The Diversity of Intelligent Software Systems

In Shenzhen, taxis drive through the city without a driver, and the people there have grown used to it. In our companies, by contrast, much is still considered too important to be left to software. This post shows what the intelligent software systems of spring 2023 have in common technically, in which areas of a company they already prepare or make decisions, and which tasks remain with people.

What Shenzhen teaches about acceptance

Since 2020, robotaxis have been running in pilot operation in Shenzhen in southern China, often called "China's Silicon Valley." Since August 2022, a regulation of the special economic zone has also permitted trips on designated roads entirely without a safety driver and defines who is liable in the event of damage. The vehicles signal with green lights at the front and sides that they are operating in autonomous mode. Providers such as WeRide are expanding their operations in the region, as the South China Morning Post reports. Anyone driving through Shenzhen today experiences not a sensation but everyday life: traffic without distraction, without fatigue, without alcohol at the wheel.

For me, that is the real news. The technical feasibility of self-driving systems was foreseeable; what matters is that a society accepts them. In my book "Das selbstfahrende Unternehmen" (Springer Gabler 2021, in German), I described traceability and predictability of decisions as the most important requirements for intelligent algorithms. Beyond the technical prerequisites, emotional and socio-political requirements must also be met, otherwise these systems will not be accepted. The green light on the robotaxi is exactly that: a simple signal everyone understands, and it creates predictability.

Once these conditions are met, habituation is quick. With the smartphone it took a few years; with self-driving traffic in Shenzhen apparently just as long. For companies, a simple consequence follows: everything that can be self-driving will sooner or later be self-driving, controlled by learning, reliable systems deployed where they are superior to people and create real value.

Computing power, neural networks, and reinforcement learning

Behind all of these systems lies the same technical foundation. Deep learning methods recognize patterns in large volumes of data, and for that they need considerable computing power. The development of that power has followed exponential curves for decades: a current smartphone delivers what required an entire data center only a few years ago. For the training of the largest AI models, OpenAI calculated that the compute used doubled roughly every 3.4 months between 2012 and 2018 (Amodei and Hernandez 2018, "AI and Compute"). Nothing suggests that this curve will flatten in the coming years.

Three learning methods shape today's systems:

• Supervised learning. A neural network learns from labeled examples, for instance from millions of images, which features make up a traffic sign or a defective workpiece. The image recognition of self-driving cars is based on this; with every trip the data set grows, and the error rate falls.
• Unsupervised learning. The system finds structures in data without anyone providing the right answer: customer segments, anomalies in machine behavior, topics in large bodies of text.
• Reinforcement learning. The system tries out actions and is rewarded when the outcome is good. This is how DeepMind's AlphaGo beat the world-class player Lee Sedol in 2016, and this is how the language models behind ChatGPT are refined with human feedback (Ouyang et al. 2022, arXiv:2203.02155).

As impressive as the progress is, in spring 2023 we are far away from a "general intelligence" that comes close to the breadth of human thinking. Almost all intelligent systems are specialized in one task or one field of application; they cannot be deployed as flexibly as human intelligence. From this follows an architectural decision that I justified in Cognitive software is more than artificial intelligence: the intelligent company does not emerge from one super-algorithm but from the networking of many subsystems. Each of them combines rule-based programming with learning methods; the rules provide traceability, the learning provides adaptation. How such systems arrive at their decisions is described in How self-driving companies decide.

For classification within a company, the purpose of use helps more than the technical method. In the book I proposed the RAPD framework for this: recognition evaluates historical data, assistance prepares decisions, prediction extrapolates trends, and detection decides in real time, for example in anomaly detection on the conveyor belt or in autonomous driving. The following fields of application show how these four types are distributed in everyday business.

Four fields of application in the company

Market analysis: recognizing patterns, calculating scenarios

Intelligent software collects and processes large volumes of data from market and competitor observation in order to recognize trends and patterns. Posts on social media can be evaluated continuously to understand the opinions and needs of customers; the activities of competitors are monitored and provide the basis for the company's own differentiation. Scenario models turn this into forecasts of future developments on which campaigns and product adjustments build.

Customer relationships: individual instead of segmented

In customer relationship management, algorithms predict on the basis of earlier interactions and purchases which products or services are relevant for individual customers. Chatbots answer inquiries automatically, increasingly with speech input and output as well; the language models of 2023 make such dialogues attainable even for smaller companies that could not afford their own development until now. Reviews and complaints are analyzed to derive improvements. Leads are evaluated and prioritized according to their potential, within the limits of data protection and on the basis of the behavioral patterns that can be captured online.

Production: continuous optimization instead of periodic inspection

In production, systems monitor the performance of machines and plants and predict when and how maintenance must be carried out (predictive maintenance). Products are checked automatically for defects and deviations, and quality improves continuously. Decisions made in real time shorten lead times; energy consumption and material requirements are optimized continuously. Detection dominates here: the system decides while the plant is running, not in the next shift.

Sales: evaluating opportunities, steering prices

In sales, systems forecast sales opportunities and revenue potential from historical data and adapt the sales process accordingly. Personalized messages are generated automatically from the analyzed interests and preferences. Pricing strategies are developed from purchase history and buying behavior and ideally aligned with the contribution margin accounting from controlling, so that profit rather than revenue is optimized.

The common denominator: data and an understanding of the system

As different as the fields of application are, the prerequisites are the same everywhere. First, the company needs a data strategy: which data is created where, who owns it, in what quality is it available, and how does it reach the systems that process it? Second, it needs an understanding of how its own internal and external workflows interact. Only then can the subsystems access the right data from one another. A pricing algorithm without contribution margins and a maintenance model without production planning each optimize past the whole.

These foundations are the real effort. Building the systems initially costs a great deal of investment budget and work. Individual use cases, such as anomaly detection on an expensive plant, pay off quickly, but the platform behind them only over years. Based on my consulting experience with large companies, I recommended in the book a calculation period of around 15 years for business-critical software systems; a return on investment for such systems is not to be expected in under five to ten years. Anyone who treats intelligent software like a quarterly project will be disappointed; anyone who plans it like an investment in machinery or real estate has the right yardstick.

People remain at the center

In the book I named the threshold: if around 80 percent of operational functions and decisions are automated and controlled by sufficiently intelligent algorithms, a self-driving company exists. The remaining 20 percent are not a residual but the more demanding part of the work:

• Decisions in complex and unexpected situations for which there is no training data.
• Collaboration and communication with major customers, with partners, and in innovative projects.
• Strategy and innovation, that is, the guidelines within which the systems optimize in the first place.
• Supervision and control of the AI systems: people make sure that the systems follow ethical principles, do not discriminate, and can justify their decisions.

Behind this stand two of the seven central theses from the book: people will continue to work in and for companies and fulfill the empathetic and creative tasks; intelligent software algorithms are better suited to repetitive tasks than people. Creativity, empathy, judgment, and strategic thinking remain indispensable in order to cope with ever faster change and to find new forms of value creation. And the last of the seven theses applies in Shenzhen just as it does in Linz: companies will still serve people in 2035.

The next step

You will find the vision of the self-driving company and the levels of autonomy on the way there on the page The self-driving company. If you want to know which of the four fields of application has the greatest leverage in your company and which data strategy it requires, book an expert meeting.