Cognitive Software Is More Than Artificial Intelligence

Since ChatGPT became available to everyone at the end of November 2022, two questions keep coming up in our conversations with boards and IT leaders: Is this the artificial intelligence that will run companies in the future? And if not, what is still missing? In my book "Das selbstfahrende Unternehmen" (Springer Gabler 2021), I use a term of my own for the software that will make most of the decisions in the company of the year 2035: cognitive software. This article explains what distinguishes it from the artificial intelligence available today, what ChatGPT already anticipates, and the two requirements every decision-making software must meet before a company hands decisions over to it.

Artificial intelligence: an old term without a sharp boundary

The term artificial intelligence is as old as computer science itself and still has no exact definition. In 1950, Alan Turing proposed a pragmatic criterion: artificial intelligence is present when I can no longer tell whether I am dealing with a human or a machine. Joseph Weizenbaum showed how fragile this criterion is as early as 1966, with a program that merely rephrased questions and played them back, yet still prompted genuine emotions in some of its users.

For practical purposes, a different distinction is more helpful. Classical software consists, at its core, of a very large number of if-then decisions that people have defined in advance. An algorithm is a recipe: preheat the oven, weigh the flour, bake for a certain time. Learning algorithms, by contrast, change their decisions based on the data they see. The simplest example is the learning spam filter I programmed myself years ago: for every word in an email, it records how often that word appeared in messages that users had marked as unwanted, and from this it derives a spam probability for each new email.

Almost everything that runs under the label of artificial intelligence today is so-called weak AI or narrow AI: a learning algorithm is trained for a single task, such as recognizing speech, classifying images, or predicting machine failures. What it learns cannot be transferred to the next task. Yet this transfer is exactly the rule in everyday business: whoever solves a customer problem draws on knowledge about products, contracts, suppliers, and people. The long-term goal of research is therefore a general intelligence that imitates complex human thought processes and transfers insights between tasks.

What ChatGPT shows and what it does not show

With ChatGPT, an AI system meets the general public not as a feature inside an app but as a conversation partner that answers questions on almost any topic. According to a UBS analysis cited by Reuters in February 2023, it reached around 100 million users within two months. Yesterday, OpenAI already presented the successor model, GPT-4, which according to the company also understands images as input and performs significantly better than its predecessor in professional exams.

What is remarkable about ChatGPT is not any single capability but its breadth: it answers follow-up and detailed questions in almost every field of knowledge, corrects itself when pointed to a mistake, challenges false premises, and declines inappropriate requests. In doing so, it steps out of the narrow problem space of narrow AI and shows, in some areas, exactly the kind of contextual common sense we need for cognitive software.

The second observation is just as important. In the book, I described that artificial intelligence will at its core be based on individual algorithms, but that its effect will emerge from the networking of subsystems, just as voice input on a smartphone only became usable through the interplay of phonetic decoding, interpretation of meaning, and data center. ChatGPT confirms this: a large language model, training on dialogues, human ratings of the answers, and a simple interface together produce something that none of the components delivers on its own.

What ChatGPT does not show is equally clear. It knows neither the contracts nor the bills of materials nor the customer history of a company. It has no access to the company's systems. When it does not know an answer, it invents sources and figures that sound plausible, and it cannot explain how it arrived at a statement. As a decision-making system for a company, it is therefore unsuitable; as a building block of such a system, it is impressive.

What distinguishes cognitive software from artificial intelligence

Cognitive software, as I use the term, is the combination of learning algorithms, the company's rules and data, and the interfaces to its systems that makes decisions autonomously, proactively, and traceably. The difference from today's typical artificial intelligence lies in four points:

• Context instead of a single task. Cognitive software imitates intellectual abilities flexibly and applies them across tasks, rather than recognizing one single pattern.
• Dynamics instead of a snapshot. It handles challenges in dynamic systems, adapts to changing conditions, and remains capable of acting when disruptions occur.
• Company data instead of world knowledge. Its raw material is the company's own data, supplemented by external sources such as weather data, market prices, or credit information.
• Justification instead of a black box. Every decision is traceable and its effect predictable; more on this below.

The computing power for this has long been available; for individual areas of a company, cognitive software can be built today. Its development, however, does not begin with the algorithm but with an analysis of how people in the respective area actually decide: which information they draw on, which rules they apply, where they exercise discretion. Only this analysis turns a model into software that a company can entrust with decisions.

Four fields of application: recognition, assistance, prediction, detection

In the literature, learning algorithms are usually classified by method or learning behavior: symbolic AI, neural networks, simulation. For decision-makers, a classification by field of application is more useful, because use cases in the same field share similar data, risks, and integration requirements. Out of our consulting practice, the RAPD framework emerged for this purpose, which I describe in the book:

ChatGPT falls into the field of assistance and shows how quickly this field is currently maturing. The other three fields are already in productive use in many companies, but mostly as isolated solutions of individual departments. Cognitive software emerges when the four fields are connected with the company's data and with each other: it continuously analyzes large volumes of data, recognizes patterns in the behavior of customers and markets, interprets correlations, evaluates opportunities and risks based on data, creates individualized offers, accelerates decisions in acute situations, and meets compliance requirements without exception.

Two requirements: traceability and predictability

In the book, I justified the most important requirements for cognitive software not technically but socially: decisions that no one can trace will not be accepted by those affected, neither by employees nor by customers, neither by supervisory authorities nor by courts. Traceability means that for every decision, it can be stated which data and rules it is based on. Predictability means that it can be said in advance how the software will decide in a given situation. The draft AI regulation presented by the European Commission in April 2021 aims at the same properties, and we expect them to become the benchmark for every deployment in a company.

For dealing with systems like ChatGPT, a clear rule follows: they belong where a human checks the result before it takes effect, that is, in the preparation of texts, research, and decision proposals. In a credit decision, a claims settlement, or a price approval, they have no place as a black box. There, cognitive software is needed whose rules are explicit and whose learning components are embedded in such a way that their contribution can be logged and verified. Why secure software means secure decisions and how self-driving companies decide are topics we have already covered in this series.

Why people decide worse than they think

The objection is obvious: why should software decide better than experienced executives? The answer comes from behavioral economics, which, since the work of Daniel Kahneman and Amos Tversky, has systematically documented how irrationally people decide under information overload, time pressure, and stress, in other words, under the conditions of normal office life. Six biases are particularly consequential for business decisions:

• Information overload. When there is too much information, people draw on only a few pieces because of limited cognitive resources, and often the less relevant ones.
• Stereotypes. New information is assigned to familiar patterns even when it does not fit; because this happens unconsciously, it is rarely noticed.
• Anchoring effect. The first assessment shapes all subsequent ones, even if it turns out to be wrong.
• Halo effect. A single, particularly striking aspect outshines all others and distorts the overall judgment.
• Risky shift. In groups, the striving for consensus and harmony dominates; risks are systematically ignored, and decisions turn out riskier than individual decisions based on the same data.
• Mental accounting. Someone who loses an opera ticket worth 100 euros usually does not buy a new one; someone who loses 100 euros in cash on the way buys the ticket anyway. The loss is identical, the decision is not, because we book expenses in separate accounts.

Added to this are the motives that Steven Reiss (2004) described in his theory of the 16 basic desires, such as power, acceptance, status, and vengeance, which shape countless decisions in hierarchical organizations without ever appearing in a decision memo.

Cognitive software is cognitive in a precise sense. People often only believe they are acting cognitively and rationally.

Cognitive software does not know these biases. It has no prejudices, holds no grudges, does not tire, and processes the complete data situation rather than an excerpt. This does not mean it is free of errors: its errors come from the data it was trained on and from rules that someone formulated incorrectly. That is precisely why traceability and predictability are not optional extras but the condition under which software may replace human decisions.

What companies lose by waiting

Companies that cling to analog decision paths are caught, as data volumes grow and change accelerates, in a gap that widens from year to year:

• They become slower and more error-prone in analyzing and evaluating opportunities and risks.
• They can only handle a limited density of information and leave signals unused that have long been present in their data.
• The manual effort of collecting information and transporting it to decision-makers grows faster than the organization.
• On every path through the hierarchy, information can be altered or suppressed by errors, emotions, and self-interest.
• The time lag of human transmission becomes critical, for example with individualized offers or in acute crises.

The shortage of skilled workers intensifies this development, because exactly those people are missing who have so far carried out the manual transport of information. Cognitive software is therefore not a project for the year 2035 but the lever with which companies will increase their value creation and reduce their coordination costs in the coming years. The first step in that direction is due now.

The next step

The first step is not a technology decision but an inventory: Which decisions does the company make every day, which data are they based on, and in which of the four fields of application could the first cognitive software solution with traceable rules be built? My book "Das selbstfahrende Unternehmen" describes the foundations: About the book. In an expert meeting, we will jointly identify the area in which your company should begin with cognitive software.