The most important AI skill is knowing when not to use AI.
When a company discovers agents, it wants them everywhere. That's understandable, agents can do a lot. But for many decisions, an agent is needlessly expensive, slow and unpredictable. A good system doesn't use as much AI as possible, but the right kind of decision-making in the right place.
Four different “workers” make decisions in a company.
Every recurring decision can be made in one of four ways. It isn't a ladder with the agent at the top. Each has its role.
| WHO DECIDES | WHEN IT FITS | EXAMPLE |
|---|---|---|
| A person | High risk, accountability, an unclear situation | Hiring someone, legal interpretation, a large investment |
| A rule in a program | The logic is clear and doesn't change | Price calculation, due date check, access rights |
| A statistical model | Patterns in data you can't see with the naked eye | Demand forecast, estimating customer churn |
| An AI agent | Unstructured inputs that need to be understood and put into words | Reply to a free-form email, contract summary |
With a person
Your company has a policy: “invoices up to a certain amount are approved by the head of department”. Nobody thinks about it, everyone follows the rule. You call an experienced colleague only for a case the policy doesn't cover. And you want the managing director's signature on things someone has to answer for.
With AI
Automation should be divided exactly the same way. Clear cases are handled by a rule. Unclear ones are assessed by an agent. Where a lot is at stake, a person decides.
A rule always gives the same answer. An agent doesn't.
A rule is simple: if a condition holds, take an action. If the return period since the order hasn't passed and the goods weren't on sale, the complaint is approved. Same input, always the same result.
An agent works differently. It gets a goal, materials and tools and works out the procedure itself. The language model chooses among several likely answers. It may approve the same request once and reject it another time. That isn't a bug, it is its nature.
A rule has three more advantages that are easy to forget. It can say exactly why it decided, which helps with audits. It can be tested. And it costs nothing extra, it doesn't call any model.
An agent makes sense where the logic can't be written in advance.
A rule handles only what someone foresaw. But a customer writes a long email, sends a photo of damaged goods or explains a situation nobody thought of. This is where an agent excels: it reads the text, looks at the photo, looks up the order history and proposes a decision.
Even with AI, the rule applies: what can be calculated should be calculated by a calculator. If you let AI add up a thousand bank transactions, the result may not be right. Let a program do the calculation and AI write a clear summary of it.
“An e-shop with complaints. A rule handles most of them immediately: within the period, no exception, approved. Complaints with free text and a photo are assessed by an agent, which proposes a solution. If the proposal exceeds a set amount, it goes to a person for approval. People deal only with what really needs judgment.”
A combination in the right order works best.
The recommended approach has three steps, and the order isn't random.
- Rules first. They are fast and cheap and handle most clear cases.
- Then the agent. It gets only what the rules can't decide: incomplete data, free text, images, unexpected situations.
- A check at the end. The agent's proposal is compared with fixed limits. Below the limit it applies immediately, above it a person approves it.
This way the agent doesn't get the last word where it's risky, and people don't work on things a simple condition could handle.
“A dental clinic and cancelled appointments. Cancellation more than a day in advance: no fee, a rule decides. A patient writes that they were in hospital and attaches a report: an agent assesses the text and proposes waiving the fee. The decision on an exception for a long-term patient with a larger amount stays with the head of reception.”
When not to use an agent.
If you can write the decision as a simple condition that changes once a year, an agent is unnecessary. If a mistake isn't acceptable, as with payments or access rights, an agent doesn't belong there. And if someone must personally answer for a decision, a person decides, and an agent can at most prepare the material.
What we see in practice holds: most AI projects don't fail because of a bad model, but because someone chose the wrong type of solution.
Ask yourself
Which decisions at your company do people make by hand, even though a simple rule could handle them?
