When does an AI agent make sense?
We build AI agents and we regularly recommend against them. Not out of modesty, but because an agent in the wrong place is expensive, disappointing and ruins your confidence in automation for years. Here's how we rate it: so you can.
THE CORE QUESTION
Repetition without judgment: that is the domain of an agent
An agent is strong in work that occurs frequently, has a predictable form and rarely requires real consideration. As soon as judgment, relationship or exception takes center stage, a human wins: and you lose money to automation that makes correcting more expensive than doing.
GREEN
Many, similar, factual
The same question about delivery time, status or opening hours a hundred times a month. The answer is somewhere in a system and is not open to discussion. This is where an agent earns his return the fastest.
ORANGE
Many, but with exceptions
Quotations and requests: the form is predictable, the content not always. Works well when the agent prepares and approves a human: not when sending independently.
RED
Little, or full of judgment
Complaints, negotiations, sensitive files, or simply ten questions per week. Automation here costs more than it yields, and the error is more expensive than the profit.
FIVE SIGNALS BEFORE
Then an agent is probably a good idea
Do you recognize three or more? Then there is a good chance that an agent will pay for himself within six months.
- The same questions, every week. You could recite the answer from memory: and you do.
- Outside office hours it remains stationary. Customers ask questions in the evening and don't hear anything until the next morning.
- The answer is already there somewhere. In your order system, your terms and conditions or your price list: only someone has to look it up.
- Retyping between systems. The same data moves from one screen to another by hand.
- Reaction speed determines whether you get the job. Whoever offers first wins: and that's not always you.
- Volume is there — at least several dozen similar actions per week.
- The source is digital — the data is in a system, not in someone's head.
- The rules can be written down — you can explain when which answer is correct.
- There is a human available — someone who picks up exceptions and watches.
- You dare to measure — you know approximately how much time the process takes now.
FIVE SIGNALS AGAINST
Then we don't recommend it: even if you ask for it
In these cases we say no. Sometimes with an alternative that costs a fraction, sometimes with the advice to do nothing at all.
- Too little volume. Construction does not pay for itself without approximately twenty similar actions per week.
- Every case is different. If there is no pattern, an agent cannot learn one.
- The knowledge is not fixed anywhere. If everything is in someone's head, write it down first - that often yields the most.
- An error is unacceptable. Medical, legal or financial advice does not belong to an agent.
- The process itself is flawed. Automating a bad process makes it wrong faster, not better.
- Alternative: a good FAQ page often captures half of the repeat questions.
- Alternative: one connection between two systems solves retyping without AI.
- Alternative: a booking system retrieves appointments from the mailbox.
- Alternative: automatic emails for status changes save a lot of "how are you doing" questions.
- Alternative: first write down the process and clean it up: sometimes that makes the problem go away.
THE arithmetical test
Calculate it yourself, in four lines
You don't need a consultancy to determine whether it will work out. Enter your own figures and compare the outcome with the construction price.
| Rule | What you fill in | Example |
|---|---|---|
| 1. Number of actions per week | how often does it happen | 120 customer questions |
| 2. Minutes per action | including lookup and switching | 4 minutes |
| 3. Share that an agent can handle | realistic: 60–80% | 70% |
| 4. Hourly costs of the employee | gross, including employer charges | €35 |
| Result per month | (1 × 2 × 3 ÷ 60) × 4.3 × 4 | ≈ €840 |
CALCULATION EXAMPLE WITH INDICATIVE NUMBERS: ENTER YOUR OWN NUMBERS FOR A REAL OUTCOME
FREQUENTLY ASKED QUESTIONS
Doubtful cases
I don't have much volume, but I do have a lot of evening questions. Does it make sense then?
Sometimes yes. Accessibility has a value that is separate from hours saved: a customer who receives an answer in the evening will not call your competitor tomorrow. Then don't count on saved time but on lost turnover: if one extra assignment per month covers the costs, it can be done.
Can't I just use a standard chatbot?
For simple FAQ questions: fine, and cheaper. As soon as the answer has to come from your order system or CRM, a standard chatbot crashes: that's exactly the point where customers drop off.
What if my employees are resistant to it?
Justified concern, and usually the real risk. That's why we first let the agent shadow with approval by a human. In practice, resistance changes when people notice that the boring part in particular disappears: exactly what happens in our lives portal case happened.
You make money from agents. Why would you advise against one?
Because an agent who produces nothing is switched off after six months: and then we lose a customer and a reference. An honest no costs us one assignment; a bad project costs us more.
Are you unsure whether it applies to you? comes true?
Tell us which process you have in mind. We calculate it and honestly say whether it is worth it.