes your AI system will need to learn from? Are there data access restrictions (privacy, legal, contractual) that will affect how the data can be used?
If your data situation is messy — and most organizations' data is messier than they'd like — say so. The vendor's job is to work with what exists, and experienced AI teams are used to starting with imperfect data. What they need is an accurate picture of the challenge, not a sanitized version that makes the project look simpler than it is.
Specify Constraints, Not Solutions
Rather than specifying a particular model architecture, platform, or technology stack, specify your constraints. These might include:
Budget constraints: Your total budget range (not just the amount you want to spend, but the range you could approve for the right proposal).
Timeline constraints: Your go-live target and any hard deadlines driven by business events.
Integration constraints: What systems the AI must integrate with, including CRM, ERP, or industry-specific platforms. List the specific software and version if possible.
Compliance constraints: Regulatory requirements that affect data handling — PIPEDA, PHIPA, SOC 2, industry-specific standards.
Infrastructure constraints: Whether you require on-premise deployment, specific cloud providers, or Canadian data residency.
Team constraints: Whether you have internal technical staff who will maintain the system, or whether you need the vendor to provide ongoing support.
Specifying constraints rather than solutions gives vendors the information they need to propose appropriate architectures while preserving their ability to bring expertise to the approach. You'll get better proposals.
Ask the Right Questions
Your RFP should include specific questions for vendors to address. Here's a set that consistently surfaces the information needed to evaluate AI vendor proposals:
Technical approach: How would you approach this problem? What model types or architectures are you considering? Why?
Data requirements: What data will you need, in what format, and how much historical data is required before the system can be trained effectively?
Accuracy and performance: What accuracy or performance benchmarks are you committing to? How will you measure performance before and after launch? What happens if the system underperforms?
Integration plan: How will you integrate with [specific systems]? What API documentation or technical contacts do you need from us?
Team composition: Who will be working on this project? What are their relevant credentials and prior experience with similar projects?
Timeline: Break down your timeline by phase. What are the milestones and deliverables at each phase?
Ongoing support: What support model do you offer post-launch? What does the handoff look like if we decide to manage the system internally?
References: Provide two to three client references for comparable engagements, including contact information.
Pricing Structure Matters
Ask vendors to quote using a structure that makes comparison possible. For AI projects, this typically means:
A fixed fee for the project phases through launch (discovery, development, testing, deployment), with clear deliverables at each milestone. Milestone-based payment schedules reduce vendor risk and keep projects moving.
A separate ongoing support or maintenance fee, quoted as a monthly rate. Understanding the total cost of ownership — not just the implementation cost — is essential for building an accurate business case.
If vendors have consumption-based costs (API calls, model inference costs), ask them to estimate these based on your described usage volume. Hidden API costs have derailed more than a few AI project ROI calculations.
Evaluation Criteria
Tell vendors how you'll evaluate proposals. This disciplines your own evaluation process and signals to vendors what matters to you. A typical weighting for an AI project RFP might look like:
Technical approach and feasibility (30%): Does the proposed solution actually solve the problem? Is the technical approach sound?
Relevant experience (25%): Has the vendor done comparable work? Are their references credible?
Team quality (20%): Are the specific people who will work on your project qualified?
Price and value (15%): Is the pricing fair and transparent? Does the cost-benefit calculation work?
Company stability and support capability (10%): Is this a vendor you can count on for the next three to five years?
Common Mistakes to Avoid
Setting a deadline that's too short: Quality AI vendors are often booked out 4–6 weeks. If your RFP timeline gives vendors 5 business days to respond, you'll hear only from shops that aren't in demand.
Not issuing a pre-bid conference: A one-hour call where all vendors can ask questions (with answers distributed to all) produces better proposals and signals that you're a serious buyer.
Asking vendors to sign NDAs before sharing the RFP: If the problem isn't sensitive enough to describe to a vendor without an NDA, it's probably too sensitive for an RFP process. Save NDAs for shortlisted vendors before detailed discovery.
Ignoring the pilot: For complex or novel AI projects, consider including a paid pilot phase as part of your evaluation. A $10,000 to $25,000 paid pilot with two shortlisted vendors is often more informative than 20 pages of proposals.
The goal of an AI RFP is to create the conditions for a successful vendor relationship. That means being honest about your constraints, clear about your success criteria, and respectful of vendors' time. The organizations that write the best RFPs attract the best partners.