Separate judgement from assembly
A proposal contains commercial judgement, client context, scope, risk, and repeated explanatory language. AI can help with the repeated and structural parts while the team keeps ownership of the promise.
The dividing line is usually obvious once you draw it: pricing, scope commitments, and unique client context require a human's read of the relationship; boilerplate methodology sections, standard capability descriptions, and formatting do not.
Getting this line wrong in either direction causes a specific failure: too cautious, and the team keeps hand-typing the same boilerplate every week; too permissive, and a commercial commitment ends up phrased by a model that has no stake in whether it can actually be delivered.
A worked example: a scope section
Take the "our approach" section that appears in nearly every proposal. It rarely changes in substance between clients, but it is often rewritten from scratch each time because the last version lived in an old email or a colleague's drafts folder. AI can draft that section from three or four approved versions and match the tone of the rest of the document.
The pricing table next to it is the opposite case. Even if the numbers follow a formula, the decision to discount, bundle, or hold the line belongs to whoever owns the relationship — a place where a fast draft is not the constraint, and speed is not the goal.
The same split shows up in the executive summary. Its structure and opening framing can be templated; the specific reason this client should choose you over the alternative they are also evaluating cannot be, because that reason changes with every deal.
Build from approved examples
The best starting point is a small library of approved sections, tone examples, discovery notes, and delivery constraints. AI can draft from that material instead of inventing a fresh argument every time.
Keeping that library current matters more than the drafting tool itself. Stale approved language quietly reintroduces the exact problem it was meant to solve — text that sounds right but no longer reflects current pricing, delivery capacity, or positioning.
A library does not need to be large to be useful. Five or six well-chosen examples covering the most common proposal types beat fifty examples nobody has time to keep current.
Someone should own that library the way a style guide gets owned — reviewed on a schedule, not only when a mistake surfaces in a sent proposal.
Keep review explicit
Proposal automation should surface assumptions, missing inputs, and risky claims. The final decision belongs with the person who understands the client and the commercial trade-offs.
A simple rule works well here: every AI-drafted section carries a visible marker until a human has read and approved it. That marker disappears at send time, but its presence during drafting stops assembled content from slipping into a proposal unread.
This matters most on the sections most likely to be skimmed rather than read — the ones near the end of a long document, after the reviewer's attention has already been spent on the pricing and scope up front.
Who checks the checker
Review explicitness only works if someone is actually accountable for it — a named reviewer, not a general expectation that "someone will look at it." Proposals with an unnamed reviewer tend to get the least scrutiny of all, because everyone assumes someone else already checked.
Where this breaks down
AI-assisted proposal writing works less well on new-logo pitches, unusual scope, or any client where the standard sections do not fit. Forcing a templated structure onto a genuinely unusual deal produces a proposal that reads competently and answers the wrong brief.
The fix is not to abandon the approved-sections approach. It is to flag, before drafting starts, whether this proposal fits the standard shape or not — one judgement call, made early, that prevents most of the mismatch.
A short pre-check — does this deal resemble the last ten, or not — takes less time than drafting the first section, and it tells the team upfront whether the approved-sections approach even applies.