An AI agent or plain automation? How to choose for a business process

“AI implementation” can mean three different things. The choice between them decides the cost, how often something goes wrong, and whether anyone can explain why the system did what it did.

Three levels

1. Rule-based automation. “When an invoice arrives at invoices@, save the attachment to a folder and add a row to the spreadsheet.” The steps are known in advance and the system performs them the same way every time. Tools such as n8n, Make or Zapier do this cheaply and predictably.
2. Automation with an AI step. The same workflow, but at one point a language model does something a rule can’t: reads a message and classifies it, pulls data out of a non-standard document, drafts a reply. You still decide the order of the steps.
3. An agent. The model gets a goal and a set of tools (searching a database, sending an email, creating a task) and decides by itself which to use and in what order, until it judges the goal has been reached.

The higher you go, the more flexibility you get, but also less predictability and a higher cost for every run.

How to choose

Ask five questions about your process:

• Are the steps always the same? If so, rule-based automation is enough. An agent adds nothing here but risk.
• Is the input messy? Emails written by people, scans, meeting notes. Then an AI step that turns them into structured data is useful.
• Does the order of steps depend on the case? For example, a customer enquiry sometimes needs an order lookup, sometimes the documentation, and sometimes a handover to sales. Only here does an agent start to make sense.
• What does a mistake cost? A wrongly assigned category in the CRM is a trifle. A wrongly sent quote or refund isn’t. The more a mistake costs, the lower on the list you should stay, or the more human checking you need to add.
• Do you need to explain decisions? Processes subject to audit or regulation need a trail of why something happened. Rules give you that out of the box; an agent has to be prepared for it.

In practice, level two is most often the best: a predictable workflow with one or two AI steps and human approval where mistakes are costly.

If it is an agent

An agent makes sense when cases really do differ and handling each one by hand takes a lot of time. Then a few things are worth building in from the start:

• The least privilege possible. An agent that needs to read orders doesn’t need the right to cancel them. Every tool that changes something or sends something outside is a separate decision.
• A human at irreversible steps. Sending a message to a customer, a change in the accounting system, a payment: the agent prepares, a person approves. Especially at first.
• Limits. A maximum number of steps and a maximum cost per task. An agent stuck in a loop can produce a surprising bill.
• A log of every step. What the agent saw, which tool it chose and with what result. Without it, errors are hard to fix and the team’s trust is easy to lose.
• A set of test cases. A few dozen real, anonymised cases with the expected outcome. After every change to the prompt or model, you check that results haven’t got worse.

Language models can be wrong with complete confidence. We wrote about how to limit that in our article on AI hallucinations.

How to measure it

Before the rollout, measure the current process: how many cases a month, how long one takes, how many errors there are. After the rollout, measure the same, plus the share of cases the system had to hand back to a person and the cost per case. If you can’t measure it, it will be hard to say whether the rollout paid off.

Where to start

With a list of processes, not with the technology. Describe what happens step by step, where the messy data is and what a mistake costs. That list will tell you where a rule is enough and where something more is needed. We described how to prepare it in our article on the AI process audit.

Every implementation we do starts with that audit. AI Boost starts at $3,200 / €2,950 net. Read more on our AI Boost service.