Prompt, Workflow, Agent or Agentic System: Which One Does Your Process Need?

2026-11-20来源 dev.toEnglish

"We want to build an agent" is how most automation conversations start now.

Half the time, the process doesn't need one. It needs a workflow. Sometimes it needs a better prompt.

Picking the wrong level costs you twice: more to build, and more to keep running. So before opening n8n, I ask teams to place their process on a four-step ladder.

The four levels

Level What the person does What the AI does
Prompt Asks one question Answers once
Workflow Defines the steps Runs those steps on new inputs
Agent Sets the goal, the tools and the limits Chooses the next step and uses tools to finish
Agentic system Reviews and approves at set checkpoints Several agents and workflows pass work to each other

Many teams get most of their value at the workflow level. That's worth saying out loud, because "agent" is the word everyone wants on the slide.

How to choose

Ask three questions about the process.

Are the steps the same every time? If yes, build a workflow. Fixed steps are easier to test, cheaper to run and easier to hand over.

Does the next step depend on what the AI finds? If the path changes with the input, such as deciding whether to search, read a file or update a record, that's a job for an agent with tools and limits.

Does work pass between several of these? Then you're designing an agentic system, and the design work is mostly about checkpoints.

Good first candidates are repeated weekly tasks with predictable inputs and clear outputs: reports, research summaries, ticket triage.

Build the first one in n8n

n8n is a visual builder, so non-technical teams can do this. Developers can take the same process further with Claude Code.

1. Map the process. Break one weekly process into steps. Mark which steps an agent can run and which need a person.

2. Build trigger, AI step, output. Connect a trigger, one AI model call and an output that lands in the right place. Keep it to those three until it works.

3. Give it tools, within limits. Let the AI search, read files or update records. Decide what it is not allowed to touch.

4. Add a human approval step. Nothing should reach a customer or a manager unchecked.

5. Add a fallback. Decide what happens when the AI is unsure. Usually that means routing to a person.

6. Test on real inputs. Not the tidy sample. Log the failures, fix them, and write down how the owner keeps it running.

Where the human goes

This is the part teams skip. A few typical patterns:

Function Typical workflow Human review point
Marketing Collect campaign data, draft the weekly summary Before sharing with leadership
Operations and MIS Pull figures, flag exceptions, draft the report Exceptions above a set threshold
Sales Research an inbound lead, draft a first reply Every message before sending
Customer support Classify tickets, suggest answers from the knowledge base Refunds, complaints, unclear cases
HR Screen policy questions, route with a draft answer Sensitive or personal cases

The rule of thumb: steps that are judgement-heavy, legally or financially risky, or customer-facing get an approval point.

Guardrails to agree before you build

  • Approval thresholds for exceptions
  • A fallback path when the model is unsure
  • Only company-approved AI tools and data
  • An owner for every workflow

Measure three things

  • Time saved. Weekly hours the workflow removes, estimated by the team that owns it.
  • Usage. How many workflows are still running after 30 days.
  • Quality. How often a reviewer has to correct the output before it goes out.

The 30-day number is the honest one. Plenty of automations work on demo day and are switched off a month later.

Why this is on every roadmap

Microsoft's 2025 Work Trend Index reports that 81% of leaders expect agents to be moderately or extensively integrated into their company's AI strategy in the next 12 to 18 months.

Expectation isn't readiness. Readiness looks like mapped processes, working workflows and human checkpoints.

I saw this at NPCI BHIM in 2025, where marketing, UI/UX and executive leadership teams joined a closed-door workshop and each turned its own repetitive processes into role-specific n8n use cases. The starting point was each team's own process, not the word "agent".

A one-line test

Before you build, finish this sentence: "A person approves this at the point where ___."

If you can't fill the blank, you're not ready to automate it yet.

Originally published on Special Forces of AI: https://specialforcesofai.com/blog/agentic-ai-trainer-india

原文:https://www.a8q9.cn/en/article/prompt-workflow-agent-or-agentic-system-which-one-does-your-process-need-228618.html

相关文章