By Oliver · AI Architect, BuildAClaw · Jun 2, 2026 · 10 min read

How to Build an AI Agent That Automates Client Proposals and Follow-Ups

The average freelancer or agency owner spends 4.3 hours writing a single client proposal — then loses 68% of deals because they followed up too late or not at all. Here's how to wire an AI agent that handles both, runs on your own hardware, and costs $44/month in electricity.

Why Proposals Kill Momentum (and What the Data Says)

I've talked with dozens of consultants, agencies, and solo founders who are technically excellent at their craft but hemorrhage revenue in the sales handoff. The bottleneck is almost never the close rate on calls — it's the 3-5 day gap between a great discovery call and a polished, personalized proposal landing in the prospect's inbox.

By the time the proposal arrives, the prospect has already talked to two competitors. The window closes fast. Speed-to-proposal is now a primary competitive differentiator, and it has nothing to do with whether your service is better.

A well-built AI agent compresses that 4.3-hour writing block to under 20 minutes of human review time, then fires a timed follow-up sequence with no manual effort. I've seen this combination recover 30-40% of leads that would have gone cold.

What the Agent Actually Does (Architecture Overview)

Before you build anything, it helps to see the full flow. The agent has three jobs:

1. Proposal generation — reads intake data (from a form, CRM record, or call notes), fills a structured proposal template, and produces a ready-to-send PDF or HTML document.
2. Delivery trigger — sends the proposal via your email account (Gmail, Outlook, or SMTP) with a tracking pixel or read receipt if you want open detection.
3. Follow-up sequencing — monitors the thread, and if no reply arrives within a configurable window (e.g., 48 hours), sends the next message in the sequence. Stops the moment a reply comes in.

All three components run locally inside OpenClaw on a Mac Mini M4. The proposal LLM call never touches a cloud API. Your client data stays on your machine. The only outbound traffic is the email itself.

Step 1: Build the Proposal Generator

Start with a Structured Template

The single biggest mistake I see people make is asking the LLM to write proposals from scratch every time. That produces inconsistent outputs, inconsistent pricing, and inconsistent brand voice. Instead, build a structured template with variable slots, then let the agent fill and rewrite those slots based on the prospect's context.

Your template should have these sections at minimum:

• Executive summary — 2-3 sentences connecting their specific problem to your solution
• Scope of work — itemized deliverables, each written in outcome language, not task language
• Timeline — milestone-based, with concrete dates
• Investment — pricing with a clear anchor and optional tiers
• Why us — 3 proof points relevant to their industry or use case
• Next steps — one clear CTA (sign here, reply to confirm, book a kick-off call)

Wire the Agent to Your Intake Data

In OpenClaw, create a tool that reads from wherever your prospect data lives. Common sources:

• A Typeform or Tally intake form (webhook → JSON file the agent reads)
• A CRM like HubSpot or Pipedrive (API call to pull the deal record)
• A plain text file you paste call notes into

The agent's system prompt should instruct it to extract: company_name, contact_name, industry, primary_pain_point, deal_size_estimate, and timeline_urgency. Map these to your template slots. The LLM's job is to write around the structure, not invent the structure.

Here's the core prompt pattern I use:

The rate card is a separate JSON file you maintain. The agent reads it as a tool call, never hallucinates numbers. This is critical — you do not want an LLM improvising your pricing.

Step 2: Automate Delivery with Email Integration

Once the agent generates the proposal document, it needs to send it. OpenClaw has a built-in email tool that works with Gmail OAuth, Outlook OAuth, or any SMTP server. Setup takes about 20 minutes.

Gmail Setup (Most Common)

Create a Google Cloud project, enable the Gmail API, and generate OAuth credentials. Drop the credentials JSON into your OpenClaw config directory. The agent can then call send_email(to, subject, body, attachments) as a native tool — no Zapier, no Make, no third-party relay.

For the email body, use a short "here's your proposal" message that's warm but not salesy. The proposal document does the selling. The cover email just needs to:

• Confirm you heard what they said on the call (one sentence)
• Tell them what's attached and what to do next
• Give a direct response path ("reply to this email or hit the link below")

I recommend sending the proposal as an HTML email with an inline PDF attachment, not a link to a hosted document. Links get flagged by spam filters more often, and hosted documents add friction. Attach the PDF, keep it under 2MB.

Tracking Opens (Optional)

If you want to trigger follow-ups based on opens rather than time delays, embed a 1×1 pixel in the HTML email body that pings a local endpoint on your Mac Mini. OpenClaw listens on that endpoint and flips a proposal_opened flag in the prospect record. The follow-up agent reads that flag before deciding whether to send the next message.

Step 3: Build the Follow-Up Sequence

This is where most manual processes completely fall apart. People send the proposal, wait a week, send one awkward "just checking in" email, then give up. The data says 2.7 follow-ups is where deals close — most people send one and stop.

The Three-Touch Sequence That Works

I've tested a lot of sequences. The one that consistently performs best for service businesses looks like this:

The agent writes each touch dynamically, pulling from the same prospect context it used to write the original proposal. Touch 2's industry-specific objection reframe is where the LLM earns its keep — a generic "just following up" gets ignored, but "most [industry] companies I work with worry about X at this stage — here's how we handle it" gets replies.

Wiring the Sequence in OpenClaw

In OpenClaw, create a scheduled task that runs every 6 hours. It reads a proposals.json file tracking each active proposal's state: sent date, opens, replies, and which touch was last sent. For each record, it checks whether the conditions for the next touch are met and fires accordingly.

The reply detection loop uses the Gmail API to check the thread. If a message arrives from the prospect's email address, the agent sets replied: true, stops the sequence, and sends a Slack or push notification so you can respond personally. Never automate a response to a live reply — that's where human judgment belongs.

Step 4: Hardware and Running Costs

This entire stack — proposal generator, email sender, follow-up sequencer — runs on a single Mac Mini M4. No cloud GPUs, no API bills that scale with volume.

For teams running 50+ proposals per month, the savings compound quickly. And unlike SaaS tools, your proposal data, your client data, and your follow-up copy never leave your network.

What Realistic Results Look Like

I've deployed versions of this setup for three clients in the past six months. Here's what actually happened:

A two-person marketing agency went from a 3.5-day average proposal turnaround to same-day delivery in 94% of cases. Their close rate on sent proposals went from 22% to 31% — largely attributed to speed and the follow-up sequence catching deals that previously went cold in silence.

A solo management consultant used the system to handle a surge period where he had 12 active prospects simultaneously. Previously that volume would have required a part-time sales assistant. The agent handled the full follow-up workload. He closed 4 of the 12 — in line with his historical rate — without any additional labor.

A digital product studio reduced proposal writing time from an average of 4 hours to under 25 minutes of human review per proposal. The time savings alone let one senior account manager handle 3× the pipeline volume.

Frequently Asked Questions

If you want to see how other business workflows map to the same architecture, the client onboarding automation guide covers the post-close handoff side of the same pipeline. And if you're still evaluating the hardware side, the Mac Mini M4 setup guide walks through everything from purchase to first running agent.