AI Sales Academy
en

A sales agent architecture

Automation 13 min
Business problem

A workflow does repetitive steps perfectly, but trips when a decision is needed: 'what to do with a lead of an unusual profile?', 'which research to run first?', 'is this reply really interest?'. An agent adds a decision layer: it chooses the tools and the action order itself. The risk: without control an agent does things you didn't want — hence human approval.

Proven method

The agent architecture in 5 elements:

  1. TRIGGER: an event that starts the agent (a new lead, a reply, a status change, a cron),
  2. AGENT: the decision core (LLM) with access to tools: SEARCH (research, news), CRM (deal data, updates), EMAIL (sending — limited), RESEARCH (briefs), LLM (generation, analysis),
  3. DECISION: the agent picks a plan: 'this lead looks qualified, but signals are missing — I'll run research and check hiring',
  4. ACTION: execution (research → brief → CRM update → a draft proposal),
  5. HUMAN APPROVAL: key actions (sending, contact, big CRM changes) are approved by a human — the agent prepares, the human decides. Rules:
  6. an agent = LLM + tools + a loop (not a single call),
  7. every action is logged (what it did/why),
  8. scope: an agent has clear boundaries (what it does NOT do without consent: sending, promises, prices),
  9. human approval for actions with consequences,
  10. start with 1 agent with a narrow scope (e.g. a research agent), not with 'everything at once'.
Market example
BAD
Manual, inconsistent execution or chaotic automation lacking data validation.
GOOD
The 'Prospecting Assistant' agent (a narrow scope): Trigger: a new qualified lead. Agent: 1) decides 'I'll run research' → Search (website, news), 2) decides 'I'll check hiring' → job posts, 3) builds a Prospect Brief, 4) updates the CRM (a note + score + tags), 5) proposes an L4 opener draft. A Decision at the end: 'the lead is ready for the campaign — I need approval'. A human reviews (brief + opener) and approves. The agent never sends without consent. The agent did 4h of research work in 10 minutes — the human spent 3 minutes approving.
Ready-to-use prompt 
You are an AI agent architect.

Context: I want to build an agent for: [SCOPE — research/qualification/inbox]. I have: [TOOLS: SEARCH/CRM/EMAIL/LLM/API]. Existing workflows: [WHICH].

Task: Design a sales agent:
1. Trigger: what starts the agent (a list of events)
2. Scope: what the agent MAY do (boundaries) and what it does NOT do without consent
3. Agent tools: what access it has (Search, CRM, Email, Research, LLM) + for what
4. The decision loop: how the agent picks a plan (a sample scenario step by step)
5. Actions: what the agent executes + how it logs
6. Human approval: which actions a human approves (a list) and how to organize approval
7. Risks: what can go wrong (unwanted actions, costs, hallucinations) + mitigations
8. Start: which ONE agent you'll build first (a narrow scope) + a 30-day plan

Format: 8 sections.
Model rekomendowany: Claude 3.5 Sonnet / GPT-4o Copy and paste into ChatGPT, Claude or Perplexity
ALTERNATYWNE SCENARIUSZE

Variations & alternative scenarios

No-API version: an agent = prompt + workflow (a pseudo-agent on Make/n8n).

Full version: an agent with tools (API, search, CRM) + a decision loop.

Practical exercise

Define the scope of your first agent (what it MAY, what it does NOT). Write the trigger and tools. Design the decision loop on 1 scenario. Set 2 actions requiring human approval.

Zapisano w przeglądarce

Steps to take

0/4