How to write an RFP for an AI project
Most AI RFPs ask for the wrong things and get incomparable answers. What to specify so responses can actually be evaluated.
AI RFPs tend to specify a solution and ask for a price. That produces bids you cannot compare, because every vendor has interpreted the ambiguity differently and priced their own interpretation.
Specify the outcome and the constraints, not the design
Say what the system must achieve, what it must not do, and what it must comply with. Leave the architecture to the responses — that is precisely where you want vendors to differentiate, and where their answers reveal how they think.
Define success numerically before you ask for a price
If you cannot say what accuracy, latency or throughput would make this a success, nobody can price it responsibly. Ambiguity here gets loaded into contingency by honest vendors and ignored by the rest, which systematically favours the wrong bidder.
Describe your data honestly
Volume, format, quality, access path and known gaps. Overstating data readiness is the most common cause of AI projects that come in late — the estimate assumed data that does not exist in the state described.
What to put in, what to leave out
Specify
- The outcome, numerically — accuracy, latency, throughput
- Constraints: residency, compliance, systems you must integrate with
- Your data honestly, including the gaps
- How you will evaluate the responses
Don't specify
- The architecture — that's what you're buying judgement on
- A named model or vendor stack
- A fixed price before anyone has seen your data
- Requirements you can't measure at handover
Ask questions that separate vendors
The answers to these differ sharply between teams who have shipped and teams who have demoed:
- How will you evaluate this, and what will you show us weekly?
- What will the system do when it is not confident?
- Who specifically works on this, and what else are they on?
- What are the ongoing costs after handover, itemised?
- Describe an AI project of yours that underperformed, and what you changed
Ask for a paid discovery, not a free proposal
A serious estimate requires looking at your data. Vendors who quote a fixed price for a complex AI build without that are either padding heavily or planning to change the number later. Pay for two weeks of scoping and you will get a plan worth having — and you can stop there if it disappoints.
What a vendor's answer tells you
- Quotes a fixed price without seeing your dataPadding, or a number that will change
- Offers a paid discovery firstHas done this before
- No evaluation plan in the responseIntends to ship on impressions
- Names the actual team and their availabilityPitch team is the delivery team