A draft went out on Tuesday. By Thursday the client was asking when the mobile app would be ready. Nobody in your team had agreed to a mobile app. The draft came out of a model, someone skimmed it on a phone between meetings, and a line about “cross platform delivery” survived into the version that got signed. Now it is scope, because it is in writing.
That is the real risk with AI generated proposals. Not bad grammar. Not a stiff tone. Invented scope, invented timelines and invented capability, written fluently enough that nobody reads them twice.
The model writes, it does not decide
A language model is very good at turning notes into prose. It is very good at keeping a consistent voice across nine sections. It is very good at producing a first draft in four minutes that a human would take ninety minutes to produce.
It is not good at knowing whether your fabrication shop can hold a three week turnaround in March. It has no idea that this client has an ageing invoice from last year. It does not know your firm has never done an on site deployment outside Delhi NCR.
So the rule is simple. The model composes. Humans supply the facts and humans approve the commitments. Every problem with AI generated proposals comes from blurring that line, usually because the draft looked confident.
What context to feed the model
A model with no context invents context. That is not a fault, it is what it is built to do. It fills gaps with the most plausible thing. Your job is to leave fewer gaps.
Feed it three categories of fact, and nothing vague.
Facts about the client
Company name and legal entity, exactly as it should appear. Industry. Size, roughly. The named contact and their role. What they actually asked for, in their words, not your summary of it. If the enquiry came through a form, paste the form. If it came from a call, paste the call notes.
This is where a CRM earns its place. If your leads live in a spreadsheet, the context is in someone’s memory or an email thread. On BizvoraOne, the enquiry, the source, the stage and the call notes sit against the record, so the drafting step starts from what was actually said rather than what someone half remembers. If you use a voice agent to make the first contact, the notes it captures become drafting input too, which is one more reason to be strict about what that agent is allowed to promise on a first call.
Facts about the job
The scope items, priced, from your scope library. Not a description of the scope. The actual line items with quantities and rates. The exclusions. The assumptions the price depends on, for example client supplies drawings by a fixed date, or site access is available on weekends.
If your scope library does not exist yet, this is the point where the absence hurts. A model asked to invent scope items will invent good sounding ones. They will be unpriced and unbuildable.
Facts about you
Two or three relevant past jobs, with real details you are allowed to share. Your standard payment terms. Your standard timeline structure. Your GST position. The certifications you actually hold.
Never let the model write the credentials section from scratch. That is where invented awards and invented client logos come from.
What to lock as fixed text
Some parts of a proposal should never be generated. They should be inserted verbatim, every time, from a template the model cannot touch.
Lock these:
- Payment terms. Advance percentage, milestone splits, credit period, late payment position. One approved wording. If a client negotiates different terms, a human edits it and a human initials it.
- Taxes and GST treatment. Whether the quoted figure is inclusive or exclusive, the applicable rate, place of supply where it matters. A model that paraphrases a tax line is a liability.
- Exclusions and assumptions. These exist precisely to prevent the scope creep an AI draft can cause. They must be identical across proposals so your team recognises them at a glance.
- Validity period. Thirty days, fifteen days, whatever you use. Fixed.
- Warranty, support and liability. If your standard support window is ninety days, that sentence is fixed text and nothing else.
- Legal boilerplate and signature block. Obvious, but worth stating.
Everything else can be drafted: the understanding of the requirement, the approach, the phasing narrative, the relevance of past work, the covering note. Those are places where fresh writing genuinely helps, because they should read as if written for this client.
The split is easier to hold if you already know which sections a proposal needs and what each one has to answer. Generated sections carry argument. Locked sections carry commitment.
A six item review checklist before anything leaves the building
One person reads the full draft against this list before it becomes a PDF. Not a skim. A read. If it takes longer than fifteen minutes, the draft was fed bad context and should go back rather than be patched.
- Every deliverable named in the draft exists in the priced scope. Read the narrative sections and underline each thing the proposal says you will do. Check each one against the line items. If the narrative mentions training sessions and the pricing does not, one of them is wrong. This single check catches most invented scope.
- Every number appears once and matches its source. Totals, quantities, rates, tax, timelines, team size. A model will happily restate a total in the summary that does not match the table. Pick the pricing table as the single source and verify every restatement against it.
- Every named fact about the client is correct. Entity name, spelling, contact designation, city, what they asked for. Getting the client’s own company name wrong loses more deals than a high price.
- Every claim about you is one you can evidence. Past project references, years in business, team strength, certifications, sector experience. If you cannot point to the proof in under a minute, cut the sentence.
- Locked sections are intact and unparaphrased. Read the payment terms, tax treatment, exclusions and validity word for word against the template. Models smooth text they are not meant to touch. Look for it.
- Dates and durations are ones operations has agreed to. A start date, a delivery window or a milestone schedule in a proposal is a commitment. Someone who controls capacity has to have said yes to it. If nobody has, the proposal goes out with a duration and no fixed dates.
That is the definition of done. Six checks, one named reviewer, then it becomes a branded PDF and goes out.
Who reviews, and how long it should take
One reviewer, not a committee. Committees are how a two hour proposal becomes a four day proposal again.
For most SMB teams the reviewer is the person who owns the number, usually a sales head or a partner. For technical scope, the reviewer may need a five minute confirmation from delivery, and that confirmation should be a message, not a meeting.
A reasonable cadence looks like this. Drafting happens the same day the scope is settled. Review happens within one working day. Anything sitting in review for more than two days gets escalated, because the delay is now costing you more than the proposal quality is worth.
If you generate proposals in a tool that keeps the CRM record, the quote lines and the output in one place, the review is faster because the reviewer can see the source data next to the draft. That is the useful part of running AI Proposals inside the same workspace as the pipeline rather than in a separate window with copy and paste in between.
When AI generated proposals are the wrong tool
Be honest about the cases where this does not help.
Highly regulated submissions with prescribed formats do not benefit. Tender responses to a government format, a compliance filing, a court facing document: fill the format, do not generate prose.
First proposals to a client you have never spoken to properly do not benefit either. If the context is thin, the draft will be thin, and the model will pad it. Have the conversation first.
Very small, very standard jobs also do not need this. If you sell the same annual maintenance contract forty times a year, a fixed template with three variables is faster and safer than any draft.
AI drafting pays off in the middle: bespoke enough that a template does not fit, frequent enough that writing from scratch every time is wasteful.
What to do first
Do not start by turning on generation. Start by building the two things that make generation safe.
First, write your locked text. One document, six blocks: payment terms, tax treatment, exclusions and assumptions, validity, warranty and support, signature block. Get whoever handles your accounts to read it. This takes an afternoon.
Second, list your ten most common scope items with rates. Not fifty. Ten. That covers most of what you sell.
Then draft one proposal with AI, run the six item checklist on it slowly, and note every place the draft went beyond its source. Those places tell you exactly what context you failed to supply. Fix the input, not the output.
