Many companies start their AI project with a tool. Somebody sees a demo, buys a subscription and connects it to the inbox. A few weeks later, the result is confusing. The AI does things, but nobody is sure they are the right things.
The problem is rarely the tool. The problem is that the company automated a process that nobody had described. This article explains why that makes things worse and which order works better.
The E-Myth idea
There is a well-known small-business book called The E-Myth. Its main idea is simple. Many small companies run on the owner's memory and energy. The owner works in the business every day, so nobody works on the business.
The book's answer is to turn the work into a system. Each repeated task has a written method. A new employee can follow it. The company does not stop when one person is on holiday.
This idea is decades old. It is more important now, because AI can only follow a system that exists. If the rules live in one person's head, the AI has nothing to follow. It guesses.
What happens when you automate a mess
A messy process has a few typical signs:
- Three people do the same task in three different ways.
- The price list exists in two versions, and nobody knows which one is current.
- Exceptions are handled "the way we always do it", but nobody can say what that is.
- Work waits because it is not clear who decides.
When you put AI on top of this, four things happen.
1. The AI copies the mess faster
AI learns from your examples: old emails, old quotes, old invoices. If those examples are inconsistent, the AI is inconsistent too. Now it produces inconsistent output at the speed of a computer.
2. Exceptions multiply
Every unclear rule becomes an instruction in the prompt. "If the customer is from Tartu, use the other price list. Unless it is a repeat customer." After a few months, the instructions are long and contradict each other. Nobody dares to change them.
3. Errors spread before anyone notices
A person who makes a mistake makes it once. An automated step makes the same mistake in every case until somebody notices. Without logs and checks, that can take weeks.
4. Nobody owns the result
If the process had no owner before, the automation has no owner either. When it fails, everyone points to the AI. The AI cannot explain itself in a meeting.
The order that works: map, simplify, automate, keep running
We use four steps, always in the same order.
Map
With your people, we follow how the work happens today. Not how it should happen, but how it really happens. Where does the input come from? Who touches it? Where does it wait? Where do errors appear?
The output is a process map. It is a simple drawing of steps, people and decisions. Most owners see something new in it. Often it is a step that nobody knew about, or a wait that takes longer than all the work together.
Simplify
Next, we remove steps you do not need. We write the rules and the exceptions in plain language. We choose one price list, one template and one owner for each step.
This step often saves the first hours without a single line of code. It also makes the next step cheaper. A simpler process is easier to automate and easier to test.
Automate
Now AI gets the repetitive part. It reads, sorts, extracts and drafts. A person approves every step that touches money, a customer or a decision. You decide what the AI may do alone.
Because the rules exist in writing, we can test them. For each rule, we prepare test cases: an input and the result we expect. If the AI gives a different result, we see it before your customer does.
Keep running
A workflow is not finished when it starts to work in production. Models change, the APIs of your software change and your people change. An API is the interface that lets two programs exchange data. We monitor the workflow, run the tests again after each change and update what needs updating. We explain this step in Models change every few months.
An example
For example, a 20-person wholesale company wants AI to answer stock and price questions from customers. During mapping, we see that two salespeople use different discount rules. One gives 5% to every repeat customer. The other gives 10% to customers above a certain annual volume.
If we automate now, the AI must pick one rule or invent a third one. Instead, the owner decides on one rule. The sales team agrees. We write the rule as one sentence. Only then does the AI draft replies with that rule.
The numbers in this example are illustrative. The pattern is very common. When you try to automate, you find the decisions the business never made.
Signs that your process is ready for AI
- You can describe the process in ten steps or fewer.
- You know who owns it.
- The input arrives in a predictable form: an email, a PDF, a web form.
- The rules fit on one page, with a short list of exceptions.
- You have examples of good results: sent quotes, posted invoices, answered emails.
Signs that it is not ready yet
- Every case is "different".
- The only person who understands the work is the owner.
- Nobody can say what a good result looks like.
- The process changes every month.
A process that is not ready is not a dead end. It means that the first value comes from mapping and simplifying, not from AI.
What AI does well after the process is clear
When the process is clear, AI is very useful. It reads documents that arrive in different formats. It sorts messages by type and urgency. It finds the right line in a long price list. It drafts text in your tone and your language. It does this at night, at weekends and during the holiday season.
But it does all of this inside the rules you gave it. The quality of the rules sets the quality of the result. That is why we say: process first, then AI.
What you can do this week
Pick one process that takes too much time. Ask the person who does it to describe the last five cases. Write each step on a sticky note. Then ask three questions:
- Which steps can we remove?
- Which rules can we put in writing?
- Which step repeats the most?
You will learn more from this hour than from most AI demos. If you want to know why demos often fail, read A demo is not a system.
Start with the process audit
Our process audit is this method in one week. We map one process with your people, simplify it on paper and choose what to automate first. You get a process map and 3–5 automation opportunities with time saved, risk and price. You also get a fixed-price offer for the first workflow. The audit costs €1,490 + VAT. If you continue within 30 days, we deduct it from the build.