A scripted chatbot follows a relatively fixed conversation tree. An agentic lead engine can decide which step to take next, which permitted tool to call, what context to retrieve, and when to hand the conversation to a human. n8n’s 2026 Agents release makes this separation explicit: the agent receives a model and a selected set of tools or workflows, and workflows can also call the agent as a reusable capability.

The production principle is simple: the model should not receive unrestricted access to every system. It should receive narrow capabilities such as “check availability,” “create or update lead,” “retrieve approved service information,” or “request human handoff.”

DeepSeek can occupy the model layer if that is your chosen inference provider, but the architecture should keep the model replaceable. Business logic belongs in workflows, databases, schemas, permissioned tools, and explicit policy checks — not in one oversized prompt.

The retrieval layer should provide approved facts: service scope, current pricing rules, business hours, qualification criteria, supported locations, calendar rules, and escalation policies. The model’s job is to use those facts conversationally, not to create new commercial policy.

OpenAI’s agent guidance recommends automatic guardrails plus human review for sensitive side effects. The same architectural principle applies regardless of the model provider. A WhatsApp agent can safely answer routine questions, collect requirements, classify intent, retrieve policy, and prepare drafts. Consequential actions should pause for approval.

n8n’s own engineering guidance stresses that agents must be treated as software systems, not as prompts that magically supervise themselves. Messaging workflows are especially exposed to duplicated events, race conditions, concurrent replies, and malformed tool calls.

Production flows therefore need conversation locks, message deduplication, idempotency keys, durable session state, tool-level validation, and clear error handling. These controls determine whether an automation survives real traffic.

The model should extract observable evidence — service fit, geography, budget range, urgency, decision role, and missing information — into a structured object. Deterministic workflow logic can then calculate the score and choose the next action.

{
  "service_fit": "high",
  "location_supported": true,
  "budget_band": "known",
  "urgency_days": 14,
  "decision_role": "owner",
  "missing_fields": [],
  "recommended_action": "book_call"
}

No responsible production design should claim “zero hallucinations.” Retrieval, structured outputs, guardrails, and deterministic checks reduce the probability and impact of model errors; they do not make probabilistic models infallible.