AI in proposals — what works, and what makes you look automated

Most advice about writing proposals with AI is advice about prompts. That is looking for the keys under the streetlight. The prompt is the easiest thing to change and the least important to the outcome. What decides proposal quality is duller: how much real context you gave the system, and what that system does when the context runs out.

The test every proposal either passes or fails

Before changing anything in your process, do one thing. Take the last AI-generated proposal you sent and highlight every sentence that could be pasted into a proposal for a different client without changing a word.

If more than a third of the text is highlighted, the client received a form. They will not recognize a language model — they will recognize the absence of specifics, and that is something people detect reliably even when they cannot name it.

The test is brutal because it catches exactly the sentences that feel best to write: "a comprehensive solution tailored to your needs", "years of industry experience", "an individual approach to every client". All of those are true. All of them are interchangeable, which makes their value to the reader zero.

The input ranking that actually matters

I have run this across dozens of proposals in different industries and the ranking is remarkably stable.

Input	What comes out	Effort
Call transcript	A proposal in their language, with their objections	None — you already have the recording
Notes + price list	Correct proposal, missing the decision context	Low
The client's website	Good industry background, nothing about their problem	None
An industry description in a prompt	Text interchangeable between clients	Low, and no payoff

The gap between the first row and the last is larger than the gap between any two language models on the market. That is the most important sentence here.

The reason is simple. The model has no access to what you did not give it. If it does not know the client mentioned two churned accounts last quarter and that the decision needs board sign-off on 15 September, it will not write about that — it will write something statistically plausible. And "statistically plausible" is the definition of an interchangeable sentence.

Why a transcript beats notes

Notes are your interpretation of the call, filtered through what struck you as important at the time. A transcript contains what you did not write down because it did not seem relevant: an offhand line about the previous vendor, a hesitation on the budget question, the word the client used for their own problem.

That last one is the most valuable. When the client says "handoffs between teams are a mess" and your proposal says "optimization of operational process flows", you have just replaced their sentence with yours. Theirs was better — because they are the one who has to recognize it.

Four places where AI is wrong and must not be trusted

A language model has no way to distinguish a fact from a sentence that sounds like one. In a commercial proposal that is expensive in exactly these places:

1. Numbers. "Reduces order handling time by 40%" — where did 40 come from? If not from your data, then from nowhere. That number surfaces at the first meeting where somebody asks how you know.
2. Dates. "Implementation takes 6 weeks", written by an AI, becomes a commitment nobody in your company confirmed.
3. References. An invented case study is the fastest way to burn a relationship. And models invent case studies readily, because it is a section where "something has to go".
4. Commitments. SLAs, guarantees, payment terms. These are sentences you are legally answerable for.

The common thread: wherever the proposal promises something on your behalf, the content has to come from you. AI can arrange it and phrase it. It cannot invent it.

Worth noting that buyers already know this. In Gartner's survey of 645 B2B buyers, 51% said they are more likely to encounter misleading information from GenAI than from a sales rep, and 69% prefer to validate AI-generated insight with a rep. Suspicion of AI-produced content is your reader's default setting right now. A proposal with visible automation marks starts from that baseline.

What to do instead of hunting for a better prompt

Record calls and use the transcripts. This single change does more than everything else combined. It needs no new tool — it only requires that you stop discarding recordings after the meeting.

Build a gate on gaps, not on style. Before generating, check four things: who decides, what objections came up, what the budget or its order of magnitude is, and what proof you can show. If one is missing, one email asking the client beats a proposal that guesses.

Tag the source. When every section shows whether it came from the transcript, from research, or is an AI assumption, pre-send verification takes five minutes instead of forty. And you never send a sentence you did not check, because you can see which ones those are.

Cut what you cannot defend. A "why us" section with no concrete case hurts more than it helps. Six strong sections beat seven with one filler.

Where AI is genuinely good

This piece reads sceptical, so for balance — the things AI does better than most people, myself included:

• Structure. Ordering the argument the way a reader makes a decision, rather than the way you happen to think about it.
• Completeness. A human writing at 10pm forgets the risks section. A system does not.
• Translating jargon. Turning your technical description into a sentence the client's finance director will understand.
• Uncomfortable questions. AI will put an objection into the FAQ without flinching — the one you subconsciously skip because it is awkward.
• Consistency. The tenth proposal of the week is as good as the first.

That is real value, and a lot of it. It just is not the value of "AI writes the proposal for you". It is the value of being better prepared than you would be alone, under time pressure.

One sentence to take away

If you change one thing about how you use AI for proposals: stop describing the client to the system, and start giving it the conversation with that client. The rest is detail.