Short answer
AI proposal preparation is the use of a language model to turn a request for proposal and a firm’s past work into a first draft: a compliance matrix, matched project examples, tailored CVs and method text. In New Zealand professional services firms it works best as an assembly step, with a bid lead owning the decision to bid, the price and every factual claim.
- AI drafts the compliance matrix and first pass; people own the bid decision, price and claims.
- Every experience claim needs a library source; the Fair Trading Act bars unsubstantiated representations.
- A clean, tagged library of past projects and CVs matters more than the model you pick.
- Expect assembly time to fall and review time to rise. Measure both.
What AI proposal preparation covers
AI proposal preparation means using a model for the assembly work of a bid: reading the tender documents, listing every requirement, matching past projects and CVs, and drafting sections in your house style. It does not decide whether to bid, set the fee or choose the team. Assembly is where most bid hours go, and it is the part least likely to win work on its own.
It suits engineering, planning, architecture, consulting and IT services firms. For other uses in the sector, see AI for professional services.
The workflow, step by step
- Inputs
- The tender documents and buyer notices; your library of project sheets, CVs, method statements and past submissions, tagged with sector, client type, date and approval status.
- Systems
- SharePoint or Google Drive for documents, a retrieval index over the approved library, a language model, and the bid tracker or CRM where opportunities live.
- Steps
- Extract requirements, evaluation criteria, returnable schedules and word limits into a compliance matrix; retrieve matching projects and CVs; draft each section with a library reference beside every factual claim; list requirements the library has no evidence for.
- Human review
- The bid lead approves the go/no-go and the compliance matrix before drafting; the project director rewrites win themes and method; a second reviewer checks every claim against its reference before submission.
- Outputs
- A compliance matrix, a referenced first draft, a gap list, and CV extracts trimmed to the page limits.
- Measures
- Hours to a review-ready draft, requirements missed at final check, claims corrected in review, and win rate over two quarters.
- Risks
- Invented or inflated experience, stale CVs, another client’s confidential detail carried over from an old bid, and missed mandatory requirements.
The retrieval index is ordinary RAG development with one strict rule: only approved, dated documents go in.
Illustrative example: a council tender at a 25-person consultancy
Illustrative example: a 25-person Wellington engineering and planning consultancy answers about six requests a month, half of them public-sector tenders. The figures are assumptions for one tender with a 10-page response limit, not client results.
| Task | Before (hours) | With AI drafting (hours) | Assumption |
|---|---|---|---|
| Build the compliance matrix | 3 | 1 | Model extracts; bid lead checks every line |
| Find projects and tailor CVs | 5 | 1.5 | Library of 120 tagged project sheets and 25 CVs |
| Draft method and experience | 8 | 4 | Director rewrites method; AI drafts the rest |
| Claim checking and final review | 2 | 3 | Review grows because every claim is checked |
| Total | 18 | 9.5 |
At an assumed internal cost of NZ$95 an hour, 8.5 hours saved is about NZ$810 per tender, or NZ$4,850 a month across six. Take off 6 hours a month of library upkeep (NZ$570) and the net is closer to NZ$4,280. Model usage at an assumed NZ$15 per tender adds NZ$90 a month. A stale library erases the saving.
New Zealand rules a proposal draft must respect
The firm, not the model, makes every statement in a proposal.
- Fair Trading Act 1986. Section 9 prohibits misleading or deceptive conduct in trade, and section 12A of the Fair Trading Act prohibits unsubstantiated representations: claims made without reasonable grounds at the time, whether or not they prove true. A model writing “more than 40 similar projects delivered” with no source creates that exposure.
- Government tenders. Opportunities are listed on GETS, a free service whose staff do not run or evaluate tenders, so questions go to the buying agency. Agencies covered by the Government Procurement Rules (5th edition) must incorporate the Supplier Code of Conduct in tender documents. Declarations against it are sign-off items, never drafting tasks.
- Privacy Act 2020. Proposals carry staff CVs and referee details. Under section 11 of the Privacy Act you remain responsible for information a provider processes for you; the Privacy Commissioner’s third-party guidance says to find out whether it uses that information for its own purposes. Check each tender’s confidentiality terms before uploading buyer documents.
What stays with the bid lead and directors
- The go/no-go decision, including conflicts of interest and capacity to deliver.
- Price, assumptions and exclusions. A model never fills a fee schedule.
- Named personnel. A CV in a proposal is a promise about who will do the work.
- Win themes and the method choices that distinguish the firm.
- Sign-off of every claim, referee and declaration.
Where proposal drafting goes wrong
| Failure | What it looks like | Control |
|---|---|---|
| Invented experience | Plausible project counts with no source | Approved library only; a claim without a reference fails validation |
| Copy-over from old bids | Another client’s name or rates in the new draft | Strip client detail on library entry; search drafts for other client names |
| Stale CVs | Departed staff or outdated qualifications | Each CV carries a review date; older than 12 months is flagged |
| Missed requirement | A returnable schedule left out | Bid lead signs off the matrix; final check against it |
| Generic voice | Every response reads alike and scores low on understanding | Director writes the understanding and method sections |
Making “no reference, no claim” enforceable is a job for structured outputs and validation.
What it costs and where to start
Running cost is mostly people: library curation and review time. Build cost depends on how messy the library is and whether drafting must happen inside Word.
- Test the compliance matrix. Run the extractor over ten past tenders and count requirements it missed against what your team found.
- Clean and tag the library. Approve the project sheets and CVs the model may cite, with dates and client permissions.
- Add referenced drafting. Start with experience sections, where evidence matters most, and count claims corrected in review.
- Compare two quarters. Check hours and win rate before extending drafting to method sections.
Smaller firms can start manually with our lesson on creating a business proposal with AI; to have it built and run, see AI workflow automation. Client research for the understanding section is covered in research briefing preparation, and kickoff actions after a win in meeting actions into project systems.
Sources
- Fair Trading Act 1986, New Zealand Legislation, accessed 19 October 2026
- Government Procurement Rules, New Zealand Government Procurement (MBIE), accessed 19 October 2026
- Supplier code of conduct, New Zealand Government Procurement (MBIE), accessed 19 October 2026
- GETS: Government Electronic Tender Service, New Zealand Government, accessed 19 October 2026
- Privacy Act 2020, New Zealand Legislation, 2020
- Working with third-party providers: understanding your privacy responsibilities, Office of the Privacy Commissioner, 26 November 2024
Written by the Levy Street team in Wellington. Last updated 19 October 2026.