What artificial intelligence can and can´t do

For years, artificial intelligence has fluctuated between the terms “Terminator”, “technology of the future”, or “new superpower”, and “doesn’t work”, “is disappointing”, or “no confidence”.

It is not surprising that there have repeatedly been so-called “AI summers” and “AI winters” over 70 years. More specifically, there have been two AI winters since the 1950s, when AI was being researched and developed: In the early 1970s and late 1980s, the high expectations for the technology were dashed. Investors pulled out, and researchers got less funding.

What changed? The technology got better, we have powerful servers with lots of storage, cloud computing, and there is Big Data. A new opportunity for the potential of artificial intelligence! Yes, we are currently living in an AI summer. AI is again a superpower, the hope of the future.

You can already tell: the expectations are enormous. What is needed for us to use AI successfully? Realistic expectations! How do realistic expectations come? By having enough information.

You will learn what AI can and cannot do in the following article.

What artificial intelligence can do

We already encounter it constantly in everyday life: Amazon or Netflix suggestions according to one’s interests, navigation systems, Alexa & Co. that understand speech and much more.

All these systems fall under the category of “weak AI”, or a bit nicer in English, “narrow AI”. This definition refers to systems that solve a very specific problem. AI applications are very specialized. That means they can do one thing very well, quickly, and efficiently.

For example, our predictive sales software can accurately forecast customers’ future buying behaviour quickly and efficiently from hundreds of thousands of sales data. However, the software is not capable of composing texts or understanding speech. That would require a different kind of AI algorithm.

Nor can AI systems apply general goals to different situations. For example, a goal-oriented instruction such as “make sure you don’t hurt yourself or anyone else” is not sufficient for autonomous driving. No, every conceivable traffic situation, in different variants, must be recreated for the machine so that it knows how to react.

Artificial intelligence can therefore always be put to good use where there are data sets in which the human can first show the machine how to do it. This is also called “labelling.” The machine then gradually learns from this “pretending, or labelling” and can then apply what it has learned to new data.

How well does that work today?

These are realistic expectations of an AI

In special applications, AI systems already work extremely well today – just think of ChatGPT, it is amazing how far AI systems have come in the last year. In a business context, AI systems are mainly used for routine tasks.

But artificial intelligence has its limits. Even before the introduction of ChatGPT, short stock market reports were already being partially written and published by artificial intelligence—texts with a high degree of routine and clear structure. Since then, large language models like ChatGPT have evolved rapidly. In terms of language capabilities, the boundaries are shifting almost daily.

In the beginning, ChatGPT and similar models struggled with complex chains of reasoning and multifaceted questions. Their sense of humor also left much to be desired. Today’s models have made significant progress in these areas—they can argue more coherently, express ideas more subtly, and even write with a sharper sense of humor.

Still, even the most advanced models have their limitations. They can produce factual inaccuracies, imprecise source references, or entirely fabricated information—commonly referred to as “hallucinations.” LLMs also face challenges when dealing with structured data such as tables or databases—they are optimized for text, not for exact figures or analytical precision. Additionally, maintaining logical consistency across longer texts remains a challenge, often leading to contradictions or breaks in coherence.

To keep your expectations of an AI system realistic, you should inform yourself about it beforehand. Here are our top 3 questions you should ask:

1. From what data does AI learn from?

An AI can’t know anything it doesn’t have as a basis. If an image recognition program has only seen dogs and cats, it won’t recognize a tractor.

2. How well is the AI already trained, or what is its hit rate?

No system will always be 100% correct. That is simply the nature of machine learning (a subfield of AI).

In principle, machines are constantly making predictions through mathematics and statistics. This includes predictions about whether a product might appeal, predictions about whether there is a cat in the picture, and predictions about what is the fastest route.

Machines calculate probabilities. The more data and training the systems have, the better they become.
Professional AI systems produce up to 98% accurate results.

3. For what purpose should AI be used?

As I said, AI systems are “expert idiots”. You will not be able to use a single powerful AI tool that coordinates and manages an entire company, for example. Too many different capabilities are needed for that. Today’s AI systems are usually helpers for routine tasks or analysis.

This also applies to AI-based predictive sales analytics software. The software makes reliable sales predictions, such as which customer will soon churn, which price is most appropriate for which product, or where there is cross-selling potential. However, the sales team still has the rudder in their hands and decide for themselves whether and which sales actions follow thereafter.

CALCULATE NOW THE ROI OF QYMATIX PREDICTIVE SALES SOFTWARE

What can artificial intelligence (AI) do and what can’t it do? – Conclusion

Today, there are no “Terminator-like” AI systems. Most AI applications have tremendous utility as support tools in everyday work: namely, for routine yet time-consuming tasks.

Our tip is to learn about the rough workings of the various systems and ask the right questions. Only then can you also trust an AI-based software in everyday work and use it correctly.

Overall, we can fully agree with the statement of Dr. Martin Klarmann: “Modern technologies, such as artificial intelligence, are overestimated in the short term and underestimated in the long term.”

I WANT PREDICTIVE ANALYTICS FOR B2B SALES.
Further Read:
“Was kann KI”: Interview mit dem Physiker und Neurobiologen Christoph von der Malsburg (2019) German language