Is Generative AI Ready to Write Enterprise RFPs?

Yes, but you’re going to need guardrails.

If you’ve read this far, you’ve probably had to deal with a proposal response.

Some are straightforward questionnaires while others are long, complex documents with hundreds of requirements. Some even have an urgent deadline attached to them. You might even need multiple stakeholders to help you complete the response.

We call that the nightmare RFP. It comes out of nowhere and force an unprepared sales team to scramble. We were curious how enterprises are tackling these responses so we asked Pradeep Nayar at WalkMe for his advice… how do you avoid the nightmare RFP?

Prad’s wisdom tells us the key ingredients are preparation and automation.

Prepare your content, loop in the right stakeholders, and automate as much as you can.

On the automation side, you’ve already seen tools that can help you write responses. Maybe you tried using ChatGPT, CoPilot, or Claude to respond to an RFP.

If you’ve used an LLM, by now you’re familiar with some of the drawbacks such as answer quality, accuracy, hallucination, speed, and a laundry list of data and privacy concerns.

And if you haven’t, then read on…

Generative pre-trained transformers (GPTs) are, in essence, an advanced form of autocomplete text generation. Good for writing responses to RFPs. And also bad for writing responses to RFPs.

That’s because generative AI is prone to hallucination and making things up.

This is completely unacceptable when it comes to responding to sales questionnaires, especially when you’re asked about important product capabilities, security, or compliance requests. The worst part is – AI sounds really confident when it’s lying to you. This makes it even more difficult to spot the inaccuracies and address them.

Getting answers wrong is bad enough. Making it difficult to teach your AI new things can make matters even worse.

Most language models are based on a large set of pre-trained data. Training a custom model is an expensive process that most businesses will never touch. Fine-tuning is much more accessible, but it’s unrealistic to do this every time you have new information.

For example, your answers to a compliance questionnaire might change on a quarterly basis. Product updates and new features might impact your responses as well. How do you teach your AI all of these updates in a fast, easy, and cost-efficient manner?

So you’re left with augmenting the AI with information from your business in the form of files, URLs, and 3rd party websites. Traditional copilots don’t make it easy to pull this data in, or to correct wrong answers when they come up.

When responding to an RFP questionnaire you often need to be able to have multiple projects running at the same time. For example, a compliance questionnaire, security assessment, and product capabilities matrix might be multiple sheets in a single document or separate parts entirely.

What’s great about humans is that they can multi-task. This makes it all the more surprising when you find that that Generative AI chatbots are often single threaded. Filling out hundreds of rows in an excel sheet or responding to pages of questions in a word Doc is not what they’re designed to do.

So you end up having to ask a single question at a time, hoping you’re getting accurate responses and then having to fact-check them anyway. A single threaded chatbot is not a great user experience when it comes to proposal management.

It’s odd that this needs to be called out but it goes to show how bad this problem really is. Honest answers are table stakes when it comes to a sales team responding to questionnaires. It’s one thing to provide fluffy answers… but fabricating entire features or compliance certifications is completely unacceptable.

A questionnaire automation tool should not be able to guess what an answer is. It should not be allowed to lie, and significant measures should be in place to prevent hallucination. This is where guardrails come into play such as:

On the security and privacy side, if your team is worried about uploading sensitive data to an LLM then you can forget about automation. Sensitive information is the foundation of any good RFP response. If your team is hesitant about using it, you won’t get very far.

Speed is a big factor too. You should be able to have multiple questionnaires running in parallel without having to worry about a tool’s capacity. This is important for demonstrating the value of an AI over a human – multi-tasking is expected of any software solution.

Check out our RFP software buyer’s guide, where we dive deeper into performance benchmarks and other qualities to look out for.

To solve the “learning” issue, allowing a human in the loop is the way to move forward.

This ability is surprisingly difficult for AI chatbots, who do a great job of memorizing short-term inputs but can’t be easily corrected when they get something wrong.

When it comes to RFPs and sales questionnaires, correcting AI-generated responses should have multiple modes:

1up combines these learnings to generate professional enterprise RFP responses while giving users the ability to:

With 1up, teams are able to automate the response process while ensuring a high standard of data quality and accuracy. For example, check out how FusionAuth automated the process to complete RFPs faster.

You’re just going to have to try it out for yourself.