Can GenAI Actually Write a Winning Proposal? Real Examples of What Works & What Doesn’t
I respond to RFPs for a living at an RFP software organization. Meta, I know.
Naturally, working in tech, an industry consumed by AI chatter, people started asking whether generative AI can write a winning proposal… and I needed to know.
That’s why I spent a week trying to fully replace my proposal workflow with Claude, uploading our real content library, running it against a 36-question RFP, and pushing it to do everything from drafting responses to assigning tasks and exporting deliverables.
The goal was simple: find out where generative AI excels and where it falls short. So here’s the breakdown.
TLDR: Can Generative AI Help in Creating More Effective RFP Responses?
Sometimes, but there are serious risks to relying solely on generative AI. When examining the pros and cons of Claude as a replacement for our current response process, the list of cons far outweighed the pros.
| Where Claude Excels ✅ | Where Claude Falls Apart ❌ |
| Summarizing Content: Answers multi-question RFPs using an uploaded content library, with confidence ratings per answer. | Hallucinations: Invents plausible-sounding but false details, even when explicitly told not to. |
| Response Quality: Produces natural, human-sounding answers, close to what a proposal writer would actually develop. | Reliability at Scale: Not trustworthy enough for high-volume, deadline-driven proposal work without heavy human review. |
| Personalization: Identifies prospect-specific themes and weaves them consistently across all relevant answers. | Setup Complexity: Requires hours of prompt engineering to function. |
Where Claude Impressed Me (And It Did Impress Me)
Before we get into where things went sideways, I want to be fair because dismissing what Claude can do would be just as misleading as overhyping it.
Two capabilities in particular stood out, and they’re the tasks that eat up the most time in a real proposal workflow.
Summarizing Knowledge Into Answers
I handed Claude a 36-question RFP in Excel format with a clear prompt: answer using the uploaded content, flag uncertainty, leave blanks rather than guess, and it answered all 36 questions.
The answers it generated felt calibrated and readable, closer to something I would actually write than the stilted outputs I’d seen from other AI tools. Claude even flagged which answers it felt confident in versus which ones warranted a second look from a human. That awareness is super useful when you’re working against a deadline.
Personalizing Responses to my Prospect
This is where things got exciting. I fed Claude notes pulled directly from the Salesforce opportunity for the prospect we were targeting, and the results were impressive.
Claude didn’t just restate our standard differentiators; it identified patterns in the prospect’s priorities (translation capabilities, in this case) and suggested weaving that theme through every relevant answer.
So yes, generative AI can meaningfully improve RFP responses for summarization, personalization, and first-draft speed.
Where GenAI Falls Apart: The Three Questions to Ask Before You Make Any Changes
Despite the wins above, my experiment surfaced three failures that any proposal leader should take seriously before restructuring their process to use general-purpose AI.
1. Who Owns the Workflow When Things Break?
The prompt that made our Claude workspace function took my Sales Enablement teammate, Bety Garcia, seven hours to build. Seven hours. And the result, while impressive, was something only she fully understood. It is extremely complex, and difficult to edit.
On the flip-side, purpose-built proposal software exists precisely to solve the challenge of building—and maintaining—a system.
Anyone with access can jump into an answer, a prompt, and quickly make edits. And it’s eons more reliable than the break we saw with Claude, which brings me to…
2. What Happens When Things Break?
Hallucinations aside, I also ran into functional breakdowns during the experiment. Task assignment required re-clicking the text box for each character. Exporting completed responses as a CSV failed outright.
With proposal software, there’s an infrastructure and a support team behind reliability. With a custom AI workflow, you’re the support team.
3. Can You Catch the Hallucinations?
Even with explicit instructions not to fabricate information, Claude told me that our prospect’s customers had used Loopio’s Portal Chrome Extension. Claude has no way to track that, but it was so confident that it would have looked completely plausible to anyone less familiar with our product.
I caught it because I know our product and our content library well enough to recognize the error. But what about a newer team member under deadline pressure? What about a highly technical RFP? The risk scales with complexity and time pressure.
When I pushed back, Claude did acknowledge the mistake.
Specifically, Claude admitted: “I invented a claim that combined two separate facts.”
The trouble is, transparency after the fact doesn’t protect you from sending a proposal with false claims to a prospect. You need a system you can truly trust when responding to RFPs.
Where I Landed With My Claude Experiment
Generative AI, and specifically Claude, can do genuinely impressive things in a proposal workflow.
It’s particularly good with fast first drafts, baseline personalization, and confidence comparisons. But making it work consistently, reliably, and at scale requires infrastructure that doesn’t come bundled with the tool.
Would I use Claude again as part of the process? Absolutely. Would I hand over the keys to my proposal workflow? Definitely not.
If you work in proposals, I’d encourage you to run your own experiments—and validate the results. It’s no question that AI can produce an answer, it’s whether you can trust that answer when the deal is on the line.