AI agent or plain automation? How to choose the right one

Everyone wants to sell agents right now. It is understandable. The word sounds bigger than automation. It implies autonomy, intelligence and the future. But in practice, many "AI agents" are just ordinary workflows with a language model step in the middle. What is bad is when you use an agent where a simple rule would have been better.

Three levels: workflow, copilot and agent

Workflow: when the rules are clear. If an order is missing a field, send it to the right queue. If a return has a certain reason code, create a task. If stock drops below a threshold, notify purchasing. This does not need AI. It needs good integrations, clear rules and stable operations.

Copilot: when the human still owns the decision. If customer service needs to summarise order history, product data and return policy, AI can help. If purchasing wants to see anomalies and forecast suggestions, AI can suggest. If the product team needs to improve copy and attributes, AI can do the first pass. But the human approves.

Agent: when the system can read, reason, act and follow up within defined boundaries. An agent can, for example, handle recurring order exceptions, suggest actions, write back to the correct system, escalate when confidence is low and log decisions. But it should not be free. It should have mandates, rules, traceability and stop points.

How clearly defined is the work?

The important question is not how smart the model is. The important question is how clearly defined the work is.

If the flow is stable, use a workflow. If the flow requires judgement but human accountability, use a copilot. If the flow is recurring, data-heavy and can be bounded by rules, use an agent.

A good rule of thumb:

• If the same thing should happen every time: workflow.
• If someone needs to think faster: copilot.
• If the system should drive work forward: agent.

Start with observation, not autonomy

This also makes implementation safer. You do not start with autonomy. You start with observation. Then suggestions. Then approved actions. Then limited automation. Only after that, agent operations.

Many AI projects fail because they skip those steps. They start with "the agent should do everything" and then discover that data, process and accountability are not ready.

lights-out.ai builds in the opposite direction. First we define the work. Then we build the smallest solution that delivers value. If a deterministic workflow is enough, we use that. If AI is needed, we use AI. If agent operations are right, we build in guardrails before it is allowed to act.

It is less spectacular than promising a self-governing organisation. But it is how you get something that actually runs on Monday morning.