Almost everything written about AI in proposals is about AI writing. Is the draft generic? Are the claims supported? Did it invent a past performance reference? Those are fair questions and I ask them constantly.

But a lot of us are not only drafting with these tools. We are pasting Section L (Instructions) into a chat window and asking what it says and then building the compliance matrix from whatever comes back. That step gets almost no scrutiny, and it should get the most, because of where it sits in the process.

A weak AI draft dies at pink team. Somebody reads it, winces, and fixes it. An AI summary of the RFP does not get reviewed by anyone, because it isn’t content, it’s the thing everyone else’s content gets measured against. It becomes the matrix, the matrix becomes the outline, the outline becomes the assignments, and three weeks of writing happens against a requirement that was softened before a single author opened a file.

Here is what it looks like.

The requirement, and the summary of the requirement

Representative language, not a real solicitation, but it is written the way real ones are written.

The narrative shall address, at a minimum: (a) proposed transition staffing, including the number and labor categories of personnel dedicated to transition and the date each is expected to be on site; (b) the three most significant transition risks, the mitigation for each, and the party responsible. The Offeror shall not exceed five (5) pages for this factor.

And the summary a tool gave back:

This section asks offerors to describe their transition approach, including staffing, risks, schedule, and communications. Recommended length: five pages.

Four things moved.

“Shall address” became “asks offerors to,” which converts an obligation into an invitation. “At a minimum” disappeared, and those two words were carrying real weight, they make the list a floor. Without them your team reads the list as the whole job and stops asking what else the customer wants to see.

Then there’s subpart (b). The RFP asked for three specific risks, a mitigation for each, and a named party responsible for each one. The summary says “risks.” That compression deleted the hardest part of the section, the part where somebody has to put their name next to a mitigation, and nobody would ever know it was there.

And “shall not exceed five pages” came back as “recommended length: five pages.” There is no such thing as a recommended page limit.

Nothing in the summary is invented. There’s no hallucination in it. Every sentence is defensible on its own, which is exactly why it’s dangerous, and the matrix you build from it is wrong in four places before anyone starts writing.

I remember one year during fiscal year close; the team was working more bids than we had people to manage them, and I was reviewing the past performance volume on one of those proposals. I had read the solicitation and the most recent amendment that morning, and something in the compliance matrix didn’t line up with what I’d just read. It hadn’t been built from the RFP. It had been built from a summary of the amendment, and the recency requirement had changed from five years to three. We were at red team, and we had to replace two of our five past performance references.

Why it gets past people who are good at this

Not carelessness. Two other things.

We are trained to review writing, and our instincts are tuned for it. Reading a draft, you’re already asking where the proof point is, whether this could have been written for any customer, whether an evaluator would score it. Reading a summary, you check that the topics are present. Staffing, risks, schedule, communications, all four are in there, so it looks complete. The instincts never fire because nothing about a summary looks like content.

The second thing is that the tool did what we asked. Summarize means shorter and smoother. Obligation language is the first thing a summarizer throws out, because “shall address, at a minimum” is ugly and “asks offerors to” reads better. It isn’t malfunctioning. We handed it a job where success and accuracy point in opposite directions, and it chooses success.

What to do on your next proposal

Build the compliance matrix from the RFP text, not from a description of the RFP text. That one sounds too obvious to say out loud until you go look at how your last matrix was actually made.

If you want AI help pulling requirements, don’t ask for a summary. Ask for extraction:

Extract every requirement from the text below verbatim. Do not shorten, paraphrase, or combine them. Preserve every instance of shall, must, will, and shall not exactly as written. Output one row per requirement.

Verbatim is doing all the work in that prompt. You’re asking for clerical labor, matching text against a list, moving it without touching it, which is the thing these tools are genuinely good at, and close to the opposite of what most of us ask them to do.

Then check the output against the source:

Compare your summary to the RFP text I pasted. List every instance where an obligation (shall, must, will) became a suggestion, where a limit became a recommendation, or where a specific deliverable became a general topic.

The low-tech version works too, and honestly, I still do it. Spot check the obligation words yourself. Search Section L for shall, must, will, and shall not. Then find each one in your matrix.

We’ve spent over two years asking whether AI wrote something good. The question I’d rather my team ask is whether it read something right – because a bad paragraph costs you a rewrite, but a bad matrix can cost you the bid.