Overcoming Cognitive Biases in Agile Estimation Methods - Developing a Taxonomy of Psychological Influences and Artificial Intelligence Solutions Approaches

Context
The increasing complexity of software systems necessitates an agile approach to realise such projects efficiently and successfully. As part of these agile processes, estimation methods are essential to determine the amount of work scheduled for individual requirements of the system and thus enable the planning of the project. These estimation methods are often based on human judgment and decision-making, meaning that they are susceptible to various psychological biases.

Objectives
The objective of this thesis is to examine the interaction between agile estimation methods and psychological factors and to assess how the latter can be mitigated through the integration of artificial intelligence (AI) in the former. The research concentrates on identifying the psychological aspects inherent in agile estimation processes, analysing the vulnerability of these methods to such psychological influences, and exploring the potential of AI to minimise them effectively.

Method
To answer these questions, a comprehensive taxonomy was developed over five iterations. Initially, the analysis included 64 agile estimation methods and 177 psychological aspects. By applying specific inclusion and exclusion criteria, this number was reduced to nine agile estimation methods and 20 psychological aspects. For these selected methods, susceptibility to the identified psychological aspects was assessed using defined criteria. The taxonomy also elucidates the manner in which the impact of these psychological factors can be mitigated through the deployment of AI.

Results
The iterative process resulted in the matrix below. The x-axis represents the 20 psychological aspects, while the y-axis represents the nine different agile estimation methods. The cells in the matrix indicate whether a particular psychological aspect cannot (green), can (orange), or always (red) occurs in the agile estimation method. The matrix makes it clear that all the agile estimation methods analysed are influenced by psychological factors and can therefore negatively affect the estimation process. However, there are significant differences in both the specific psychological aspects that influence each method and the extent to which they are susceptible to these aspects.

Furthermore, the research identified eight key measures that AI must fulfil in order to reduce the psychological influences on agile estimation processes: (1) protecting the anonymity of estimators to allow for unbiased opinions, (2) presenting all initial estimates to all stakeholders simultaneously, (3) providing and presenting consistent and coherent information so that all estimators are working from the same information base, (4) offering alternative estimates to encourage a broader perspective and reduce reliance on a single estimate, (5) comparing current estimates with historical data from completed projects to assess their realism, (6) making the production of each estimate traceable to promote transparency and understanding, (7) encouraging the active participation of all estimators to reduce reactance, and (8) evaluating the current emotional state of estimators to recognise and address emotional biases.

In light of these findings, it is recommended that agile teams be made more aware of the effect of these psychological aspects in the short term. In the long term, quantitative studies should be conducted on this topic, and the development of AI systems that can be integrated into agile estimation processes should be promoted.