Intelligence, Ambiguity, Responsibility

Intelligence: "the ability to understand and learn well, and to form judgments and opinions based on reason" Cambridge Dictionary

When I started working with SAP software in 2006, I learned about a module called Business Intelligence. At first, I was fascinated by the term. Later, I was disappointed. There did not seem to be much “intelligence” in it.

To me, BI was mostly a large data collector. The actual intelligence was still provided by humans: by consultants who programmed extraction logic, by business users who interpreted reports, and by managers who turned spreadsheets and presentations into decisions.

In the years that followed, the word “intelligence” in IT evolved. Forecasting, predictive analytics, Bayesian models, demand planning, sales predictions. Especially in retail, we used more and more sophisticated tools to look at the past and estimate the future.

But any decision maker knows: data is ambiguous.

In fact, almost anything relevant in business and in life is ambiguous.

Yes, the electricity is either on or off. The internet connection works or it does not. But already with something as simple as an “empty fridge,” ambiguity begins. It may not be empty. But whether anything inside is still edible is a different question. The color of the yoghurt will not reveal itself until it is opened.

So what is intelligent about AI?

Technically, in the case of large language models, it is prediction. It predicts words, structures, likely meanings, likely answers. By that definition, one could argue: not intelligent.

But in practice, it does something that, until recently, only humans could do.

With limited context, it derives probable meaning. It interprets intent. It works with incomplete information. It gives an answer without first forcing us through confidence intervals, degrees of freedom, and every possible alternative. And sometimes, when the ambiguity is too high, it asks for clarification in a surprisingly human-like way.

AI deals with the ambiguity of human language: our incomplete sentences, our missing context, our typing errors, our hidden assumptions, our mysterious tendency to omit exactly the information that would have made the question easy, because we believe it to be self-evident.

It deals with the ambiguity of the artifacts we throw at it.

Of course, it can reach dubious conclusions. But it does reach conclusions. And those conclusions can be challenged.

Unfortunately, an LLM can be manipulated by giving it selected context, or unwillingly by omitting context. But that is not unique to AI. Humans can be manipulated in the same way. Give people selective information for long enough, and many will eventually treat it as reality.

The question is therefore not whether AI is fallible.

The question is: who should be allowed to make decisions when the context is incomplete, ambiguous, biased, or even manipulated, and the stakes are high?

For human decisions, companies and societies have built systems of checks and balances. An employee asks a manager. A manager discusses with a board. People and companies are bound by contracts, laws, regulations, audits, and courts. If a human decision maker fails badly enough, there are consequences: loss of position, loss of trust, financial damage, legal action.

An AI does not care about consequences.

It may be updated. It may be switched off. But it will not seriously care.

This matters deeply in business software.

Take a simple retail example: a customer wants to return an item without a receipt.

A strict rule could be: always accept it. Another strict rule could be: never accept it. Both are bad.

If a customer walks out of the store, throws away the receipt, comes back two minutes later and says, “I picked the wrong size,” most salespeople will accept the return. It is reasonable. It protects the customer relationship.

But if someone comes in three times a month trying to return goods without receipts, and these goods have gone missing from the store, the situation changes.

The rule is not enough. The context matters.

This is exactly the kind of ambiguity AI is good at processing. But it is also exactly the kind of decision where we should be careful about handing over authority without supervision. An AI will need guardrails for its decisions. But once the guardrails become so well-defined that we can trust the AI with it, we might not need the AI anymore, but a deterministic system will do.

At acceletail, this is one of our core beliefs:

AI will be extremely powerful with ambiguity. It is already changing the way people work today. In process automations, IT projects, in support, in monitoring, in maintenance, in custom code, in master data, in process exceptions.

The art of productively employing AI will be to employ it where ambiguity cannot sensibly be avoided and applying human governance corresponding to the stakes of the decision.

In many cases, the wise decision will be to use AI to disambiguate the process and implement a deterministic and testable process instead.

In many others, it will be to use AI to clarify context, surface patterns, formulate hypotheses, and help experts reach better decisions faster.

In the following chapters we will look at various situations in the IT life cycle and in business processes as to where and how AI will be sensibly deployed in the future, covering IT projects, IT maintenance and support as well as retail processes.

All texts in this series were written in English by us personally. Translations were done by AI and checked manually. AI was used as an editorial assistant: to improve style, structure and, hopefully, readability. Pictures are AI-generated.