I keep seeing invoice processing built as a crew of AI agents that burns hundreds of dollars a month in tokens, when a script plus one LLM call does the exact same thing for less than 10 shekels. Or lead routing “upgraded” to an autonomous agent, and suddenly leads go to the wrong teams because of a small hallucination, while the Zapier flow it replaced did the job with zero extra cost.

Part of the problem is naming. Every simple automated process, including scripts that have been running for 20 years, now gets the title “AI agent”. The mislabeling leads to bad decisions: the wrong tool gets chosen, and real money gets wasted.

Using a full agent for a task that a script or a regular workflow solves in 10 minutes costs more (reasoning loops, multiple model calls, retries), runs slower, is harder to debug, and is often less reliable because it “thinks too much” about simple things.

Most of what we automate today needs boring, safe automation, not an agent that tries to act like a human. There are 3 tiers, and picking the right one is the whole game.

Workflow automation

A rule-based process that performs repetitive tasks predictably: if X happens, do Y. Scripts that process files, ETL in Airflow, Zapier flows connecting apps. No thinking, no adaptation, just reliable and cheap execution.

AI-enhanced workflow

The same linear, predefined flow, with one step that calls an AI model as a black box. In an email processing flow: step 1 scans the message, step 2 classifies it with a model (spam/urgent/order), step 3 sends a notification. The tools are an OpenAI or Google Cloud AI API inside a flow like Make or n8n.

The AI adds the ability to handle unstructured content (classification, extraction, sentiment), but the process itself stays linear and predefined. This tier covers most of the cases people reach for agents to solve.

Autonomous agent

Receives a goal, not a script. It plans, decides the order of actions, chooses tools, corrects mistakes, and asks for clarification when stuck. It can handle a request like “plan me a family trip on budget X”, including weather checks, bookings, and adjustments along the way.

Checklist: do you really need an agent?

  • Does the process repeat in 95%+ the same way every time? → Workflow automation
  • Is most of the data structured, with clear if-then rules? → Workflow automation
  • Do you need AI in only one specific step (classification, extraction, generation)? → AI-enhanced workflow
  • Does the system need to decide the order of steps on its own? → Autonomous agent
  • Does it need to pick tools dynamically based on what it discovers? → Autonomous agent
  • Does it need to recover from failures on its own (try a different approach, switch tools, ask a human)? → Autonomous agent
  • Does the “right” way to perform the task change mid-process as new information arrives? → Autonomous agent
  • Is the task so variable or creative that fixed rules are impossible or need endless maintenance? → Autonomous agent (start with a prototype)
  • Are you willing to accept higher costs, longer debugging, occasional weird behavior, and the need for guardrails? → Only then build a full agent
  • Can you get 80% of the value with simple rules plus one LLM call in a regular flow? → Do that first

In short: start boring. Build the simplest thing that works. Move to an autonomous agent only when the workflow has proven it is hitting its limit, and you have the budget to monitor token spend.