When people hear "proposal agent," they picture a machine that reads an enquiry and fires back a quote. That product exists. It is also the reason a lot of owners will not go near this category, because the one thing they will not delegate is the document that sets the price and the promise.
So here is a different version, followed all the way through. One assignment, nine steps, and a named human at the end who presses send.
The reason to walk through it in this much detail is that every step is a place where an ungoverned system quietly does the wrong thing.
The assignment
An owner finishes a discovery call. The call was recorded, with the other party's agreement. Back in the team channel, the owner types roughly this:
Draft a proposal from this morning's call. Same structure as the one we sent last month, keep it to two pages.
That is the whole instruction. Everything that follows happens because the surrounding system is built, not because the sentence was clever.
Step 1. Resolve who is asking, before interpreting what they asked
The first thing the system establishes is identity: which person, in which company, with which permissions. Not because the owner might be an impostor, but because the answer to "our standard terms" is different for each business the system serves, and because a request from a junior team member should not silently return the same material.
If identity cannot be resolved, nothing else happens. It asks.
Step 2. Turn the message into a scoped assignment
"Draft a proposal from this morning's call" is not yet a specification. The system expands it into one: which recording, which client, what the deliverable is, what format, what length, which audience, and what a good result looks like. Anything genuinely ambiguous gets asked about now, in one message, rather than guessed at and discovered three paragraphs into a bad draft.
This is the step most systems skip, and it is why their output feels generic.
Step 3. Take the smallest route that will work
A two-page proposal from one recording and one template does not need a committee of agents debating it. It needs a transcript, a template, the client's context, and one competent pass.
Multi-agent coordination is not a sign of sophistication. It adds cost, delay, and new ways to fail. The right structure is the simplest one that reliably produces the work, and for most assignments that is a single specialist with good context.
Step 4. Retrieve only from approved sources
The system pulls the transcript of that specific recording, the previous proposal used as the structural model, and the current service descriptions from wherever the company keeps them.
What it does not do is search everything it can technically reach. The list of retrievable sources is a written decision made during setup. This is what stops another client's numbers from wandering into your document because both happened to live in the same drive.
Step 5. Treat connectors as access paths, not permissions
Worth stating separately because it is the most common real-world failure. Connecting a shared drive typically grants everything the connecting account can see. That is almost never the same as what should be in scope for this assignment. The boundary is built on top, deliberately.
Step 6. Produce the work, and keep the evidence
Now it drafts. Structure from the reference proposal, substance from the call, terms from the approved source of record.
Alongside the draft it keeps what it used: which recording, which timestamps informed which claims, which template, which version of the service descriptions. Not as an audit ceremony, but because the reviewer's first question is always "where did this number come from," and the answer needs to take five seconds rather than five minutes.
Step 7. Separate drafting from acting, absolutely
The draft lands in the channel. It is not sent. It is not saved into the CRM as the live version. It is not shared with the client. Nothing leaves.
The system produced a document and stopped. This is not a limitation that gets engineered away in version two. It is the design.
Step 8. Deliver to a named reviewer in the channel they already use
The draft appears where the team already works, addressed to a specific person who owns the review, not posted into a general channel where everyone assumes someone else is handling it.
That person reads it, fixes what is wrong, and sends it. Their name is on it. That has not changed and should not.
Step 9. Measure accepted outcomes, not impressive attempts
The number that matters is not how many drafts were produced. It is how many went out with light edits, and where the heavy edits clustered. If the pricing section gets rewritten every time, the specification for pricing is wrong, and that is fixable. Attempts are vanity, accepted work is the metric.
What this actually saves
The time saved is not the typing. It is the gap between the call ending and a draft existing.
In practice, proposals that used to wait for the owner to find a clear evening now come back from meeting notes in about twenty minutes. The owner still spends real time on the document. They spend it editing judgement calls rather than reconstructing what was said and hunting for last month's version.
That is the honest description of the win. Not "AI writes your proposals." The first eighty percent arrives while the conversation is still fresh, and the person who owns the relationship spends their attention on the twenty percent that requires them.
There is a compounding effect too. Because the draft exists immediately, follow-up happens the same day rather than the next week, and same-day follow-up is worth considerably more than a marginally better document sent late.
The part that is easy to miss
Nothing in those nine steps is about the model. Any capable model will write a decent proposal from a good transcript and a good template.
The work is in step one, four, five, and seven: knowing who is asking, restricting what can be read, not confusing a connector with a permission, and refusing to send. That is the difference between a demo that impresses people in a meeting and a system a business will actually leave switched on.
And the person at the end is not a transitional arrangement to be removed later. They are the reason the whole thing is safe to use.
If you want to see which of your workflows this shape actually fits, that is what the two-day audit is for. Apply at advisy.com.