Academic Writing (AW)
Neu-Ulm University of Applied Sciences
November 30, 2025
Introduction sets expectations for entire thesis. A good introduction promises: ’here is an important problem, here is how I will solve it, and here is the new knowledge the research will produce.’
If these points are unclear, readers disengage—even if your methods and findings are excellent.
And: a strong introduction makes the rest of your writing easier.
From good to great — perfecting your thesis introductions.
To demonstrate areas for improvement, we will use real examples from your cohort.
Every introduction needs:
Key principle: Each element gets its own paragraph(s)
Most introductions follow the formula,
with some doing so particularly well.
Others make a good attempt, but have the following issues:
As artificial intelligence (AI) becomes increasingly integrated into organizational decision-making processes, its benefits and associated risks have become a top priority for users and researchers. AI systems now determine who gets hired and who is evaluated for performance, and it’s doing this through a new form of algorithmic accountability and ethical exposure (Bauer & Gill, 2024; Rhue, 2024). With the increasing popularity of Responsible-AI frameworks seeking to promote fairness, transparency, and trustworthiness (Papagiannidis, Mikalef, & Conboy, 2025), making employees aware of and adhering to these principles becomes increasingly important. Yet, recent insights indicate that, even when employees use AI systems that they have only a partial understanding of, they tend to suffer from ambiguity and moral strain (Jussupow, Benbasat, & Heinzl, 2024). This increasing divergence between organizational AI governance aspirations and employees’ perception of risks associated with AI use highlights a new issue for the sustainable and ethical implementation of AI.
What works here? What could improve?
Could add more urgency—what’s at stake right now?
AI systems now make over 50% of hiring and performance decisions in Fortune 500 companies [Source], yet 73% of employees report they don’t understand how these systems work or when to question their recommendations [Source]. This opacity creates a critical governance challenge …
Consider structuring your first paragraph as follows:
From context (what is happening in the world?)
to importance (why is this timely/important now?)
to the puzzle (what is problematic/surprising?)
Technological innovation is widely regarded as a primary catalyst for competitiveness, growth, and long-term corporate development (Damanpour 1991). In knowledge-and technology-intensive industries in particular, a company’s ability to develop and successfully implement new products and processes is a determining factor in its adaptability and survivability in dynamic markets (Teece 2007). Whilst innovation research has historically concentrated on large, professionally managed companies, in recent years the innovation performance of small, owner-managed companies has attracted increasing attention (Kearney et al. 2017; Massis et al. 2017).
The owner-managed small and medium-sized enterprise (OMSME) is a unique entity characterised by a high degree of personal and cognitive influence of the owner on strategic and operational decisions. The owner’s dual roles as manager and central decision-maker result in the presence of short communication channels, informal processes, and a pronounced personality-driven organisational culture (Hambrick and Mason 1984; Zahra 2018).
What is the probleme here?
The opening immediately launches into literature—this is background (what research tells us), not a hook (what’s the problem?).
The reader is confronted with theoretical statements before understanding why this topic matters and/or what organizational challenge exists.
Possible revision:
Owner-managed small and medium-sized enterprises face a critical paradox: they must continuously innovate to survive in dynamic markets, yet they operate under constraints that make innovation systematically difficult [Source]. Unlike large corporations with dedicated R&D departments, professional management teams, and deep resource pools, owner-managed SMEs rely heavily on one person—the owner—who must simultaneously maintain daily operations, manage customer relationships, and drive innovation [Source]. This dual role creates constant tension: short-term operational demands crowd out long-term innovation investments. The stakes are high—70% of German SMEs are owner-managed [Source], forming the backbone of the economy, yet many struggle to sustain innovation beyond the founder’s initial ideas. The puzzle is: how do these resource-constrained, personality-driven organizations develop systematic innovation capabilities rather than relying solely on the owner’s sporadic insights?
Innovation research offers insights … Foundational work establishes …
Your first paragraph is a proper hook if it:
In research, […] generally called speech emotion recognition (SER). SER has moved from theory to practice in recent years. For example, Yurtay et al. (2024) trained deep learning models on large volumes of real call-centre audio in Turkish and showed that it is possible to classify broad affective states. Similarly, Martín-Doas et al. (2024) presented an industrial pipeline for Spanish call centres that uses modern self-supervised speech models to extract emotion-related information at scale. However, emotion in a call is rarely static. […] Feng and Devillers (2023) respond to this limitation by modelling emotion as something that changes continuously over time. […] Macary et al. (2023) combined acoustic descriptors—how something is said—with linguistic descriptors—what is said—to estimate customer satisfaction continuously across a call. […] There is also a very practical challenge. Real service audio includes background noise, poor microphones, people talking over each other […]. Parra-Gallego and Orozco-Arroyave (2021) […] introduce phonation, articulation and prosody features that are specifically designed to survive harsh acoustic conditions.
What is the probleme here?
Reads like a comprehensive literature review, not a synthesized background paragraph.
What came to my mind is: When will this end? What’s the main point?
Possible revision:
Speech Emotion Recognition (SER) has evolved from static classification to dynamic, context-aware analysis. Early industrial applications classified entire calls with single emotion labels (Yurtay et al., 2024; Martín-Doas et al., 2024), but recent work recognizes emotion shifts throughout conversations, modeling continuous temporal changes rather than fixed states (Feng & Devillers, 2023). Researchers now combine acoustic signals (how something is said) with linguistic content (what is said) for richer analysis (Macary et al., 2023), and address real-world complexities like multilingual code-mixing (Abhishek & Bhattacharyya, 2023) and noisy environments (Parra-Gallego & Orozco-Arroyave, 2021). However, these advances share a limitation: they focus on technical accuracy without addressing how emotion-aware systems should ethically support service interactions.
The background is properly synthesized if it:
Synthesis patterns to use:
[…] Despite these technological advancements1, two critical gaps remain in literature. First, there is limited research on how service ticket data can be systematically analyzed to detect workforce skill gaps and […]. Organizations continue to rely heavily on lagging indicators […] while overlooking the wealth of frontline knowledge embedded in support documentation. Second, […] there remains a lack of comparative studies focused on their application in enterprise learning contexts. It is still unclear which method offers the best trade-off between accuracy, explainability, and integration feasibility […]. These gaps […] leave practitioners without clear guidance on how to adopt AI tools for learning analytics. This thesis aims to address these issues through the following research question: What AI-based framework should be designed and evaluated to identify training gaps […], and how do its constituent methods (classic NLP pipelines and local LLMs) compare in effectiveness and integration suitability? To investigate this question, the study follows a Design Science Research (DSR) methodology […]. The research involves building […]
What is the probleme here?
The gap and solution run together in one paragraph—no boundary between “what’s wrong” and “what I’ll do about it”
The gap feels like two disconnected problems forced together:
The gap’s coherence—and thus the introduction—could be strengthened by focusing on one core problem.
Almost all factual claims without any evidence
Possible revision (without necessary references):
Despite advances in learning analytics, organizations continue identifying skill gaps reactively through surveys and performance reviews, detecting problems 6-12 months after they emerge. This costs organizations in repeated service incidents that trained employees could prevent. Meanwhile, thousands of daily support tickets document exactly which knowledge gaps frontline teams encounter, yet this operational intelligence remains systematically underutilized. Without methods to extract training insights from service interactions, organizations miss opportunities to proactively address emerging skill deficiencies before they cascade into customer-facing failures.
To address this gap, this study develops an AI-based framework that transforms support ticket data into predictive training recommendations using DSR methodology. The framework compares traditional NLP pipelines with locally-deployed LLMs to provide evidence-based guidance for organizations choosing between approaches for learning analytics applications.
[…] Consequently, data products today primarily provide the technical foundation for scaling advanced services but not the dual value logic required by complete servitization. Achieving this would mean integrating external customer outcomes and internal process optimization in a continuous feedback process.
Based on these observations in current servitization, data product, and Industry 4.0 literature, the following research question is posed: How should data products be designed in terms of their concept, monetization strategy, and organizational anchoring to fulfill the full servitization logic as “factory-integrated substituting services” and achieve a two-sided value contribution for both customer outcomes and internal processes?
What is the probleme here?
The paragraphs are technically separate, but the boundary is weak.
The vague transition (“Based on these observations…”) doesn’t clearly signal the move from gap to resolution.
Possible revision (sources need to be added):
[…] Consequently, data products today provide technical infrastructure for scaling services but lack the dual value logic of complete servitization. This limits their potential: organizations can improve internal data management but cannot […]. Without integrating servitization principles, data products remain internal efficiency tools rather than revenue-generating service offerings, leaving organizations unable to […].
To address this gap, this thesis investigates how data products can evolve beyond internal management tools to become servitized, outcome-oriented offerings that fulfill the complete servitization logic. The research question asks: How should data products be designed in terms of concept, monetization strategy, and organizational anchoring to achieve two-sided value contribution for both customer outcomes and internal processes? Using conceptual research that bridges servitization and data product literature, the study develops design principles for organizations establishing service-oriented business models around data assets.
Literature review discusses:
After 5-6 paragraphs of literature review without synthesis:
Together, these works highlight a common problem: evaluation methods focus either on deterministic adversarial examples or qualitative jailbreak demonstrations, but rarely provide probabilistic estimates of model robustness under varied attack conditions. However, these two approaches have so far developed separately […]
What is the probleme here?
The gap doesn’t have its own paragraph—it emerges implicitly after extensive literature review.
The reader must piece together what’s missing from scattered hints throughout the background section, rather than encountering a clear, dedicated gap statement.
Possible revision (sources need to be added):
Adversarial robustness research has evolved from early image-based attacks to LLM-specific vulnerabilities like prompt injection and jailbreaking. Current evaluation approaches fall into two camps: deterministic adversarial testing that measures specific attack success rates, and qualitative jailbreak demonstrations that show proof-of-concept exploits.
However, these evaluation methods share a critical limitation: neither provides probabilistic estimates of overall model robustness under varied attack conditions. Deterministic tests measure narrow scenarios but don’t quantify the full attack surface. Qualitative demonstrations show vulnerabilities exist but don’t measure their likelihood or severity. This leaves organizations deploying LLMs unable to estimate: “What is the probability this model will fail under adversarial pressure?” Without probabilistic robustness metrics, safety teams cannot make risk-informed deployment decisions or compare model resilience systematically.
The consequences of algorithmic persuasion in health situations have recently been the subject of scholarly investigation. According to Liu-Thompkins et al. (2022), emotionally intelligent algorithms can mimic empathy to increase user engagement, but if used without ethical guidelines or transparency, they may also have unforeseen psychological effects. […] A significant gap remains where the literature provides little information about the impact of emotional targeting in AI-driven health-related advertisements on users’ anxiety or perceived autonomy in the area of making health decisions, in spite of advancements in associated fields like algorithmic bias (Mittelstadt et al., 2016), misinformation (Bridgman et al., 2020), and chatbot-based mental health tools (Li et al., 2023).
What can be improved here?
The intro only states what’s missing from literature, not what’s broken in practice.
No explanation of real-world harm or organizational/individual cost
Possible revision (sources need to be added):
Emotionally intelligent algorithms now target vulnerable health decisions through fear and hope appeals, yet we cannot predict or prevent the psychological harm this creates. Research shows AI can mimic empathy to boost engagement (Liu-Thompkins et al., 2022), but without understanding how emotional targeting affects anxiety and autonomy, platforms deploy these systems blind to their psychological impact. The consequence: users experiencing health crises—cancer diagnoses, chronic conditions, mental health episodes—encounter AI-optimized ads designed to exploit their emotional vulnerability. These users cannot distinguish genuine health information from emotionally manipulated persuasion, leading to anxiety-driven decisions (hasty treatments, unnecessary purchases, delayed care) that worsen health outcomes. Platforms profit from emotional targeting while remaining unaccountable for the psychological harm, leaving users and regulators unable to establish ethical boundaries for AI-driven health advertising.
Your tension and resolution paragraph(s) are properly separated if:
Reframe absence framing to problem framing (e.g., “no research exists on […]” to “organizations cannot […] because […]” )
After stating your gap, ask yourself so what? until you reach a tangible real-world-impact.
How does emotional microtargeting in AI-powered social media health advertisements affect users’ decision-making and health anxiety?
What is vague here?
Multiple undefined terms that need operationalization:
Reading the introduction, I cannot determine what you’ll actually focus on (and how you will measure it).
Possible revision
RQ1: How do fear-based versus hope-based emotional appeals in AI-targeted health advertisements influence individuals’ tendency toward immediate action or deliberative consideration in treatment decisions?
RQ2: How does this relationship differ based on baseline health anxiety levels?
How can an AI-based assistant system be designed to support junior project managers in the practical application of project management methods during the planning and executing phases?
Sub-questions:
What is the problem with the (multitude of) questions?
Combines design, perception, effectiveness, and adoption questions—this indicates at least 3+ separate studies:
Also vague:
More focused question:
How does an AI assistant providing real-time risk identification prompts during project planning improve risk coverage completeness among junior project managers compared to traditional checklist-based planning?
How do different explanation designs affect advice-seeking users’ appropriate reliance on Al career guidance?
What’s vague here?
Multiple undefined terms:
More operationalized version:
RQ1: Do SHAP feature importance explanations (versus simple confidence scores) improve students’ alignment with expert career counselor judgments2 when evaluating AI-generated major selection advice?
RQ2: How does this effect vary by students’ career decision-making self-efficacy?
Test if your RQ is well-operationalized by asking following questions:
This study contributes both theoretically and practically. Theoretically, it extends information systems research on XAI and human-AI collaboration in financial contexts. Practically, it offers design principles for organizations implementing cloud cost forecasting systems.
What’s too generic here?
Multiple generic claims without specifics:
Missing:
More specific version:
This study makes three specific contributions. Theoretically, it extends XAI effectiveness research by demonstrating that explanation utility is moderated by decision stakes and user expertise—challenging the assumption that more transparency always improves trust. The study contributes a contingency framework specifying when SHAP explanations help versus harm decision quality in financial forecasting. Methodologically, it provides a validated instrument for measuring explanation comprehension in financial contexts. Practically, it delivers six design principles for cloud cost forecasting UIs, specifying: (1) when to show SHAP values vs. confidence intervals, (2) expertise-based explanation scaffolding, and (3) progressive disclosure patterns for complex forecasts.
The expected contribution of this thesis is threefold. First, it delivers design knowledge for AI-powered assistance tools tailored to the needs of junior project managers. Second, it offers empirical insights into user acceptance, task alignment and human-AI collaboration in project environments. Third, it contributes to the academic discourse on AI in project management by applying and integrating TAM, TTF and Human-AI Complementarity into a coherent and practically tested design approach.
What’s too generic here?
Lists contribution areas but no concrete deliverables:
Missing:
More specific version (aligned to the refined RQ):
This study makes three contributions. First, it provides evidence that real-time AI risk identification prompts improve risk coverage completeness compared to traditional checklists among junior PMs. Second, it challenges the assumption that more information aids novice decision-making—demonstrates that interactive prompting outperforms comprehensive checklists because it reduces cognitive load during planning (supporting cognitive load theory in PM contexts). Third, we provide a validated prototype demonstrating effective prompt timing, prompt specificity, and prompt density that organizations can implement in existing PM tools.
Strong contribution statements specify:
Use specificity test and ask
Great introductions aren’t written, they’re rewritten!
Find slides here:
(only for your personal use, do not share)
Sidenote: the sentence before does not discuss advantages.
Alignment with expert careed counselor judgements is measured as agreement rate with expert-endorsed recommendations minus agreement rate with expert-rejected recommendations.