The End of Processes, Long Live the Algorithms

For decades, companies have organized their work in end-to-end processes: from order intake to invoice, from job application to employment contract. Digitalization adopted this structure and merely rebuilt the chains in software. I consider that an intermediate step. Once self-learning algorithms communicate with each other across departments and in real time, the process chains dissolve and are replaced by a networked overall system. This article explains why that is the case, how the transition unfolds, and what it means for decisions today.

The process as an organizing principle has its limits

An end-to-end process has a beginning, an end, and in between a sequence of steps that runs through several departments. For a long time this form made sense: it made work plannable, divisible, and auditable. At the same time, it shaped the organization. Departments emerged along the process steps, boundaries along the departments, and at every boundary an interface where information is handed over, reformatted, and not infrequently lost.

Thinking was constrained along with the structures. Anyone who has worked for a long time in a larger, hierarchically organized company knows the pattern: over time, every department develops its own system logic and optimizes for itself first. Information is withheld, metrics are polished, errors are reported only when they can no longer be hidden. The daily fight for a small advantage within one's own boundaries crowds out the overall optimum. In "Das selbstfahrende Unternehmen" (Springer Gabler 2021, in German), I described this behavior as a typical trait of the analog company.

The digitalization of the past twenty years has changed little about this. It transferred the process chains into workflow systems, accelerated the handovers, and made the individual core processes a few percentage points more efficient. The basic structure remained: linear, compartmentalized, with a beginning and an end.

The comparison with the automobile

The automotive industry shows how an organizing principle can shift. For roughly a hundred years, the car was conceived and built around the combustion engine. This centerpiece with its hundreds of components was the focus of research and development; power, torque, consumption, and emissions were the decisive metrics.

With the electric drivetrain, a battery and a comparatively simple motor are essentially enough. Attention shifts as a result: away from the mechanics, toward the people on board, their comfort and safety, and toward the software that will steer the vehicle in the future. The manufacturers have to change their thinking, not just their technology.

The same applies to companies. Until now, attention has gone to the core processes of individual departments, from which a few percent of efficiency were squeezed out. The far greater potential, namely networking and steering the company as a whole, was lost from view. Relieving people of routine work through software brings that view back.

From process to networked system: three stages

Processes are not replaced in a single step, but in the same sequence in which a company evolves into a self-driving company. I distinguish three stages.

The first stage is often underestimated. An incoming invoice that is scanned at reception and forwarded as an image by email is stored digitally, but it is as opaque to software as the paper before it; the workflow stays the same. Only when the content and purpose of the data are understood by machines can the second stage begin.

The third stage is the actual break. The former processes then no longer exist as chains with laborious interface work, but as subsystems that exchange information around the clock and continuously align their decisions with one another. The company becomes, as I call it in the book, an organism whose functions are connected in real time with each other and with the relevant environment.

Algorithms are the bosses, data is the currency.

This motto from the book is deliberately pointed. It does not describe a rule of machines, but a shift: operational control moves into software, and the quality of the data determines how well that control works.

An example: approving an expense receipt

How this shows up in everyday work can be described with the four-eyes principle of dual approval, a classic end-to-end process. A hospitality receipt for 1,000 euros is submitted from the field sales team. The responsible manager checks whether the associated business purpose is plausible and forwards the receipt to an authorized signatory for approval. Then it is paid, posted, and at some point shown in the monthly report. Three people, three handovers, several days.

In the self-driving company, the receipt and all data linked to it, meaning the customer appointment, the project, the budget, and the travel policy, are available in machine-readable form. Rule-based and AI-based algorithms make the decision immediately and trigger all follow-up actions at the same time: the payment instruction, the posting, the update of the liquidity plan and of the running project calculation, and if needed the simulation of scenarios, for instance what effect another customer visit would have on the project margin.

The difference is not that one process is completed faster. Several decisions are made at once and several actions are triggered at once, and the system status can be retrieved at any time, not only after month-end in a laboriously prepared report. The manager only sees what actually needs her or his attention: the exception.

How algorithms decide and who sets the goals

That several decisions can be made in parallel is due to the fact that software decides differently than a human being. Algorithms compute correlations from large volumes of data, which are perceived from the outside as a decision. In doing so they process far more facts in far less time. Humans decide even when facts are thin, intuitively and based on values, and afterwards often have to justify the decision laboriously.

One point that is frequently lost in the discussion matters here: algorithms do not make their decisions on their own. They pursue goals that people have set for them. Technically this happens through thresholds and defined outcome parameters within which the decision in the individual case is made. The programmers, and with them the management, make a meta-decision that sets the frame for all subsequent individual decisions. Whoever sets that frame runs the company.

Making mistakes is part of deciding. A learning algorithm frequently makes wrong decisions at the beginning; humans or other algorithms detect them and report them back. From this feedback the algorithm corrects its parameters, and the changed configuration already applies to the next decision. As a rule, it does not make the same mistake a second time. This feedback cycle explains why every algorithm has to be trained at the beginning, and it also explains why an individual decision becomes hard for humans to trace after some time of learning. It remains verifiable in the larger context: by its results, measured against the goals that were set.

An example of how quickly people place trust in algorithmic decisions when the results are verifiable is investing. Robo-advisors managed assets of around 260 billion euros worldwide in 2017; for 2025, a volume of about 2.3 trillion euros is expected (Statista, Digital Market Outlook "Robo-Advisors worldwide", accessed November 30, 2021). Whether they outperform experienced portfolio managers in the long run is debated; that capital of this magnitude is entrusted to them is measurable.

The leverage lies in integration

Over the next five to ten years, companies will train more and more algorithms for individual specialized tasks: receipt checking, scheduling, pricing, maintenance planning. Each of them will soon perform its task more reliably than a manual workflow. But that is only the first half of the development. The second half consists of these algorithms continuously exchanging their status with all other areas. This creates a decision network that runs through the entire company.

The technical advances of recent decades were sums: more processors, more storage, more powerful devices, isolated AI applications. The leverage of this stage arises because all of these achievements are fully integrated for the first time. I therefore expect a transformation that is larger than any previous step of the four technical revolutions. Only with this integration can all decisions, the large ones as well as the small ones, be aligned with the overall optimum in terms of the strategic goals. The contradictory logics of individual systems, the withholding of information, and the covering up of errors are then overcome not through appeals but through transparency: every result is visible as soon as it arises.

That this step is becoming necessary also has to do with the volume of data. According to a study by the International Data Corporation, cited by Statista 2020, the global volume of data grows from 33 zettabytes in 2018 to 175 zettabytes in 2025. The problem for companies is not storing this data but actively using it. That can no longer be done with linear processes and manual analysis.

The organization follows the technology. When operational control lies in networked software, organizational charts and hierarchies lose their justification; departmental boundaries and the boundaries to partners and suppliers become permeable. Companies that cling to them will face competitors that decide faster, more precisely, and at lower cost.

What this means for decisions today

Hardly any company we advise is on the third stage today. Most are working on the first and second. The dissolution of processes is nevertheless not a distant vision but a question of architecture that is being decided now. I consider five consequences central:

• Make data machine-readable, not just store it digitally. A PDF in the inbox is no progress if software does not understand its content. Without this foundation, none of the further stages can begin.
• Design processes for their dissolution. Anyone automating a workflow today should describe it as states, events, and rules, not as a chain of handovers. That way it can later be transferred into a networked system instead of being built again.
• Make meta-decisions explicit. Thresholds, target values, and decision latitude need to be documented and owned before software applies them. That is leadership work, not a programming task.
• Plan for feedback loops. An algorithm whose errors nobody reports back does not learn. Whoever deploys it needs a way to check results and feed deviations back.
• Dismantle departmental boundaries as data boundaries. The overall optimum is only achievable if the systems of the individual areas see the same status. Every silo created today is an interface of tomorrow.

My assessment is that the productivity gap between companies that take this path and those that keep optimizing process chains will be larger than at any previous stage of digitalization. End-to-end processes have done their job. They belong in economic history, not in the target architecture.

The next step

The foundations of this article, the autonomy levels and the chapter on the victory of the algorithms, are described in detail in the book "Das selbstfahrende Unternehmen"; you will find a summary in the first article of this series. More about the book and the vision is on the page The self-driving company. If you want to know which stage your company is on and which processes can be dissolved first, book an expert consultation.