A Strong AI Draft Is Not a Proposal Workflow

The first draft is no longer the hard part.

This deserves to be said plainly, because a lot of software marketing depends on pretending otherwise. ChatGPT, Claude and Microsoft Copilot can produce a genuinely good first draft of a consulting proposal. Given the context, a couple of reference documents and a clear instruction, the output is structured, fluent and close to the register a client expects. It is not a toy. Any consultant who has tried it knows this, which is why claims that AI "cannot really write proposals" are met with justified scepticism inside firms.

So if drafting is solved, why does producing a proposal still take three weeks?

Because the draft was never the expensive part. The expense is everything around it: finding the right precedent, reconstructing what the client actually said, applying the firm's methodology rather than a generic one, tracking which version is current, coordinating partner review, and carrying the approved scope into the SOW and the delivery plan without it quietly drifting.

A general assistant helps with one step in that chain. The other steps are where the weeks go.

What is actually missing

The gap is not model quality. It is that a chat interface is a conversation, and a proposal is an artifact moving through a governed process. Those are different objects with different requirements.

Persistent opportunity context

In a chat tool, context lives in a thread. It was assembled by one person, in one session, from whatever they happened to paste. Next week, on a different machine, a colleague picks up the pursuit and starts a new thread with a thinner version of the same context — because reassembling it perfectly would take longer than approximating it.

So the context degrades with each person who touches the opportunity, and the degradation is invisible. Nobody knows what the previous person had pasted in. In a pursuit that spans four people and three weeks, that is not a small leak.

Opportunity context needs to be an object attached to the opportunity — what the client said, what was interpreted, what remains unknown — not a property of whoever is typing.

Persistent artifacts

A proposal is not text. It is a thing with an owner, a status, a version history and traceable sources. Ask a chat tool what the current status of a proposal is and the question does not parse. There is no proposal. There are messages containing proposal-shaped text.

The consequence appears at the moment of pressure. Which draft went to the client? Who changed the delivery assumption? Where did the claim about the firm's prior experience in this sector come from — was that a real engagement or something the model produced because it fit? These questions have answers in a system with artifacts. In a thread, they have archaeology.

In-place iteration

This is the most concrete difference, and the easiest to feel.

A partner reads a draft and says: make the introduction more sensitive to their restructuring — they lost people last year and this reads as if that never happened.

That is a request to change the current proposal. In a chat tool, it produces a new message containing a new version of the introduction. Now someone must decide whether to paste it back into the document, whether the rest of the proposal still coheres, and which of the seven proposal-shaped blocks in the thread is authoritative. Repeat that ten times across a pursuit and the real work is no longer writing — it is reconciling.

A request to revise should update the proposal. The artifact changes; the history records that it changed. Nothing is spawned, nothing needs to be reassembled, and there is never a question about what "the proposal" refers to.

Firm knowledge and methodology

A general model writes a generic-good proposal. It knows the shape of consulting work broadly, but it does not know that your firm always runs a two-week diagnostic before committing to a target operating model, that you never price transformation work on time and materials, or that the closest analogue to this pursuit is an engagement from eighteen months ago that went badly in a way this client should be protected from.

That knowledge exists in the firm. It sits in prior proposals, in methodology documents, and in the heads of people who cannot be consulted on every draft. Making it available at the point of drafting is not retrieval for its own sake — it is the difference between a proposal that reads well and a proposal that is yours.

Exact review and approval

This is where the absence is most serious, because it is a governance question rather than a productivity one.

A firm needs to distinguish, unambiguously, between:

• the latest draft — what someone last produced;
• the current version — what the team considers live;
• the approved version — the exact artifact a partner signed off on;
• and changes made after approval — visible as changes, not folded silently into a file.

In a chat-based flow, these four collapse into "the most recent thing in the thread," and approval becomes a memory of a conversation. That works until a client points at a clause nobody remembers approving, or until the SOW turns out to describe a scope that diverged from the proposal two revisions before anyone noticed.

Consulting sells promises. The version of the promise the firm approved should be a fact, not a recollection.

The honest summary

Generative AI genuinely removed the blank page, and firms should use it for that. What it did not do is supply the layer a professional services firm needs around the page: persistent context, real artifacts, in-place iteration, the firm's own knowledge, and approval that refers to something exact.

That layer is unglamorous. It is also where proposals actually get stuck — and why the practical starting point for firms is connecting the workflow from conversation to approved proposal rather than buying a better drafting tool. It is the same reason partner review time does not fall when drafting speeds up: the reviewer's cost was never the writing.

Solon is built for that layer. See how the proposal workflow works, or apply to run a live proposal through it as a design partner.