The Autonomous Sales Agent Playbook

<aside>
📘

Welcome. This playbook shows how to replace a $5,000/mo marketing stack with one autonomous sales system.

</aside>

Image

My name is Benjamin and I’m the co-founder of IntentBot.co

We built an AI agent to help B2B companies find and contact warm leads.

Goal: set up your agent in 10 minutes and get new warm opportunities and conversations every day.

We have a 7‑day trial available here: IntentBot.co

Here’s a quick video explaining what we do:

https://www.youtube.com/watch?v=t5YeSDBfT4g

Try 7 days for free here: IntentBot.co

This guide walks you through the exact architecture, workflows, and prompts to build an autonomous sales agent that captures intent, researches prospects, and follows up automatically.

Executive summary

Most founders end up with a messy subscription pile.

GoHighLevel, Zapier, Calendly, a VA… it adds up fast.

The punchline: you can easily spend $1,000+/month and still respond too slowly to the people who are ready to buy.

<aside>
🤖

The Autonomous Sales Agent isn’t a chatbot widget.

Think of it as an always-on operator that handles speed, context, and follow‑up without dropping the ball.

</aside>

The one‑pager architecture

•
Eyes: Intentbot.co — detects high‑intent signals.
•
Body: OpenClaw — executes workflows and moves data.
•
Brain: Claude 4.6 Sonnet — analyzes context and writes human‑level messages.
•
Memory: Markdown files — a lightweight CRM in plain text.

Visual (high-level):

Eyes → Body → Brain → Memory

Eyes = intent signals • Body = workflows • Brain = reasoning + writing • Memory = Markdown CRM

Back-of-the-napkin ROI

Close a single $1,000 deal and the economics make sense immediately:

•
model/API spend is basically paid for
•
infrastructure cost becomes negligible

Low downside, high leverage.

The cost of inaction

MetricThe old way (Human + SaaS)The new way (Agent stack)
Response time~47 hours (avg)&lt; 2 minutes
Follow‑up rate~30% (humans give up)~100% (until they clearly decline)
Monthly cost~$1,447/mo~$20/mo
MaintenanceHigh (integrations break)Low (code is stable)

Stop renting your revenue engine. Own it.

What you’ll build in this playbook

1.
Part 1: Speed‑to‑lead — the business case for replying fast.
2.
Part 2: The stack — Brain, Body, Eyes, Memory.
3.
Part 3: The Context Engine — signal → research → decide → write.
4.
Part 4: The file‑based CRM — a lightweight system you can own.
5.
Part 5: The workflow set — 5 repeatable automations.
6.
Part 6: Human‑in‑the‑loop — confidence gates, approvals, kill switch.
7.
Part 7: How to ship it — options depending on your time and skills.

Part 1 — Speed wins (the speed-to-lead case)

If you’re still replying manually, you’re paying a hidden tax: latency.

Every minute you wait turns interest into distraction.

Why speed compounds

A large study (Harvard Business Review, 2,241 companies) found the average lead response time was about 47 hours.

In practice, attention decays fast:

•
Replying within 5 minutes can make you ~21× more likely to qualify than waiting 30 minutes.
•
After the first few minutes, conversion odds drop sharply.

Reason: your prospect is in “buying mode” right now. Later, they’re back to meetings.

The bloat audit (how stacks happen)

Founders try to patch speed with tools, then patch the tools with more tools.

Soon you’re juggling a brittle chain of apps that break quietly.

At that point you’re not “doing sales”—you’re doing integration maintenance.

Stack vs agent (cost snapshot)

Old way (roughly):

•
GoHighLevel (Agency): $297/mo
•
Zapier (Pro): $74/mo
•
Calendly (Teams): $16/mo
•
Mailchimp (Standard): $60/mo
•
Part‑time VA: $1,000/mo

Total: ~$1,447/mo

Agent stack (roughly):

•
OpenClaw (self‑host): $0/mo
•
TidyCal (one‑time): $29 lifetime
•
Resend (email): $0/mo (free tier)
•
Claude API: ~$15/mo
•
VPS hosting: ~$5/mo

Total: ~$20/mo

98% reduction in overhead.

Part 2 — The agent stack (how the system is built)

Chatbots wait.

Agents act.

The difference is execution: agents wake up on signals, do the work, and push outcomes forward.

The 4 components

1.
Brain (Claude 4.6 Sonnet)
•
Why: most models sound robotic
•
Fix: Claude handles nuance and writes like a senior copywriter
1.
Body (OpenClaw)
•
Why: you need an execution environment (Zapier is expensive + breaks)
•
Fix: OpenClaw connects APIs and runs code (your “hands”)
1.
Eyes (Signal detection)
•
Why: you can’t sell to what you can’t see
•
Fix: continuously detect high‑intent signals and turn them into structured events
1.
Memory (Markdown)
•
Why: databases are often overkill and hard to migrate
•
Fix: one file per lead, readable forever

Code vs no‑code (the trap)

“No‑code” is sold by SaaS companies.

It’s fine for prototypes and terrible for production:

•
slow (each step adds latency)
•
brittle (one API change breaks the chain)
•
expensive at scale (task pricing)

OpenClaw runs on your server.

Cost stays flat.

Infrastructure (VPS)

Don’t run this on your laptop.

When you close your MacBook, your employee dies.

Run it on a small VPS (DigitalOcean / Hetzner):

•
~2GB RAM / 1 CPU
•
~$5/mo

Cheat sheet: required keys

ServicePurposeCost
Anthropic APIBrain (Claude)Pay‑as‑you‑go
OpenAI APIBackup brain (optional)Pay‑as‑you‑go
GojiberryAISignal detectionSubscription
Resend / SendGridSending emailsFree tier
Slack webhookInternal alertsFree
Search APIResearch prospectsFree tier / usage‑based

Part 3 — The Context Engine (signal → research → decide → write)

Most automation fails because people automate typing, not thinking.

A bot that says “Thanks for your message!” adds no value.

The Context Engine flips the script: collect context first, then decide intent, then write a response that actually fits.

Phase 1 — Signal detection

Speed is the only variable that matters.

We don’t “check” for leads once a day. We listen continuously.

Example flow:

1.
You publish content / run campaigns
2.
A high‑intent signal happens (comment, DM, inbound request)
3.
The system captures it instantly
4.
A structured event is created and forwarded to your workflow engine

Phase 2 — Research loop (the context file)

Before the agent writes one word, it must learn who it’s talking to.

Research workflow:

1.
Extract sender’s profile URL + company URL
2.
Pull profile + recent posts
3.
Search company name + “news” / “funding”
4.
Synthesize into a JSON dossier

Example dossier output:

JSON
{
  "name": "Alex Smith",
  "role": "VP of Sales",
  "company": "TechFlow",
  "recent_news": "Raised Series B last week",
  "pain_point": "Scaling outbound team",
  "tone_match": "Direct, professional"
}

Phase 3 — Intent triage (priority rules)

Not all leads are equal.

You need a strict system that decides speed + channel + escalation.

Intent typePriorityDescriptionAction
BuyerP0Explicit intent (“Price?”, “Sign up?”)Instant response (Slack + email)
Call requestP0Wants a meeting (“Demo?”, “Free Tuesday?”)Send calendar link
Product inquiryP1Specific feature questionAnswer + case study
Service interestP1Vague interest (“Tell me more”)Qualify + value prop
SupportP1Existing customer issueRoute to support
NetworkingP2Partnership / collab requestDraft gentle decline/delay
Spam / salesP3Someone selling to youIgnore / archive

Logic:

•
P0: wake founder (SMS/Slack), respond immediately
•
P1: draft reply → approval or auto‑send if confidence &gt; 90%
•
P2: auto‑reply with “reviewing” template
•
P3: delete/archive

Phase 4 — The copywriting engine (frameworks)

We don’t use templates. Templates are for amateurs.

We use frameworks selected based on intent.

Framework A — Context → Value → CTA (for P0 buyers)

•
Context: reference the signal (funding, hiring, post)
•
Value: 1 sentence of credible outcome
•
CTA: one clean next step

Framework B — PAS (for P1 inquiries)

•
Problem: reflect the pain
•
Agitation: underline the cost of the status quo
•
Solution: your approach + proof

Framework C — BAB (for cold/warm leads)

•
Before: current state (chaos)
•
After: desired state (clarity)
•
Bridge: how you get there

Visualizing the flow (logic)

1.
Signal received
2.
Spam? → yes: archive • no: continue
3.
Research (profile + company + news)
4.
Build dossier (JSON)
5.
Classify priority (P0 / P1 / P2 / P3)
6.
Draft reply
•
P0: send + alert founder
•
P1+: queue for review (or auto‑send if confidence is high)
1.
Log to Markdown CRM

This loop can happen in ~10 seconds.

Part 4 — The File‑Based CRM

Stop paying for a heavy CRM you don’t use.

Philosophy

If it isn’t plain text, you don’t own it.

Structure: one lead = one file

Every prospect gets a single .md file.

The filename becomes the unique ID (e.g., john-doe-stripe.md).

“Folders as stages” (state machine)

01_Inbox → 02_Qualifying → 03_Proposal → 04_Closed

Dead ends: 02_Qualifying → 99_Dead, 03_Proposal → 99_Dead

•
/01_Inbox/ — new leads land here
•
/02_Qualifying/ — active outreach
•
/03_Proposal/ — meeting booked, you take over
•
/04_Closed/ — won
•
/99_Dead/ — unqualified/unsubscribed

Moving a file triggers the next action.

Master lead template (copy/paste)

YAML
id: "lead_001"
name: "Alex Smith"
company: "TechFlow"
role: "VP of Sales"
email: "alex@techflow.io"
linkedin: "https://linkedin.com/in/alexsmith"
status: "active"
stage: "02_Qualifying"
last_contact: "2026-03-01"
next_action: "follow_up_email_3"
value_potential: 20000
# tags: ["series-b", "hiring", "python"]

# Research Dossier
- Company news: Just raised $15M Series B.
- Pain point: Scaling team, drowning in admin.
- Personal interest: Writes about outbound systems.

# Interaction Log
## 2026-03-01 — Agent (Outbound)
Sent Email 1: "Saw the Series B news..."

## 2026-03-02 — Prospect (Reply)
"Thanks. We’re swamped. What do you do?"

## 2026-03-02 — Agent (Reply)
"We automate X. Saves you Y hours/week. Free Tuesday?"

Part 5 — The 5‑Workflow Arsenal

Your agent is only as good as the workflows you give it.

Skill 1 — Outbound campaign builder

Most founders blast a list of 1,000 leads with the same template.

That’s how you get flagged.

The fix: unique outreach, per person.

Workflow:

1.
Drop a CSV into /inputs/prospects.csv
2.
For each row:
3.
research profile + company
4.
extract a specific “hook” (e.g., “just hired a new CMO”)
5.
draft a 4‑email sequence referencing the hook
6.
schedule emails via Resend/Gmail

Output: 100 emails over 48 hours, 100% unique, 0% template.

Practical prompt template:

Plain text
You are an expert copywriter.

CONTEXT:
Prospect Name: {name}
Company: {company}
News Hook: {news_hook}
Pain Point: {pain_point}

TASK:
Write a 3-sentence cold email.
1) Hook: Reference the {news_hook} naturally.
2) Bridge: Connect their {pain_point} to our solution.
3) Ask: Soft CTA (e.g., "worth a chat?")

CONSTRAINT:
Do not be formal. Write like a busy colleague.
No hype. No exclamation points.

Skill 2 — Inbound content monitor

You post a case study.

People comment: “link”, “interested”, “send me this”.

If you reply 6 hours later, they’re gone.

Workflow:

1.
Detect keyword comment
2.
Research the commenter
3.
Check ICP fit
4.
If fit: DM the link + 1 qualifying question
5.
If not: reply publicly with the link

Output: high‑intent leads captured in &lt; 60 seconds.

Regex patterns (examples):

•
/send|link|guide|pdf|doc/i (asset request)
•
/interested|how much|cost|price|demo/i (purchase intent)
•
/dm me|message me|chat/i (conversation request)

Skill 3 — Trial‑to‑paid nudger

Most free users churn because they never set up properly.

Generic onboarding emails don’t help.

Workflow (example):

1.
Trigger: Stripe webhook customer.subscription.created
2.
Wait 3 days
3.
Query usage:
•
Scenario A (inactive): send “stuck?” email + 1‑min setup video
•
Scenario B (active): send “power user” tip + advanced workflow

Skill 4 — Win‑back agent

“Dead” leads are often just dormant.

Workflow:

1.
Monthly run
2.
Pull list of closed‑lost / churned
3.
Check if champion changed jobs
4.
If yes: send a congrats + re‑engagement email

Practical prompt:

Plain text
Analyze this LinkedIn profile: {profile_url}
Compare with our CRM record:
- old_company: {old_company}
- old_title: {old_title}

Determine:
1) Have they changed jobs? (Yes/No)
2) Is the new company a fit for us? (B2B SaaS, >10 employees)
3) Draft a short congrats + re-engagement email.

Skill 5 — Market intel briefing

You can’t spend 2 hours/day reading news.

The agent can.

Workflow:

1.
Every Monday 8:00
2.
Search competitor news + funding + hiring
3.
Filter noise
4.
Cross‑reference with your active pipeline
5.
Summarize into 5 bullets
6.
Send to your Telegram/Slack

Example format:

•
Competitor alert: X raised $10M; hiring 5 SDRs
•
Prospect news: TechFlow mentioned in TechCrunch
•
Trend: “AI agents” search volume up +40%
•
Pending approvals: 3 drafts waiting

Part 6 — Deployment & Human‑in‑the‑Loop

You’re afraid:

“What if the AI hallucinates? What if it insults someone? What if it offers a discount by accident?”

Valid fear.

We build human‑in‑the‑loop systems.

Safeguard 1 — Confidence score

Before any message is sent, the model must rate confidence 0–100.

Rules:

•
confidence &gt; 90% → auto‑send (routine scheduling, simple questions)
•
confidence &lt;= 90% → send draft to approval queue

Approval workflow:

1.
Agent drafts
2.
Agent self‑scores (e.g., 75%)
3.
Bot pings you with two buttons: APPROVE / REJECT
4.
You approve
5.
Agent sends

Safeguard 2 — The kill switch

Software breaks. APIs fail.

You need an emergency brake.

Pattern:

•
global variable: SYSTEM_STATUS = "ACTIVE"
•
if anything looks weird: send /STOP
•
system freezes instantly (no emails, no DMs)
•
fix bug → /START

The new daily routine (operator mode)

You’re not an SDR anymore. You’re a manager.

You review, not do.

Morning (8:00)

•
check market intel briefing
•
see who raised / who’s hiring

Time: ~5 min

Noon (12:00)

•
clear approval queue
•
approve/edit a handful of drafts

Time: ~10 min

Evening (17:00)

•
check pipeline folder (/03_Proposal/)
•
see meetings booked

Time: ~2 min

SOP — the 15‑minute daily audit

Checklist:

•
check logs: did the cron run at 9 AM?
•
check inbox: are there urgent P0s waiting?
•
spot check: read 3 sent emails (quality control)

If quality drops, tweak prompts.

You don’t fire the employee. You retrain it.

Part 7 — Stop Renting Software. Start Owning Agents.

You now have the blueprint.

This is not about saving money.

It’s about speed.

It’s about control.

It’s about freedom.

You have four choices:

Use Intentbot as the eyes.

Even without a full agent stack, signal capture alone can multiply warm conversations.

Option 2 — Do it yourself (DIY)

You have the guide.

Spin up a VPS.

Install OpenClaw.

Plug in your signal detection.

Build the machine.

Cost: time + effort.

Start now

→ Start your warm AI engine: https://intentbot.co/ (trial)

⇒ Book a call with the founder

If you’d like extra guidance:

1.
Follow the full step‑by‑step blueprint

LinkedIn High-Intent Outreach Blueprint (Book 12 Demos in 5 Days)

Bonus

LinkedIn high‑intent outreach: the short version

A practical add‑on on how to turn intent signals into conversations (without spamming).

Why high‑intent beats cold

LinkedIn outreach works when you show up at the right moment, not when you blast a list.

Cold outreach (classic)

•
You push a message to someone who did nothing to deserve it
•
You fight defenses from line 1
•
You get 1–2% response rates

!image_blueprint_1.png

High-intent outreach (what we do)

•
You message someone who just signaled interest
•
Your timing does most of the work
•
You often get 25–40% response rates

!demo_linkedin_metrics.png

What counts as an “intent signal”?

A signal is any public action that screams:

“I’m in the market.”

Example: someone commenting “interested” on a relevant post is basically raising their hand.

!image_blueprint_2.png

Step 2 — The 300 high-intent leads framework

We generated 300 high-intent leads in a week from one agent search

Not with magic

With rules.

2A) Define high-intent (properly)

High-intent is not “fits the ICP”

High-intent is “fits the ICP and something is happening right now.”

Look for triggers like:

•
Liking or commenting on competitor or category posts
•
Announcing funding
•
Hiring for roles that imply budget and urgency
•
Changing roles (new mandate, new problems)
•
Publicly sharing pain points

Rule of thumb: if it happened this week, it’s worth a message.

2B) Turn signals into a daily lead flow (10 minutes)

1.
Create an account: intentbot.co
2.
Describe your ICP (5 minutes)
3.
Activate your agents (5 minutes)

!1512w default.png

Enter your ICP

!flutter-view → image.png

Activate your agents

!flutter-view → image 2.png

With IntentBot, you can run multiple “intent sensors” in parallel:

•
Competitor engagement
•
Influencer engagement
•
Job changes
•
Top 5% ICP
•
Funding alerts
•
Engagement & interest
•
Your company

2C) Collect fresh high-intent leads every day

!demo_linkedin_leads.png

This is the part people underestimate.

When leads come in daily, you stop “prospecting”.

You start having conversations.

Step 3 — The messaging sequence that feels human

Most people ruin high-intent leads with a high-effort pitch.

Long message.

Lots of claims.

Zero replies.

So we do the opposite.

We keep it simple.

Message 1 — One question (no pitch)

Plain text
Hey [Name],

Quick question — what’s your biggest challenge with [topic] right now?

That’s it.

No positioning.

No “we help”.

Just a real question.

Message 2 — Bridge to Loom (after they reply)

Plain text
That makes sense — we just solved that for [similar company].

Want me to send a 3-minute video showing how?

If they say yes, send a Loom that:

•
Mirrors their situation in the first 10 seconds
•
Shows 1 relevant result
•
Explains the mechanism (not the feature list)
•
Ends with a soft CTA

Message 3 — Clean follow-up (if they don’t answer)

Plain text
Hey [Name] — quick bump.

Noticed you’re [specific observation].
If [topic] is a priority, I can share a quick example.

Personal.

Short.

No pressure.

Step 4 — AI prompts (only if you need them)

You can do this manually

But if you’re running volume, you’ll want help.

Here are prompts that keep the tone human.

•
Research prompt
Plain text
    Look at this person’s LinkedIn profile: [profile URL]
    
    Identify:
    1) Their most likely current business challenge based on their recent posts
    2) Something specific they seem to be working on
    3) A genuine, specific compliment
    
    Keep it under 50 words.
•
DM writing prompt
Plain text
    Write a 30-word LinkedIn DM to [Name] who just showed this intent: (describe intent)
    
    What we learned: [paste research]
    
    Rules:
    - Sound like a real person
    - Reference 1 specific detail
    - Ask 1 simple question
    - No sales language
    - No exclamation points
    - Max 30 words
•
Follow-up prompt
Plain text
    They responded: "[their response]"
    
    Write a 25-word follow-up that:
    - Acknowledges their situation
    - Adds 1 relevant proof point
    - Suggests 1 next step
    - Stays conversational

Step 5 — Your first week (simple rollout)

Day 1

•
Pick 2–3 signals
•
Define your ICP
•
Turn on the agents

Day 2

•
Take the freshest leads
•
Send Message 1 (one question)

Day 3–5

•
Reply fast
•
Send Looms
•
Book calls

Day 6–7

•
Review what converted
•
Double down on the best signals

What to remember

You don’t need better copy

You need better timing

Intent fixes timing

And timing fixes conversion !

→ Start now

Start the 7-day trial: intentbot.co

!image_blueprint_4.png

Made with Looply