Writing proposals with AI: what to hand over, what to keep

To write proposals with AI without regretting it, split the work three ways. Hand over the words: the client's problem in your own terms, the approach, the bios, a translation. Let it draft, then check, anything that commits you: the in-scope list, the deliverables, the timeline. Keep the prices, the payment terms and every promise for yourself. Below: a part-by-part table, a worked retainer example and a red-pen checklist to run before you send.

A proposal written with AI still needs the same parts in the same order. If you want the method first, read the step-by-step method for a business proposal. This guide is about who does which part.

Why the split matters

A language model is very good at turning rough notes into clear sentences. It is not good at knowing what you charge this quarter, what your team can deliver in March, or which promises you never make. It will fill those gaps anyway, in the same confident tone as everything else.

That is the real risk. A weak sentence is easy to spot. A plausible price, a generous extra round of revisions or a launch date that sounds reasonable is not, and once the client signs, it is the deal.

So the question is not whether to use AI. It is which parts you can hand over completely, which you can let it draft and then check, and which you never hand over at all.

The part-by-part table

Two rows deserve a word. Cases: AI is happy to describe a project that never happened. If your tool picks from your own library, fine; if it is a plain chat, give it the case or leave the slot empty. Scope, what is out: an AI can list what is included, because you told it. It can't know that you never do content entry or never support old browsers, so that list is yours. How to write both lists is in the guide to writing a scope of work.

Prices: where AI drafts go wrong

A plain chat knows only the prices you paste in. Paste an old rate card and you get old prices, in a well-written proposal that looks finished.

That is not a theoretical risk. At We Do Dev Work, my agency in Bangkok, a client once signed a version with our old, cheaper rates. Several versions were going round and the prices were updated by hand. An AI working from an old rate card makes the same mistake, only faster.

The fix is the same either way: keep one price list, and let prices come only from there. The AI may choose which line goes in and how many; the amount comes from the list. If you offer tiers, build them from that list too, as in the guide to good, better and best tiers.

Promises the AI adds

Language models like to please. Ask for a proposal and you often get extras nobody asked for, because they sound generous. Look for these before anything else:

• Unlimited anything: revisions, requests, calls, support.
• Guarantees of an outcome: rankings, leads, sales, a percentage.
• Dates you did not give: "live within six weeks", "reports every Monday".
• Response times: "replies within the hour", "round-the-clock support".
• Extras that are not in the price: "including copywriting", "free training".
• Figures with no source: a market statistic, a "typical" result.

None of these is wrong in itself. Each one is wrong if you did not decide it.

Worked example: one retainer, two drafts

The studio, the client and the amounts are fictional examples: replace them with your own.

Example

Sample: a two-person SEO studio writes a retainer proposal with AI. In September the studio raised its prices. Its price list now has SEO Essentials at €950 a month, SEO Growth at €1,650 a month, a technical audit at €1,400 and extra hours at €95. The old rate card, from spring, had Essentials at €850 and Growth at €1,500.

The description: "Retainer proposal for Harbour Dental, three clinics. Monthly SEO, Essentials or Growth, the client picks. The technical audit as an option. Twelve-month term. In English."

Draft 1: a plain chat, with the spring rate card pasted in. The text reads well. Two kinds of problem, prices and promises:

Line	In the draft	Current price	Difference per month
SEO Essentials	€850	€950	€100
SEO Growth	€1,500	€1,650	€150

On Growth over the twelve-month term, that is 12 × €150 = €1,800 below the current price. The draft also promised "first-page rankings within 90 days" and "unlimited keyword requests". The studio promises neither.

Draft 2: a tool that drafts from the studio's own price list. Both plans go in as a choice the client makes, at €950 and €1,650 a month. The audit goes in as an option the client can tick, at €1,400 one-off. Extra hours show at €95, billed as used. Flagged for checking: "4 blog posts a month" on the approach slide, a number that was not in the description.

What the studio kept for itself:

• Changed "4 blog posts" to 2, its real Growth volume.
• Added the out-of-scope list: no content for the clinics' booking system, no paid ads.
• Wrote the minimum term and notice period by hand.
• Removed nothing about prices: there was nothing to fix.

The studio's own first-year sum, if Harbour Dental picks Growth and ticks the audit (excluding VAT):

Item	Amount
SEO Growth, 12 × €1,650	€19,800
Technical audit	€1,400
Total, first year	€21,200

This is the studio's own planning sum. In the proposal itself, the monthly price and the one-off audit stay apart.

The two drafts took the same description. Draft 1 failed on the prices and on promises; draft 2 failed on one number in the text, and it said so. The writing was fine in both. That is the pattern: with AI, the words are rarely the problem.

Red-pen checklist

Copy this and run it on every AI draft, on the page the client will see, not in the chat.

Template

Before I send an AI-drafted proposal

1. Client name, contact and email address are right.
2. Every price matches today's price list. No amount in the text that is not in the price table.
3. Every number in the text came from me: days, rounds, pages, posts, hours, weeks.
4. No promise I did not make: unlimited, guaranteed, within X days, 24/7.
5. The scope lists what is out, not only what is in.
6. My rate for extra work is there.
7. Payment terms and the validity date are mine.
8. Every case is real and from my own work.
9. One-off and monthly amounts are shown apart.
10. The VAT rate is right for this client.
11. The tone is consistent: the same form of address throughout.
12. The summary on the cover matches the price page.

This is general information, not legal or tax advice. For your situation, ask a lawyer or accountant.

Plain chat or a tool with your price list

In a plain chat, such as Claude or ChatGPT on its own, you supply the notes, the structure and the prices, and you get text back to lay out, total and send yourself. A tool that already holds your price list needs only the description, and takes the prices out of the AI's hands; you still read every word.

Some of these tools also work from inside Claude. How that goes, from description to sent, is in how to write and send a proposal with Claude.

How Propozit helps

In Propozit you select New proposal, then Describe it, and write what the proposal is for in 20 to 2,000 characters. The proposal AI picks the type of work, products from your price list, the quantities, the client and matching cases, and writes the text of the main slides in the deck's language. It never sets a price: every price comes from your price list, and Check these lists every amount and every unfamiliar number in the text it wrote. Not found lists what no product matches. Nothing is sent: the draft opens in the steps for you to check. The details are in describing a proposal to get a draft.

To change the draft in your own words, ask the AI for changes in the editor: it can rewrite text, add a product from your price list or translate the deck, but it can't change an amount or a discount. Your client then picks options and plans with live totals, one-off and monthly apart. Propozit is also listed in Claude's connector directory, so you can start the same draft from Claude.

Your data is stored in the EU, in Frankfurt; the AI features run on Anthropic in the US, as the privacy page explains. The Free plan has 3 AI drafts a day.

Propozit is in early access and free to start.