Drafting that can be checked, sentence by sentence.
The outline comes from the solicitation. Each section is drafted against the requirements it has to answer, from your content library and your past performance, under a prompt that says what it may not claim — and every claim carries the requirement it answers and the source it came from.
- Produces
- Drafted sections, with provenance
- Drafts from
- Your library, profile and past performance
- Owned by
- A person, per section, always
A model that will write anything produces text somebody has to check line by line.
Point a general-purpose assistant at a solicitation and it will produce a fluent proposal in minutes. It will also invent a metric, promise a service level nobody agreed to, and describe a capability in words the delivery team would not recognise. The time saved in drafting is spent, at higher cost, in review.
Meanwhile the actual work of proposal writing is not prose. It is answering a specific requirement in the words the evaluator will score against, from content that has been approved before, without contradicting what the last submission said.
The content usually exists. It is in the last three responses, in the capability descriptions marketing maintains and in the past performance library — scattered across a shared drive, in four versions, with nobody sure which is current.
- Review that costs more than writingUnattributed generated text has to be verified sentence by sentence by the one person who knows whether it is true.
- Claims that cannot be supportedA number nobody can source is a protest risk in a submission and an embarrassment in a debrief.
- The same section rewritten from nothingAnswers that were approved six months ago get retyped because nobody can find the version that was signed off.
From an extracted requirement list to a section with its sources attached.
The point of every step is that the draft can be checked against something. Nothing here is generated from a blank page.
Plan the outline from the solicitation
Sections follow the instructions to offerors and the evaluation criteria, so the response is ordered the way it will be read.
An outline with each section carrying the requirements it is responsible for.
Assign owners and states
Every section has a person and a state — not started, drafting, drafted, in review, approved.
A board where an unowned section is visible, rather than discovered in the final week.
Write the prompt as terms
Which requirements to answer, which sources may be drawn on, the length, the voice and what may not be claimed.
A prompt that is a reviewable contract rather than an instruction typed into a chat box.
Draft against the library
The section is drafted from your content library, capability profile and past performance, with requirement markers inserted as it goes.
A draft carrying its requirement references and the list of sources every claim came from.
Refuse what cannot be supported
Sentences containing a number, a superlative or a commitment with no cited source are removed, with the reason recorded.
A shorter draft, and a list of the claims that were dropped so a human can substantiate or drop them.
One section, its prompt, its sources, and the two sentences it refused to keep.
The terms the section was written under are on the left, the draft with its requirement markers in the middle, and underneath it the claims that were removed — with the reason each one failed.
3.2 Data migration approach
- C.7.1, C.7.4, L.3.2 — extracted from the solicitation
- Under 900 words; the volume limit is 25 pages across nine sections
- Content library, past performance library, company profile
- State a migration duration, an error rate or a headcount not in a cited source
- Your approved template — second person, no superlatives
- Content library — “Phased migration, legacy claims systems”
- Past performance — award 2024-0871
- Company profile — data practice
Library entries carry their own age. A section drafted from something nobody has looked at in two years says so before a reviewer reads it.
Migration runs in three passes rather than one cutover: a structural pass that maps the legacy claim record to the target schema, a reconciliation pass that resolves the records the first pass could not place, and a verification pass run against a sample the agency selects rather than one we select. Records that fail reconciliation are quarantined with the rule that rejected them, so the agency reviews a categorized exception list rather than a log. The plan, the mapping and the exception categories are delivered before the first pass runs, and no pass proceeds without written acceptance of the one before it.
Every marker is the requirement the sentence before it answers. The markers survive into the export, which is what makes the compliance matrix a derivation rather than a spreadsheet somebody maintains by hand.
Our migrations typically complete in under six weeks with a 99.9% match rate.
Neither number appears in a cited source. The past performance record gives record volumes and dates, and no match rate was ever measured that way.
We are the leading provider of claims migration to state health agencies.
A superlative with nothing behind it, and the voice rules in the template exclude them regardless.
The removed sentences are the point of the panel. A generator that will write anything is one whose output has to be checked line by line; this one says where every line came from.
Nothing above is a live query or another customer’s pipeline. Every record in it was written for this page, and no buying organization named in it is real.
What makes generated text reviewable.
Provenance, constraint and ownership. Everything else in this feature is arrangement.
Requirement markers in the text
Each claim carries the requirement it answers, and the markers survive into the export — which is what makes the compliance matrix a derivation rather than a parallel document.
Per-section prompts
An executive summary and a past performance annexe have nothing in common. Each section is drafted under its own terms, which you can read and edit.
A content library with ages
Approved answers, capability descriptions and past performance entries, each showing when it was last reviewed, so a draft built from stale content says so.
Claim refusal
Numbers, superlatives and commitments with no source in the cited material are removed rather than smoothed over, and the removal is shown.
Your voice, not a model’s
Tone, person and vocabulary come from your template, so the draft reads like the rest of your submissions instead of like everybody else’s.
A named owner for every section
No section is ever authored by the product. A person owns it, edits it and approves it, and their name is on the record.
Where drafting is the constraint.
Three teams with the same bottleneck at different scales.
A two-person bid team
- Both people write, review and price, and the volume of responses is limited by how fast they can type.
- Outline generation and first drafts on the standard sections, so their time goes to the technical approach.
- More responses answered without hiring, and the parts that need a human get the human.
A firm with a deep content library
- Ten years of approved answers spread across a shared drive, in versions nobody can rank.
- The library as the drafting source, with ages visible and every reuse recorded against the section it landed in.
- Existing content is actually reused, and the entries nobody has reviewed in two years become obvious.
A team that has been burned by generated text
- Tried a general assistant, found an invented metric in a draft, and stopped trusting the tool.
- Prompts as terms, sources on every claim, and the refusal list read as part of review.
- Review becomes checking citations rather than re-verifying prose, which is a job a reviewer can actually finish.
Drafting is metered, and every plan can draft.
The two allowances that matter here are section generations a month and the AI token allowance behind them. A shared content library across the organization arrives at Professional.
- Every plan includes section-by-section drafting with requirement markers and provenance.
- Section generations and the monthly AI allowance rise with each plan.
- Professional and above add the shared knowledge library, so approved content is an organization asset rather than a personal folder.
- Active response projects — how many pursuits can be in flight at once — is the other allowance to watch.
Plan names and allowances are mirrored from the billing catalog. Amounts are on the pricing page, which reads them from Stripe rather than from a number typed into a marketing page.
What people ask about drafting.
Does it write the whole response?
It drafts section by section against the extracted requirements and your own content, and a person owns, edits and approves every section. Nothing is submitted that somebody has not read — and the draft is built to make reading it fast, because every claim says where it came from.
What stops it inventing things?
Two things. It drafts from cited sources — your library, profile and past performance — rather than from general knowledge; and sentences containing a number, a superlative or a commitment that no cited source supports are removed before the draft is saved, with the reason shown. That list is part of what a reviewer reads.
Can we use our own writing style?
Yes. Voice, person and vocabulary rules come from your template and are part of the terms each section is drafted under. The aim is a draft that reads like your last approved submission.
Which model is used?
The platform is built against a provider-independent interface, so the model behind drafting can change without the product changing. What does not change is the constraint: drafting happens against cited sources, and every claim keeps its provenance.
Is our content used to train models?
No. Your content library, past performance and drafted responses serve your organization and are not training data for a shared model.
Draft one section against a real requirement.
Take a solicitation you are working on, let one section be drafted under its terms, and read what it cited — and what it refused to say.