Artificial intelligence automation

What is artificial intelligence automation and where to start?

Adding AI to a process doesn't make anything better by itself. First, we need to know which task really takes the team's time and what its correct output looks like. In this article, I will deal with this choice step by step.

Reza Doostipour
Reza DoostipourExpert in artificial intelligence automation and business intelligence
9 minutes read

When do we need artificial intelligence at all?

If the law of work is clear, the same automation is better and more reliable. For example, when the payment is confirmed, the order status changes. There is no need for artificial intelligence to enter the story.

Artificial intelligence is useful when the input of text, image or information is irregular; For example, it is necessary to recognize the subject of a message or to extract some specific data from a document. Usually, the best result is obtained by combining artificial intelligence with simple and clear rules.

What should we look for to begin with?

A good start is usually repetitive, takes time, and we have real examples of that. Most importantly, we can clearly tell the difference between acceptable output and wrong output.

If a mistake could have a serious financial, legal, or human consequence, the result of AI should not be implemented automatically. In such a situation, it is better for the system to make a suggestion and someone to confirm it.

An attractive example is different from a usable system

In a demo, the input is clean and everything works fine. In real work, incomplete information arrives, services are interrupted, and some results are ambiguous. The system must have a specific response for all these modes.

Logging events, retrying at the right time, controlling cost and referring suspicious cases to humans are not extra details. These determine whether the solution can be used tomorrow or not.

How do we know if the output is really good?

Before daily use, we collect a number of real and different samples and write the desired result for each one. Then we try different models or commands on the same samples so that the comparison is fair.

It is not just the accuracy of the model that matters. We need to see how much time the team has saved, how many outputs still need to be modified and how many times the whole process has stopped with errors.

Is it worth implementing this idea?

Before you start, write down how many times you do the task and how long it takes each time. Then factor in the cost of the model, infrastructure, maintenance, and time still needed for human review, aside from the real savings.

Sometimes an idea is technically attractive, but because it is not used more than a few times a month, it has no economic value. On the other hand, saving a few minutes on a task that is repeated hundreds of times every day can be very important.

A simple and low-risk start

Choose a specific task, set aside a few dozen real examples and write down what results are not acceptable. Then build a limited prototype and let someone who actually does the work work with it.

If the sample is useful in practice, increase the scope. This path may be less exciting than starting a big project, but the probability of achieving a real result is much higher.

Picture of Reza Dostipour
About the author

Reza Doostipour

I started my career in software development and later gained experience in network, operations management and enterprise sales. Today, I use the combination of these experiences to simplify work processes.

View introduction and records
Let's start here

Have a time-consuming task that you think could be done better?

It is enough to write briefly how the work is done today and which part is annoying or time consuming.

Write a message