TL;DR: A quote-drafting AI handles the 80% templated portion of a project quote in under a minute and leaves the 20% judgment work to a human. Typical quotes take 1–3 hours each, so a studio sending 5+ a week loses a workday to busywork. The build: discovery notes get fed to an LLM with retrieval against your past quotes, the model drafts scope, deliverables, timeline, and a starting price tier, then a human edits and sends. Stack: a Notion or Airtable past-quote library, OpenAI or Claude with retrieval, a small editor UI. Build cost: $1.5–4k. Time saved: 4–8 hours/week per quoter. Human stays on send.
Writing project quotes by hand is a lot of busywork that looks like high-judgment work. The first 80% of a quote is templated content the studio has written hundreds of times. The remaining 20% is the part that actually requires judgment: scope decisions, pricing tier, edge cases.
A quote-drafting AI handles the 80% in under a minute and leaves the 20% to the human. We built one for ourselves, and the same pattern applies to most service businesses that quote regularly.
What "quote drafting" actually means
A typical workflow:
- Discovery call ends. Rep has notes.
- Rep needs to send a quote within 1–3 business days.
- Quote contains: scope summary, deliverables list, timeline, price, terms, what's not included.
- Rep writes from scratch or copies from a similar past quote.
- Quote is reviewed, edited, sent.
Steps 3 and 4 take 1–3 hours per quote, depending on complexity. For a studio sending 5+ quotes a week, that's a workday gone every week to busywork.
The build
A working quote-drafting assistant has three pieces.
1. Discovery notes capture. A structured form or a transcript from the call. Either works. Notes include the prospect's situation, the requested scope, the timeline, any constraints.
2. Retrieval over past quotes. A vector store of your past quotes (or a structured database of them). Indexed by project type, price tier, scope shape. The LLM retrieves the 3–5 most similar quotes as context.
3. LLM generation with retrieval. GPT-4o or Claude Sonnet 4 takes the discovery notes and the retrieved similar quotes and drafts a new one in your voice and structure.
The output is a Markdown or Google Doc draft the rep reviews, edits, and sends. The rep stays in the loop. The AI handles the typing.
The cost
For a studio sending 20 quotes a month:
- LLM generation: ~$0.10 per quote × 20 = $2/mo. Trivial.
- Vector storage (Pinecone or local): $0–$30/mo.
- Hosting (Vercel or n8n Cloud): $20/mo.
- Total: ~$25/mo.
Build cost: $4k–$8k for a working system.
If a senior person on the studio is spending 2 hours per quote and the build saves 90 minutes per quote, the math at $150/hr blended is:
- Savings per quote: $225.
- 20 quotes/month × $225 = $4,500/month saved.
- Payback: 1–2 months.
This is the unit economics for most service businesses that quote regularly. The savings dwarf the build.
What the LLM is actually good at
Three things specifically.
1. Restating the prospect's needs in clean prose. Discovery notes are messy. The LLM reformats them into a "Here's what we understand about your situation" paragraph that's actually readable.
2. Generating the deliverables list. Given a project type and scope, the LLM produces a deliverables list that matches your studio's pattern. This is mostly retrieval-based — past quotes provide the templates.
3. Drafting the closing terms section. Payment terms, scope-change policy, what happens after launch. This is boilerplate that varies slightly per project type. AI handles the variation.
What the LLM is not good at:
- Pricing decisions. Don't let the AI pick the price. Have it suggest a range based on similar past projects, and let the rep decide.
- Strategic scope tradeoffs. "Should we include X or push it to a v2?" That's a human judgment call.
- Reading the room. The prospect's tone, urgency, and price sensitivity require a human.
The build is a typing assistant, not a sales person.
A real prompt structure
The prompt that drives this looks roughly like:
You are drafting a project quote for a web design studio.
The studio's voice: direct, specific, willing to recommend against itself.
Studio voice examples are in the retrieved context.
Discovery notes:
[paste notes]
Most similar past quotes (3 of them):
[retrieved quotes]
Draft a quote with these sections:
1. Situation summary (1 short paragraph in our voice)
2. Recommended scope (5–8 bullets)
3. What's not included (2–4 bullets)
4. Timeline (1 paragraph)
5. Price (use the format from the retrieved quotes)
6. Terms (use the standard terms)
7. Next step (1 sentence)
Output as Markdown. Keep it under 600 words.
Don't fabricate facts about the prospect. Where the discovery notes are silent, leave a [TBD: ask about X] placeholder.
That last line is critical. The AI is allowed to flag gaps for the human to fill, rather than hallucinating answers. We've seen this single instruction cut bad outputs by 80%.
The hard parts of the build
Three things that bite people building this.
1. Vector retrieval over quotes is harder than retrieval over docs. Quotes have implicit structure (price tiers, scope shapes) that simple semantic similarity misses. You may need to filter by metadata (project type, ticket size) before semantic search.
2. Maintaining the past-quotes corpus. Every new project should be indexed. The system gets better as the corpus grows. Without a maintenance habit, the system stagnates.
3. The voice consistency problem. The LLM is trying to match your studio's voice. The retrieved past quotes are the main signal for voice. If your past quotes vary in voice, the AI's output will too.
What we ship for clients
A typical Webdimonia quote-drafting build:
- Discovery on the studio's quote pattern, voice, scope tiers: included.
- Vector store of past quotes (we provide the structure, you provide the corpus): $1k–$2k.
- LLM pipeline with custom prompts tuned to the studio's voice: $2k–$4k.
- Output integration (Google Docs, Notion, Markdown to email): $500–$1.5k.
- Review queue UI: $500–$1.5k (optional).
- Documentation, 30-day tuning window: included.
Total: $4k–$9k. Pays back in 1–3 months for studios sending 10+ quotes a month.
When to skip this
- Under 5 quotes a month. Just write them by hand. The build doesn't pay back.
- Every quote is a unique snowflake. If your work is so custom that no two quotes share structure, retrieval doesn't help. Templates won't either.
- Your team prefers the writing. Some senior people use quote-writing as thinking time. Don't take that away from them.
Three questions to decide this week
- How many quotes a month? Under 5: skip. 5–15: marginal. 15+: probably worth building.
- Do your past quotes share structure? If yes, retrieval works. If no, build a template system first.
- Does the studio have voice consistency in past work? If yes, the AI can match it. If no, fix the voice problem first.
If you want a quote on a quote-drafting build for your studio (with input on the corpus structure and voice profile), send us your current quote template and a rough quote-volume number. We send a tiered proposal back within two days.