Short answer
Creating a business proposal with AI works best when you supply the facts and the AI supplies structure and wording. Give it the client’s brief, your scope, your prices and your evidence, ask for a draft that uses nothing else, then check every claim. In New Zealand, claims made in trade need reasonable grounds when you make them, so an invented case study is a liability, not a shortcut.
- You supply scope, prices and evidence; the AI supplies structure, wording and gaps.
- Make the AI list every factual claim so you can match each one to evidence.
- Never let AI set prices, timeframes or guarantees.
- State whether prices include GST, and check every calculation yourself.
- Use a past proposal as a style reference, never as content.
Goal: Turn a client’s brief and your own notes into a first-draft proposal in under an hour, with a claims register that shows the evidence behind every promise. Who it is for: Owners, project managers and business development staff in service businesses who write proposals or quotes with a narrative. Time: 40 minutes, plus a real client brief and your notes on scope and price.
Most AI proposal generators fail the same way. They write what a proposal usually says rather than what yours should say: “experienced team”, “proven track record”, “on time and on budget”. The generic filler is merely weak. The specific inventions are dangerous: a case study for a client you never had, a percentage saving nobody measured, a four-week timeline nobody agreed. This lesson splits the job cleanly. You own the facts. The AI owns the form.
What AI should and should not write in a proposal
AI is good at structure, restating a brief, plain wording and spotting gaps. It should not originate prices, timelines, guarantees or proof.
| Proposal part | Let AI draft it? | Why |
|---|---|---|
| Restating the client’s needs | Yes, from their brief | Shows you listened, and is easy to check |
| Structure and headings | Yes | Matches what evaluators expect to find |
| Scope and deliverables | Only from your scope notes | Must match what you will actually do |
| Prices, rates and GST | No, paste them in | A commercial decision, not a writing task |
| Timeline | Only from your plan | A promise you must keep |
| Case studies and numbers | Only from evidence you supply | Claims need substantiation |
| Assumptions and exclusions | Yes, then review hard | AI is good at noticing what is missing |
The last row deserves attention. Disputes about proposals usually come from scope that was implied rather than excluded, and an assistant asked “what would a careful client assume is included?” finds those gaps quickly.
Why invented claims are a real risk
New Zealand’s Fair Trading Act 1986 prohibits misleading conduct in trade, and the Commerce Commission’s guidance on making accurate claims says a business must have reasonable grounds for a claim at the time it is made; substantiating it afterwards may not be enough. An AI that adds “reduces maintenance costs by 20%” to your proposal has made a claim on your behalf. This is general information, not legal advice, but the practical rule is simple: if you cannot point to the evidence, the sentence goes.
Exercise: draft a proposal with a claims register
- Gather four inputs. The client’s brief or your call notes, your scope notes, your price breakdown, and two or three pieces of evidence you can stand behind. Remove other clients’ confidential details and anyone’s personal contact information.
- Set the style in one paragraph. For example: “Plain New Zealand English, short sentences, no superlatives, prices in NZD excluding GST.” If you attach a past proposal you liked, label it as a style reference only.
- Run the drafting prompt. Adapt the worked example below in ChatGPT, Claude, Copilot or Gemini. Keep the instruction to use only your inputs and to mark gaps.
- Work through the claims register. For every line marked NO SOURCE, either supply evidence or delete the claim. Do the same for any claim whose source does not really say what the draft says.
- Fill the gaps yourself. Replace each [NEED] placeholder with a real answer, or decide the section is not needed.
- Read it as the client would. Ask the assistant to act as the client’s evaluator, score the draft against the brief and list unanswered questions. Fix what it finds.
- Do the final human pass. Check names, dates, every price and GST calculation, validity period and terms. The person who signs reads every line.
Worked example: a grounds maintenance proposal
Illustrative example: a 12-person Tauranga landscaping firm responding to a body corporate committee’s request for a 12-month grounds maintenance contract on a 40-unit complex in Mount Maunganui. The business and figures are fictional.
You are drafting a proposal for my business. Use ONLY the inputs below.
Do not add facts, numbers, clients, awards or guarantees that are not in
the inputs. Where a section needs information I have not given, write
[NEED: what is missing] instead of inventing it.
INPUT 1, client brief: (pasted request from the body corporate committee)
INPUT 2, our scope: mowing fortnightly Oct to Apr, monthly May to Sep;
hedges trimmed quarterly; garden beds weeded monthly; green waste removed
same day; spraying only with committee approval.
INPUT 3, our prices (NZD, excl GST): $1,450 per month; one-off hedge
reshaping $2,800; extra callouts $95 per hour.
INPUT 4, evidence: two body corporate sites maintained since 2022
(permission to name them: no); spraying staff hold GrowSafe certificates.
Structure: 1. Your requirements 2. What we will do 3. Not included
4. Price 5. Assumptions 6. Next steps.
Plain NZ English, short sentences, no superlatives.
After the draft, add a CLAIMS REGISTER table listing each factual claim
and the input that supports it, or NO SOURCE.
Example output (illustrative, abridged):
4. Price
Monthly maintenance: $1,450 + GST ($1,667.50 including GST)
One-off hedge reshaping: $2,800 + GST
Additional callouts: $95 per hour + GST
[NEED: how long this price is valid, and payment terms]
CLAIMS REGISTER
Claim | Source
Mowing fortnightly Oct to Apr, monthly May-Sep | Input 2
Green waste removed on the day of each visit | Input 2
Experience maintaining body corporate sites | Input 4 (sites not named)
"A reliable, high-quality service" | NO SOURCE
Two details show the method working. The assistant flagged a missing validity period instead of inventing “30 days”, and it caught its own filler phrase as unsupported. Your job is to supply the validity period and delete the filler.
Where AI proposals go wrong
- Invented proof. Case studies, statistics and client names that appear from nowhere. The claims register is the control; run it every time.
- Scope that grows in the wording. “All garden maintenance” is a bigger promise than your list. Compare the scope section with your notes line by line, and keep an explicit “not included” section.
- Recycled content. A past proposal pasted as content brings another client’s name or pricing with it. Search the final draft for previous client names before sending.
- Arithmetic and GST. Assistants still make calculation errors. Recalculate every total and say clearly whether prices include GST.
- Confidential briefs in personal accounts. Use a business plan with data settings you have checked. Our lesson on giving AI useful context covers how much to paste and what to leave out.
For a reusable version of the drafting prompt, save it using the approach in building a reusable prompt library. Firms that write proposals every week can see how this fits a larger process in our proposal preparation workflow, and teams that want hands-on practice with their own documents can book AI training for business teams. If you are new to prompting, start with writing useful business prompts, and use your account brief from researching a customer account as the first input.
Self-check
- Every price in the draft came from your price breakdown, and GST treatment is stated.
- Every factual claim appears in the claims register with a real source.
- No line marked NO SOURCE survives into the final version.
- Scope, exclusions and assumptions match what you will actually deliver.
- No other client’s name, price or confidential detail appears anywhere.
- An evaluator-style read has been done and its questions answered.
- The person who will sign the proposal has read it in full.
Next lesson: Analysing a spreadsheet with AI.
Sources
- Making accurate claims, Commerce Commission, 2026
- Fair Trading Act 1986, New Zealand Legislation, 1986
Written by the Levy Street team in Wellington. Last updated 19 October 2026.