Expert knowledge for digital decisions
How to Start an AI Project Correctly?
Short answer
The Sequence
1. Choose a use case. Frequent, time-consuming, errors easily correctable. Not the most impressive, but the most rewarding.
2. Measure the initial value. How many processes, how much time, how many errors – today. Without this number, nothing can be proven later.
3. Collect test cases. Fifty real cases with the correct answer. This is simultaneously the basis for every subsequent model evaluation.
4. Build a pilot. Small, with a test step, for a manageable group.
5. Let it run for four to eight weeks and measure.
6. Decide. Expand, adapt, or discontinue. Discontinuation is a legitimate outcome.
What Delays the Start Without Adding Value
- An AI strategy before the first use case
- Selecting a model before defining what it should do
- Comparing all providers
- Establishing a working group
What Causes Pilots to Fail
Too large a scope. A use case affecting five departments will never move beyond coordination.
No real data. A pilot with example data says nothing.
No initial value. At the end, people discuss impressions instead of numbers.
No one responsible. Without a person who accompanies it daily, it fades away.
The Timeframe
From the decision to the running pilot, only a few weeks should pass. If it takes months, the use case was too large.
Key facts
- First Step
- Use case, not strategy or model
- Mandatory Before Starting
- Measure the initial value
- Legitimate Outcome
- Discontinue the pilot
Sources
All external claims are backed by traceable sources.- 01