Sales agents are not robo-SDRs. They're leverage for the humans who close.
Every company has tried "AI SDR" tools. Most produce generic outreach that recipients delete without reading. The pattern worked for email-blast era companies. It fails now because recipients can smell generated outreach in two seconds.
Sales agents that actually work don't try to replace the human. They do the work the human hates: research, qualification, scheduling, follow-through, pipeline hygiene. The human closes. The agent keeps the pipeline alive in between.
This module is 90 minutes of building one such agent. By the end:
- A qualification agent that reads a new lead and returns a scored brief.
- A follow-up agent that keeps no-response threads warm.
- A rule: what sales agents must never do.
Thinker.
Sales has three parts. Agents do two of them well. One is still a human job.
- Research and qualification. Agent job. Fast, thorough, cheap.
- Personalized outreach and follow-up. Agent job, with guardrails. Never fully automated.
- Closing. Human job. The emotional labor of building trust and asking for the deal doesn't transfer.
What sales agents must never do
- Send outreach no human reviewed.
- Claim to be a human.
- Make commitments (price, delivery date) the human hasn't approved.
Break any of these rules and you're building a spam cannon, not a sales agent.
The assist pattern
The agent drafts. The human edits and sends. The agent watches for replies. The agent drafts follow-ups. The human sends or skips. Two humans per loop: your salesperson and the prospect.
Talker.
The qualification prompt
You are a lead-qualification agent. Read the lead info and
return a scored brief.
Return JSON:
- fit_score: 0-10. How well this lead matches our ICP.
- urgency_score: 0-10. How urgent their problem is.
- reasoning: 2-3 bullets explaining the scores.
- research: 3 specific facts about the company we'd want
before reaching out (recent news, funding, hiring).
- opening_line: one sentence the human can use to open, NOT
a full email. Must reference one specific fact.
- red_flags: list. Any reasons not to pursue (compliance,
competitive, wrong persona, etc.)
Lead info: [paste]
Our ICP: [paste from icp-skill or context]The follow-up prompt
You are a follow-up drafter. The thread below has been quiet
for N days. Draft one follow-up message.
Rules:
- Under 60 words.
- References something specific from the earlier exchange.
- Asks one question, not more.
- Sounds like it was written by [rep_name] (voice spec:
[paste]).
- Marks itself as needs_human_review: true.
If the thread has had 3 previous follow-ups with no response,
return:
draft: "BREAK-OFF NOTE"
plus a breakoff-note template.
Break-off signals we stop following up, not more follow-up.The break-off rule matters. Sales agents without it become harassment.
Rememberer.
Sales agents run on CRM state. No CRM, no coherent agent.
[sales-agent]/
icp.md (ICP, updated quarterly)
voice/
rep-[name].md (per-rep voice spec)
playbook.md (stages, rules, breakoff policy)
logs/
qualifications.jsonl
drafts/ (every draft, by thread)The CRM as the spine
Agents read and write the CRM. Every lead has: qualification score, follow-up state, last human touch, last agent draft. If the CRM doesn't have those fields, add them.
The break-off ledger
Log every break-off. If the same company comes back 6 months later, the agent knows the history and doesn't start from zero.
Doer.
Twelve minutes. Ship a qualification agent that scores one real lead.
Step 1. Write your ICP (3 min)
If you don't have an ICP in writing, write it now. Company size, industry, role, budget signal, urgency signal. Save as icp.md.
Step 2. Build the qualification agent (4 min)
SDK app with the qualification prompt. Inputs: lead info (name, company, title, any notes). Loads icp.md. Returns JSON brief.
Step 3. Run it on 5 real leads (3 min)
From your pipeline. Real leads, not invented. Read every brief. Would you use the opening line? Would the research unlock a better conversation?
Step 4. Hand-tune (1 min)
If the scores feel off: tighten icp.md. If the opening lines are generic: tighten the prompt to insist on specificity.
Step 5. Commit and share (1 min)
Share the briefs with your sales team (if you have one) or your cofounder. Make it part of the weekly sales ritual.
Five scored leads. Three useful opening lines. A process that saves 15-30 minutes per lead research.
- Scores cluster at 5: ICP is too vague. Add specificity.
- Research is hallucinated: the agent had no data source. Give it a search tool (Module 015 MCP).
- Opening lines are generic: your prompt didn't insist on specificity. Add "must reference one specific fact from the research."
Rookie.
Failure 1. Auto-sending outreach
You wire the agent to send directly. Saves time. Deliverability tanks. Prospects report your domain. You're in a blacklist.
Fix: human sends every outreach message. The agent drafts. The human approves. Non-negotiable.
Failure 2. Pretending to be human
Agent signs messages "Alex, SDR." Prospect asks a follow-up. Agent answers. Prospect realizes it's a bot. Trust is gone.
Fix: if the agent drafts, the human signs. If the conversation involves the agent autonomously, disclose.
Failure 3. The infinite follow-up
Agent keeps following up. No break-off rule. Prospect is being nagged every 5 days by an agent that doesn't know it's overstepping.
Fix: break-off is hardcoded into the playbook. Three follow-ups then close the loop. More follow-ups require human decision.
Manager.
Sales agents inside a sales team need rules, rituals, and responsibility.
One playbook
The team agrees on: what gets agent-drafted, what doesn't, approval thresholds, break-off rules. Written down. In the repo.
The per-rep voice file
Every rep has their own voice.md (Module 013). Agents draft in the rep's voice. Prospects get consistency across touches. Reps don't feel replaced, the agent sounds like them.
The weekly pipeline review
Once a week, the team reads:
- Last 20 agent drafts, sampled.
- Every break-off triggered that week.
- Qualification scores vs. actual conversion signal.
Tune the prompt, ICP, or playbook based on what you see.
Chief.
Risk 1. Deliverability
Automated, generic outreach hurts deliverability for your whole domain. Sales agents that auto-send will poison the well for legitimate email.
Governance: no send without human review, volume caps per rep, deliverability monitoring.
Risk 2. Compliance
Regulated industries (healthcare, finance, legal) have specific rules about cold outreach, disclosures, record-keeping. Sales agents must comply.
Governance: compliance reviews the playbook. Certain industries get different rules or no agent at all.
Risk 3. Pipeline hygiene theater
Agents make pipelines look healthier than they are. Follow-ups logged, scores tracked, dashboards green. Revenue doesn't move.
Governance: measure the north star (closed-won deals, not pipeline activity). If agent activity goes up but revenue doesn't, the agent is doing motion, not work.
Founder.
Solo founder doing sales: the agent multiplies your reach without multiplying the load on you.
The solo sales stack
- Qualification agent on every new lead.
- Follow-up drafter for threads older than 5 days.
- Weekly pipeline digest: which leads are warm, which went cold, which to break off.
The 30-minute weekly review
- 10 min: read the agent's weekly digest.
- 10 min: approve or edit the drafts.
- 10 min: close out break-offs. Schedule outreach to the top 3 warm leads.
Sales agents don't close deals. They keep pipelines alive.
Closing is a human act. Research, qualification, drafting, and follow-up are not. The agent does the 80% that drains humans. The human shows up to the 20% that only a human can do. That ratio is where the leverage lives.