When you outsource system development, the first deliverable is the RFP (Request for Proposal). The quality of the RFP largely determines the quality of the proposals you receive, the accuracy of the estimates, and ultimately the success or failure of the project. This article explains what an RFP should contain, how to draft one efficiently with generative AI, and “why you should never submit an AI draft as-is”—from the perspective of a development firm that is also on the receiving end of these proposals.
What an RFP should contain
The format is flexible, but the minimum information a vendor needs in order to produce an estimate is fixed.
- Background and objectives — Why this system is needed. Current operations and the problems with them
- Scope — Target operations, target users, and the range of functionality you want built (write down what is out of scope too)
- Distinguishing must-have from nice-to-have requirements — Marking everything “required” sends estimates soaring
- Non-functional requirements — Number of users, data volume, security, operating hours, and so on
- Existing environment — Existing systems that need to be integrated, and whether data migration is involved
- Budget and schedule — Even if you would rather not disclose it, showing an “upper-bound range” tends to pay off in the end
- What you want proposed and how you will evaluate it — State up front what you will judge proposals against
- Contract terms — Acceptance criteria, warranty against defects, and ownership of intellectual property
How to draft with generative AI
It is simply true that having generative AI produce a draft is far faster than writing from scratch. The trick is not to make it write everything in one go.
- Hand over your current operations as bullet points — Who does what, and how often. Only a human can write this part
- Articulate the problems and objectives through dialogue — Having the AI ask you “what exactly is difficult about this work” reduces gaps
- Have it build the section structure first, then write section by section — Fix the outline by giving it the item list above
- Have the requirements list produced as a table — Three columns: feature name, summary, and required/nice-to-have. You can reuse it directly for later comparison and scoring
Why you should never submit an AI draft as-is
This is the heart of the article. An AI-generated RFP draft has some typical problems when viewed from the receiving side.
1. Plausible but unverified requirements creep in
AI fills in “requirements a typical system might have.” In reality, unneeded features (an approval workflow no one uses, excessive permission management, and the like) slip in and can push the estimate up by millions of yen.
2. The particulars of your own operations are not reflected
Most of the reason estimates diverge is “exception handling.” A process that only occurs at month-end, a rule that differs only for a specific client—the AI does not know about these, so it cannot write them. And requirements that are not written come back after the contract is signed as additional costs.
3. No sense of feasibility or market pricing
Phrases like “in real time” or “automatically with AI” take a second to write, but the implementation cost can differ by an order of magnitude. When the required level is fixed without any sense of market pricing, you invite declined proposals or inflated estimates.
4. It cannot notice its own contradictions
Writing a budget of ¥3 million while the requirements list is on a ¥30 million scale; demanding data migration and parallel operation within a three-month schedule—the moment a vendor sees a contradiction like this, they grow wary and quote a number with a safety margin built in.
Final checklist before you send it
- Did a business owner confirm, line by line in the requirements list, “will we really use this in our operations?”
- Have the must-have requirements been narrowed to under half of the total?
- Did you interview business owners about exception handling, closing processes, and integration with other systems, and reflect the answers?
- Are the budget range and schedule balanced against the volume of requirements?
- Did you decide the evaluation axes (price, team, track record, maintenance) first and write them into the RFP?
Summary | Write with AI, finish with a professional eye
Generative AI is a powerful tool for creating an RFP, but it structurally cannot guarantee correctness. Draft quickly with AI, then have a human with a feel for development market rates review the requirements for validity, feasibility, and contradictions—this division of labor is, even now, the most rational way to write an RFP. At SHANNON, we help review and finish AI-generated RFP drafts, standing on the client’s side of the table.