Why AI didn't help the proposal flow until the workflow changed.

Twelve people, forty proposals a month. They'd adopted AI "fully." It was not helping. Here's why, and what we changed in the second week.

The founder called me because her agency's proposal flow was crushing her team. Forty proposals a month, each running four to six hours of senior time. The AI tools they were using were, she said, "supposedly the best ones." Everyone on the team had subscribed. Everyone had watched the tutorials. The output was fine. The throughput had barely moved.

She'd concluded the AI was overrated. I thought, from the symptoms, that the AI was fine and the workflow was the problem. I was right, and I'm telling you about it because the specific pattern is probably running somewhere in your business too.

The shape of the problem

I asked her to walk me through one real proposal — start to finish, from the moment the discovery call ended until the moment the proposal was sent. The walkthrough took forty minutes, and by minute ten I could already see what was happening.

The discovery call notes lived in one place. The client's prior conversations lived in another. The reference projects they'd want to cite lived in a third. The pricing spreadsheet lived in a fourth. The "voice" guidelines for how the firm wrote lived partly in a doc and partly in the founder's head. Four to seven different places, depending on the proposal. And every single time a proposal began, a human had to open all of them, reread all of them, and hold all of them in their head simultaneously while drafting.

This took an hour before the draft even started.

Then the drafter would open ChatGPT or Claude and try to get it to help. The assistant, with no context about the specific client, the specific history, the specific voice, or the specific pricing model, would produce something competent and generic. The drafter would then spend two hours rewriting it into something that actually sounded like the firm. Then another hour polishing. Then internal review, usually another hour of the founder's time.

The AI was contributing roughly fifteen percent of the output. The drafter was doing the other eighty-five percent, with AI assistance on bits of phrasing. This is not AI failing. This is AI being handed a blank page with no context and asked to produce work that matches a firm it has never been shown.

A blank page in front of an AI is the same as a blank page in front of a human: the work still has to come from somewhere.

What the leverage was

The move was not a better prompt. Prompts are downstream. The move was building a structure where the AI always started from the same place, with the same inputs, in the same order. In other words: stop treating every proposal as a blank page.

We spent two weeks, on and off, doing three things. None of them required fancy tooling.

First, we collapsed all the "places things lived" into one place per proposal. For each client, a single project folder. Discovery notes, prior conversations, reference projects, pricing inputs, voice guidelines — all copied or linked in. No more opening four tools. When a proposal started, everything the drafter needed was in one location.

Second, we defined the shape of a proposal, once. Not a template with fill-in-the-blanks — a more subtle thing. A structured prompt that asked the drafter to produce, in order, a handful of specific inputs: what is this client actually trying to do, what have they already tried, what is the one thing they'll remember from our proposal, what's the commercial shape, what's the risk. These five inputs, written by a human in fifteen minutes, became the structured context that every AI draft started from.

Third, we built a single AI workflow — not a "tool," a workflow — that took those five inputs plus the project folder's contents, plus the firm's voice guidelines, and produced a first draft that was genuinely close. Not "competent and generic." Close. Specifically close — the client's situation addressed, the firm's voice audible, the pricing reasoning visible.

The drafter's job became: spend fifteen minutes on the five inputs, let the workflow produce a draft, spend forty-five minutes refining it. Total time per proposal: roughly one hour. Quality: higher than before, not lower, because every proposal now started from the firm's structured reasoning rather than from scratch.

Numbers, for those who want them

Before: four to six hours per proposal. Founder touched every one, usually for forty-five to ninety minutes. Output ceiling around forty proposals a month, with the team stretched.

After (three months in): roughly one hour per proposal for the drafter. Founder touched maybe one in four, usually for ten minutes. Output currently running around fifty-five proposals a month, with the team less stretched than before. Close rate on proposals went up too — about eighteen percent, by her tracking — though I'd be careful attributing all of that to the workflow.

What I will attribute to the workflow: the founder personally got back something on the order of twenty-five hours a month. That is not nothing. That is a different life, over the course of a year.

The thing I want you to take from this

"AI isn't working for us" is almost never true in the way the sentence implies. What's usually true is: AI is being asked to produce specific work from generic inputs, in a workflow that was designed for a human to hold all the context in their head. The AI produces exactly what its inputs deserve. The workflow is where the leverage sits.

You can spend months tuning prompts and get a ten-percent improvement. Or you can spend two weeks rebuilding the workflow, once, and get a five-times improvement that compounds with every proposal afterward.

Which one your business needs depends on specifics. That's what the Diagnostic is for — to surface the specifics cleanly, in about seven minutes, without a call.