Bremermann’s Limit

According to Bremermann’s limit there is an insurmount-

Year-End Forecasts 2018
CONTROLLING

able, absolute limit to human knowledge, which cannot

Truth (Year-End-Closing 2018)
be overcome, no matter how great the advancements in
Manual Current Forecast (CF) digitization are. Due to the atomic structure of matter

Artificial
Predictive Analytics
there is an upper limit to information processing, which
cannot be surpassed by any computer or brain made of
matter: no system consisting of matter can process more

Intelligence
than 2*10⁴⁷ bits per gram and second, corresponding to PROF. DR. HEIMO
the speed of light and the Plank constant.⁴ As a conse- LOSBICHLER
quence not even the strongest cloud-based computer
clusters, like i.e. Hadoop, have enough processing power Is the chairman of the

in Controlling
International Association of
to make exact forecasts in today’s complex, competitive
Controllers (ICV) in Munich,
CF02 CF03 CF04 CF05 CF06 CF07 CF08 CF09 CF10 CF11 environment. Malik made an interesting comparison in
chairman of the International
his habilitation thesis, in which he calculates the theoreti- Group of Controlling (IGC)
cal limit of the information processing capacity based on based in St.Gallen as well as
Fig. 1: Machine vs. human year-end forecast (real-life example) the assumption that the entire earth’s mass since the start the head of the program for
Applications and limits of the history of the earth was a giant computer, that was Controlling, Accounting and
permanently processing information. He then compares Financial Planning and Dean
of machine forecasts this information processing capacity to the complexity of of the School of Business
and Management of the
The differences between human and machine forecast typical management decision-making situations and
FH Oberösterreich in Steyr.
can be plausibly explained by the complementarity of thereby shows the limited forecasting ability.⁵
human and machine information processing. However,
realistic expectations regarding the predictive accuracy Partial Ascertainability and Controllability
of machine planning and forecasts are in order, despite of Complex Systems
the positive experiences, because there are limits to the Figure 2 shows the structure of complex systems like that
ascertainability and planning capacity of AI in a VUCA- of our modern economic system. They consist of a multi-
The effects of digitization can be felt in two very differ- classical planning in the early 2000s. In the wake of the context as well. Even the machine forecast in the “suc- tude of elements ⁶ (a to h) and relations (arrows between
ent fields of controlling. On the one hand in the automa- 2008 financial crisis the term VUCA – Volatility, Uncer- cessful example” of figure 1 was considerably too opti- the elements), while the system is split into one part that
tion of repetitive routine tasks (Robotic Process Automa- tainty, Complexity and Ambiguity - established itself as mistic until August. These limits will be discussed from is visible (a, b, d, e, g, h) to the actor A (manager, control-
tion) and on the other hand in the support or automation synonymous with the difficulty of predicting future de- the perspective of complexity and cybernetics. ler) and one part that is invisible (c, f). An example for an
of challenging, analytical tasks (machine forecasts, AI). velopments. As an answer to this “new normal”, concepts invisible element would be the corona virus before its
While the automation of routine tasks has been pro- like Modern Budgeting, Scenario Planning, Fluctuation outbreak. This has an important consequence: we do not
ceeding quite well, especially in large companies, the Margin Planning or Rolling Forecasts have been intro- The Limits of Predictability know that certain elements exist and cannot factor them
support of analytical tasks seems much more difficult. duced, which propagated the departure from detailed, due to Complexity and Cybernetics into our decisions. Thus, the system is only partially as-
According to a study of the German Federal Ministry of pinpoint planning and prognoses in various ways. certainable and therefore it can also only partially be
Economics, only 5% of German companies use AI in one Dealing with complexity is nowadays regarded as one of modelled via AI-systems. Fig. 2: Structure
of their business units.¹ The amount of companies that But with the advent of the digitization a new paradigm the biggest challenges in management. Managers have of complex systems [7]
use AI in controlling is thus vanishingly small. At the shi�t seems to have started. The access to new data to consider an ever-increasing number of factors in cor-
same time there are high expectations for AI-systems in sources (Big Data), nearly unlimited processing power porate management, that also change at an increasingly
controlling.² This article was written before the start of and AI-systems have quickly led to buzz words like Pre- rapid pace and are highly interconnected. Key drivers of
the corona crisis. It illuminates both the limits of the dictive Analytics and the first applications of AI-based this development are globalization and paradoxically the
prognostic ability, and the possible applications, of ma- machine forecasts. Through this, the faith in the predict- rapid progress of digitization, that connects the world in A
chine forecasts. ability of the future has been resurrected – at least until real-time and increases its speed of change. The han- a b c
the start of the corona crisis. The few reports of experi- dling of complexity has been studied by cyberneticists in
ences, primarily from large companies, seem to confirm particular. Pioneers like Ashby, Beer, Forrester, Luhmann,
d f
A Paradigm Shi�t in the Area of Planning, the feasibility of making predictions via AI and the supe- Ulrich, Probst, Gomez, Malik, Dörner or Vester have al-
Budgeting and Forecasting? riority of machine forecasts. Figure 1 shows the monthly ready laid the groundwork for this a long time ago, which,
e g
development of human and machine year-end forecasts considering the limits of artificial intelligence, is now
Lamenting an uncertain environment that is hard to pre- of a large international company. Here the machine fore- more current than ever. Here Bremermann’s limit and h
dict, the premature obsolescence of plans and budget cast has indicated the downturn three months earlier the partial ascertainability and controllability of complex
policy “games” has a long history. The Beyond Budgeting than the controller, and also predicted the end-of-year systems have been selected as examples.
Round Table (BBRT) made loud demands for the end of result a little more precisely.

4 Controller Magazin _ Special 5

 Complex systems can be further divided into can be achieved with lower expenses through covered at length in literature. The long list Human-Machine Applications place or supplement human forecasts? A think that they are right, not the system. In
active elements (b, d) that can change inde- automation, on the other hand, through the of identified biases is proof of that. The fol- and Division of Tasks distinction might be made between differ- the last stage of Autonomous Intelligence,
pendently, and passive elements (a, c, e, f, h, complementarity of human and machine in- lowing examples show typical human short- ent support levels of “Assisted Intelligence, the machine forecasts replace the human
g). Complex systems have their own momen- formation processing, improvements in qual- comings in the creation of forecasts: The discussion so far has shown that Augmented Intelligence, Autonomous In- forecasts; controllers and managers rely on
tum through these active elements. They do ity can be achieved. telligence”, similar to the way it is done in the AI-system. Alongside the support level,
not wait for the interventions of the actor, but ▶ Overconfidence frequently leads to opti- ▶ AI-systems, or rather machine forecasts, autonomous driving.¹¹ In the case of Assisted the question of the requirement level for the
rather change themselves. Both the elements mistic prognoses are still not very widespread and in their Intelligence, the entire forecast process re- AI must also be considered. In analogy to
themselves and the relationships between infancy, but they are of great importance mains in the hands of the controller. The AI the stages of development of analytics, the
the elements can change without an outside
The Complementarity ▶ People subconsciously adjust their prog- and hold great future potential. or the machine forecast does support work, expectations for the AI-system might only
influence. Consequently, the input (interven- of Human and Machine noses to an “anchor” or point of reference. following the specific requirements of the be the provision of relevant information on
tions from management) is not the only thing Information Processing In the case of a forecast this can, for exam- ▶ The ideal of exact prognoses will remain controller. The controller then decides the variances, as a basis for the actual prognosis
that determines the output now. In truth the ple, be the figures from the budget or an unattainable one, even in the age of ar- result of the forecast. For Augmented Intel- (descriptive, diagnostic). But in most cases
output depends on the input and on the The question why machine forecasts could from the previous year tificial intelligence, their use, however, ligence, the forecast of the controller and the companies are not content with that and im-
states of the system. That is why their behav- be superior to human ones, can largely be can enhance human predictive power and machine forecast are created in parallel. The plement a quantitative forecast (predictive).
ior constantly surprises us. Forrester called answered from the perspective of shortcom- ▶ The receptiveness to new information in- automate or support the creation process. differences are analyzed and the controller The highest requirement level is set for an
them counterintuitive because known phe- ings of human rationality. The achievements, creases if they support the intents of the or the manager decides which result will be AI-system, that not only predicts the likely
nomena suddenly behave differently than we or rather the limitations, of the human brain decision-maker ▶ People do, in fact, have cognitive abilities used. An example for using Augmented In- result, but also the measures necessary to
would expect from experience. ⁸ This also ap- can be summarized as follows:⁹ that machines do not have (yet). telligence in the forecast process is the SAP attain it (prescriptive). But that still seems
plies to machine forecasts based on artificial ▶ Power-based distortion of information, AG. If the discrepancy between the forecasts like a vision of the future from today’s point
intelligence, which ultimately should make ▶ People can only use the information they like the loss of prestige, leads to progno- The question therefore is, how machine exceeds a certain threshold value, the con- of view. ⬛
accurate predictions for the future based on have learned or that is quickly available ses being upheld, even when the opposite forecasts can be used best. Should they re- cerned departments must explain why they
historical data (states of the system). The mo- externally (e.g. on paper). The human case is already becoming apparent.
mentum of complex systems, while taking brain shows weaknesses in the retrieval of
Bremermann’s limit into consideration, has information. ▶ Distance bias – because distant problems
profound consequences: the ideal of exact seem less important than immediate one,
Prescriptive Analytics
forecasts becomes impossible. Ultimately, ▶ T he human scope for solving problems is negative developments are not communi- (How can we
we must make do with patterns. fairly small. Only a few pieces of informa- cated right away make it happen?)
tion can be processed at the same time. No Foresight
Finally, managers in complex systems only more than 5-9 informational or sensory The examples given show, that the forecast- Predictive Analytics priority field of application
have limited management options. To units, so-called ‘chunks’, can be processed ing quality can be enhanced through the use (What will happen?) of machine forecasts
achieve objectives, the actor must change the simultaneously in short-term memory.10 of machine forecasts. On the one hand a
state of certain elements. The elements of greater amount of information can be includ-
Diagnostic Analytics
the system are broken into elements that can ▶ The brain gets tired and can only continu- ed in the forecast, on the other hand machine Insight
(Why did it happen?)
be influenced directly (dashed line from the ously solve problems for a limited time. forecasts are not subject to interest-based
actor to the elements a, d, g), elements that Constant thinking over a longer period is distortions (“emotionless forecast”). But you
can be influenced indirectly (b, e, h), and ele- accompanied by an increasing frequency should be careful with this. An essential prin- Descriptive Analytics
Hindsight
(What happened?)
ments that cannot be influenced (c, f). In ad- of errors. ciple of artificial intelligence is the ability to
dition, the elements can hardly be influenced learn and to improve itself. Optimization al-
in isolation, because they are highly intercon- ▶ The brain works relatively slow. The speed, gorithms can determine the accuracy of the Assisted Augmented Autonomous
nected and the actor is influenced by the ele- however, depends on the shape and famil- model and adapt it to heighten its future ac- Intelligence Intelligence Intelligence
ments in turn (dashed lines from the ele- iarity of the problem type: lighting-quick curacy. Even though AI-systems do not have
ments a, e, h to the actor). Thus, not only the human pattern recognition whether an self-interests, human biases can be learned Fig.3: Applications and support levels of machine forecasts
forecasting ability is limited, but the control- apple is fresh or rotten vs. the sluggishness unintentionally through the data made avail-
lability as well. when doing mental arithmetic. able to the system. Besides the limitations of 1. Feser, D., Hürden für den Einsatz künstlicher Intelligenz, ZfO, 1/2020, p. 17
the human brain, one of its major strengths is 2. Seufert, A., Treitz R., Künstliche Intelligenz und Controlling, Controller Magazin Special, May/June, 2019, p. 20
In summary, it can be deduced from these Beside those capacitive “skill-deficiencies”, worth mentioning too. The human brain reg- 3. IBM (pub.), Global CEO Study – Unternehmensführung in einer komplexen Welt, 2011
4. Bremerman, H.J., Optimization Through Evolution and Recombination, in: Yovits, M.C./Jacobi, O./Goldstein, O. (pub.), Self-Organizing Systems, p. 93-106
two fields, that, from a cybernetic and a sys- there are also behavioral shortcomings. For ularly solves problems that were not posed as
5. Malik, F., Strategie des Managements komplexer Systeme, Bern, 1986, p. 199
tem-theoretical perspective, the ideal of ex- example, people are content with reaching well. The brain does not have a static struc- 6. Reeves, M., Levin, S., Fink, T., Levina, A., Taming Complexity, Harvard Business Review, January-February, 2020, p. 115
act forecasts will remain an unattainable ide- their individual level of ambition and not ture, on the contrary it is reorganized con- 7. Forrester H., Das intuitionswidrige Verhalten sozialer Systeme, in: Meadows, D.L./Meadows, D.H. (pub.), Das globale Gleichgewicht, Stuttgart 1974
al even in the age of AI and machine forecasts. necessarily the achievable maximum, or stantly. That is why problems are spontane- 8. Dörner, D., Kreuzig H.W., Reither F., Lohausen: Vom Umgang mit Unbestimmtheit und Komplexität, Bern 1983
9. Häfner K., Psychische Mobilität mit Informationstechnik, in: Bäck H. (Hrsg.), Der gut informierte Manager- Erfolgreiche europäische MIS- und EIS-Lösungen, p. 115-130, Köln 1991
This should not mean, however, that machine they make decisions for their personal gain ously seen in a new way. This characterizes
10. Miller, G., The Magical Number Seven, Plus or Minus Two: Some Limits on Our Capacity for Information Processing, in: Alexis, M./Wilson, C., Organizational Decision Making, Englewood
forecasts could not lead to improvements in and not for the benefit of the company. The the creativity and innovative capability of hu- Cliffs 1967;
controlling. On the one hand the same result cognitive limits and behaviors have been mans and is a major difference to machines. 11. Grotenhermen, J.G., Oldeweme, A., Bruckes, M., Uhlending, L., Künstliche Intelligenz im Personalwesen, ZfO, 1/2020, p. 6

6 Controller Magazin _ Special 7