The discovery call ends and a scoped, priced draft proposal exists before the follow-up email goes out.
For agencies, consultants, and professional services firms, the gap between a good discovery call and a sent proposal is where deals cool. The typical lag is three to five days, and it is not caused by difficulty — it is caused by the proposal requiring the one person who was on the call to sit down and reconstruct it from memory and scribbled notes. Meeting capture tools (Granola, Fathom) now produce a structured transcript with the requirements, constraints, and budget signals already extracted. Chaining that into a proposal draft against your own templates and rate card collapses the lag to hours. The workflow is not "AI writes your proposal" — it is "AI assembles the 70% that is reconstruction so the human spends their time on the 30% that is judgment and pricing."
The pipeline is only as good as the input. Use the same discovery agenda every time: current state, desired outcome, constraints (budget, timeline, internal capacity), decision process, and success criteria. The AI extracts what was said; if a topic was never raised on the call, no tool can recover it. A fixed agenda is what makes extraction reliable.
Bot-based capture (Fathom) syncs cleanly to the CRM and is fine for most B2B calls. Bot-free capture (Granola) matters when a visible recorder in the participant list would be awkward — sensitive client situations, in-person meetings, or prospects who decline bots. Either way, disclose that you are recording.
This is the real work and it is done once. Take your standard proposal, list every field it needs — scope items, deliverables, timeline, assumptions, exclusions, pricing basis — and write a prompt that pulls exactly those from the transcript, with an explicit instruction to output NOT STATED rather than infer. The inference is what will burn you.
Do not let a model price your work. Extract scope items from the transcript, then map them to prices through a deterministic lookup — a spreadsheet or a table in your CRM. The AI's job is to identify that the client needs three landing pages and a migration; the rate card decides what that costs. This single separation prevents the most expensive failure mode in the workflow.
Speed is the point. The draft lands in your proposal tool or doc template with scope, assumptions, exclusions, and a pricing block populated, plus a visible list of every NOT STATED field the model could not fill.
Read the NOT STATED list first — those are the questions to resolve before sending, and they are frequently the ones that determine whether the deal is profitable. Then sanity-check the price against the effort you actually expect. Fifteen to twenty minutes, not two hours.
Track hours from call end to proposal sent, and win rate segmented by that lag. Most firms find the correlation is strong enough to justify the workflow on its own, independent of the time saved.
Use these templates as-is or customize for your business.
1. CURRENT STATE — What is happening today? What have you already tried? 2. DESIRED OUTCOME — What does success look like in 90 days? How will you measure it? 3. CONSTRAINTS — Budget range? Timeline? Who internally does the work on your side? 4. DECISION PROCESS — Who else needs to approve? What is your timeline to decide? 5. SUCCESS CRITERIA — What would make you consider this engagement a failure? Ask all five. The extraction prompt maps to these headings, so a skipped section becomes a NOT STATED field, which is exactly what you want — a visible gap rather than an invented assumption.
You are extracting proposal inputs from a discovery call transcript. Output ONLY the fields below. If something was not explicitly stated on the call, write NOT STATED. Do not infer, estimate, or fill gaps from general knowledge — a wrong assumption here costs real money.
CLIENT_NAME:
PRIMARY_PROBLEM: (in the client's own words, quote where possible)
DESIRED_OUTCOME:
SUCCESS_METRICS:
SCOPE_ITEMS: (bulleted list of concrete deliverables discussed)
EXPLICITLY_OUT_OF_SCOPE:
TIMELINE_STATED:
BUDGET_SIGNALS: (quote any number or range mentioned; NOT STATED if none)
CLIENT_SIDE_RESOURCES: (who does what on their side)
DECISION_MAKERS:
DECISION_TIMELINE:
RISKS_OR_CONCERNS_RAISED:
OPEN_QUESTIONS: (anything the client asked that was not answered on the call)
Transcript:
{{transcript}}Scope item (extracted) | Unit | Rate | Notes ------------------------------|-----------|---------|------ Discovery & audit | fixed | $X | Always included, never discounted Landing page (design + build) | per page | $X | 3+ pages: apply volume tier CMS migration | per 100pp | $X | Add 30% if source is unstructured Monthly retained hours | per hour | $X | Minimum 20 hrs/mo Rush delivery (<2 weeks) | multiplier| 1.35x | Applies to whole engagement PROCESS: AI extracts scope items -> human matches each to a row -> spreadsheet computes the price. The model never sees the rate card and never outputs a number. This is not paranoia; a hallucinated price you sent to a client is a price you now have to honor.
Get a new AI workflow every week. Prompts, tool stacks, and ROI math included.
Single agent with function-calling: one LLM with a defined toolbox (CRM, calendar, knowledge base) decides which tool to invoke at each turn. Easiest to debug; appropriate for most well-scoped business workflows.
Learn the agentic glossary →Where this workflow tends to break in production — and what to put in place before you ship it.
Model infers scope the client never agreed to
Mitigation: Extraction prompt mandates NOT STATED over inference; human reviews the NOT STATED list before sending.
Hallucinated pricing reaches a client document
Mitigation: Pricing computed by a deterministic rate-card lookup outside the model; the model never emits a number.
Transcript quality collapses on a poor-audio call
Mitigation: Check the transcript before generating; fall back to manual drafting rather than building on garbage input.
Proposal sent same-day but missing a deal-critical unknown
Mitigation: OPEN_QUESTIONS and NOT STATED fields are surfaced at the top of the draft and must be resolved or explicitly flagged as assumptions in the proposal.
Skip this if your proposals are genuinely bespoke every time with no reusable template — the extraction prompt has nothing to map to and you will spend longer maintaining it than writing proposals. Skip it if you sell a fixed productized service with a standard price, because you do not have a scoping problem, you have an order form. And do not let the model price the work under any circumstances, no matter how good the drafts get: a hallucinated rate in a document with your logo on it is a number you will end up honoring.
A phased approach to get this workflow running and delivering ROI.
Days 1–30
Foundation
Days 31–60
Optimization
Days 61–90
Scale
Four tools, one genuinely important design difference, and a free tier good enough that most small teams should not be paying at all.
AI agents and traditional automation tools like Zapier solve different problems. Here is a clear framework for when each one is the right choice.
Most enterprises do not have an AI adoption problem. They have an AI inventory problem — and the first honest count is usually three to five times what anyone predicted.
One practical AI workflow per week. No fluff.
Get the full guide with step-by-step setup, workflow templates, and copy-paste assets.