In short: Most decisions in a business process can be made with simple rules. Rules cost nothing per call, they are explainable, and they can be tested. AI is used where rules cannot get any further, and its result is stored so that the same question never costs money twice.
The reflex in reverse
Anyone starting an automation project today often begins with the model. Which model do we take, which prompt, how do we connect it. The question of whether this decision needs an AI at all rarely comes up.
In the sales tool I am currently building with a partner, one rule therefore sits at the very top of the architecture: AI is the fallback, never the first choice.
First we look in our own catalogs, in the cache and in fixed rules. Only when that finds nothing do we ask a language model. And whatever the model answers, we store. The next identical question is answered by the store again.
Three reasons that have nothing to do with technology
That sounds like a developer decision. For you as a managing director, there are three very tangible reasons.
Cost. A rule costs development time once and nothing after that. An AI query costs money on every call. Per call that is little, but automation means: a great many calls. A process that runs a thousand times a day turns cent amounts into a monthly bill.
Explainability. When a rule says “This customer has not been contacted for 6 months, which is why he is on the list”, every employee understands it. When an AI says “relevance 7 out of 10”, everyone rightly asks: Why 7? A value nobody can explain gets ignored. Or worse, believed blindly.
Testability. A rule always does the same thing with the same input. You can test it once and rely on it. An AI can answer the same input differently today than it did yesterday. That is not wrong, but it has to be monitored permanently. Every AI step in the process is a step where you have to measure quality.
An example: Which customer is up today?
The sales tool has a feature for existing customers. The requirement was: It should move in the direction of a CRM, but the user should have no work with it. The system decides when a customer becomes relevant again and puts at most one card per day on the “Today” pile.
The obvious solution would be: An AI reads all customers and decides. That would be expensive, hard to explain, and one call per customer per day.
We split it up.
Rules decide whether a customer qualifies. No reply for a long time. Contract ending in 90 or 30 days. Never asked for a referral. Seasonal business coming up. These are clear criteria that any salesperson understands immediately.
The AI picks the one from the candidates and writes the sentence to go with it. Why today of all days, what a good opener for the conversation would be. That is one call per user per day, without personal data.
The rules do the heavy, frequent work. The AI does what rules are bad at: weighing and phrasing.
A second example: Understanding wishes, filtering honestly
When searching for companies, the user can write in plain language what they are looking for. For example: “Small trade businesses in the south of Berlin that do roofs.”
AI makes sense here because it can translate language into search criteria. But it may only translate, not decide. The user sees the result of the translation as individual building blocks before the search runs. The AI suggests, the user sees it, nothing happens unasked.
And it must not promise anything the data cannot deliver. None of the sources knows how many employees a business has. So “small” does not become a filter but a sort order, and one line says openly: “Not filterable.” If the AI promises “small” and then delivers a business with 80 people, trust is gone.
Where AI really belongs
That does not mean AI is superfluous. It means it should sit in the right places. In my experience, these are mainly three kinds of tasks:
- Reading unstructured material. A legal notice, an email, a PDF from the call center. Turning it into fields that a rule then processes further. AI reads, the rule decides.
- Phrasing. A draft reply, a summary, a conversation opener. Here text is the purpose, so AI is the right first choice.
- Mapping where there would be too many rules. For example, mapping the columns of someone else’s Excel list to your own fields. The result is shown before saving, and the whole operation can be undone.
Incidentally, we do exactly the same when matching emails to customers: first by sender address, then by domain, and only at the third stage does an AI get asked. Most emails are done after stage one.
What this means for you
When someone proposes an AI automation to you, ask for every point where an AI decides:
- Could this be done with a rule too? If an employee can explain the decision in one sentence, it is usually a rule.
- What does one call cost, and how often does it happen? Scale it up to a month.
- Is the result stored? If the same question is paid for anew every time, that is money nobody is spending consciously.
- Can your team explain why the system decided that way? If not, they will not trust the system for long.
Good automation often looks unspectacular from the outside. Many rules, little AI, and exactly where it adds something.
If you want to automate a process and want to know where AI pays off and where a simple rule is enough, that is exactly what we talk about in a first call.
