13. Digital Research, Big Data, Analytics and AI

Introduction and Context

Digital technologies have changed not only the objects that business researchers study but also the evidence through which organisations can be understood. Transactions, searches, clicks, posts, sensor readings, digital communications, platform interactions and automated decisions leave traces of organisational and market activity. These traces can be analysed at a scale, speed and level of granularity that conventional surveys and interviews rarely permit. At the same time, digital abundance does not automatically produce valid knowledge. Data may be incomplete, commercially controlled, algorithmically selected, context-poor or generated by systems whose operating logic is inaccessible to researchers.

For doctoral researchers, the central issue is therefore not whether big data or artificial intelligence is fashionable, but whether a digital source is appropriate for a clearly formulated research question. A large dataset may provide precise measurement of a narrow behavioural signal while failing to explain meaning, motivation or institutional context. Conversely, qualitative digital research may reveal how people interpret platforms and algorithms but cannot support claims about population frequencies without an appropriate sampling logic.

This chapter examines digital research as a methodological field. It considers digital traces, platform data, text and network analysis, machine learning, data provenance, algorithmic bias, explainability, reproducibility and the responsible use of generative AI. It also addresses digitally transformed organisations without assuming that technology determines organisational outcomes. Technology is embedded in governance, work practices, power relations, professional norms and institutional environments. The research task is to explain these relationships rather than to treat technology as an autonomous causal force.

Learning Objectives

By the end of this chapter, the reader should be able to:

1. Distinguish digital traces, big data, platform data, administrative data and designed research data.

2. Select digital research methods in relation to research questions, units of analysis and claims.

3. Evaluate data provenance, quality, representativeness and reproducibility.

4. Apply text, network and machine-learning approaches critically rather than treating algorithms as neutral substitutes for theory.

5. Identify algorithmic bias, privacy risks, security concerns and ethical problems in digital research.

6. Assess explainability and the appropriate role of predictive models in doctoral inquiry.

7. Use generative AI responsibly in searching, coding, analysis, writing and research administration.

8. Study digital transformation as a socio-technical and organisational process.

From Digital Traces to Research Evidence

A digital trace is a record produced through interaction with a digital system. Examples include a customer review, a timestamped purchase, a web request, an employee-system login, a GPS coordinate or an interaction between users on a platform. Traces are often produced as by-products of activity rather than in response to a researcher’s question. This distinguishes them from designed data, such as survey responses or interview transcripts.

Big data is best understood through a combination of characteristics rather than a single threshold of size. Common dimensions include volume, velocity, variety, veracity and value. Volume concerns scale; velocity concerns the speed of generation and analysis; variety concerns multiple formats and sources; veracity concerns reliability and uncertainty; and value concerns the usefulness of data for a decision or inquiry. These dimensions are analytically important because increasing volume does not solve problems of validity. Millions of observations may represent the same narrow group, contain duplicated records or measure platform activity rather than the underlying construct.

Platform data are generated within digital infrastructures that organise interactions among users, firms, advertisers, workers and regulators. The platform is not merely a neutral container. Its interface, ranking rules, moderation policies, application programming interfaces and commercial incentives shape what becomes visible and recordable. A study of online reputation, for example, must consider that reviews may be solicited, filtered, ranked and removed according to platform rules. The observable dataset is therefore partly a product of platform governance.

Administrative and operational data can be valuable because they capture actual processes, such as procurement, customer support or employee turnover. Yet they may reflect institutional classifications and reporting routines rather than objective reality. A “resolved complaint” may mean that a case was closed in a database, not that the customer’s problem was solved. Researchers must investigate how categories are created, who enters them, for what purpose, and what incentives affect recording.

The distinction between data and evidence is fundamental. Data become evidence only through conceptualisation, measurement, contextual interpretation and an explicit warrant connecting observations to claims. A researcher studying employee engagement should not assume that message frequency is engagement, or that sentiment in an internal forum is an unbiased measure of morale. The construct, indicator, population and inference must be justified.

Designing a Digital Research Study

A defensible digital study begins with a research question, not with access to an attractive dataset. The question should specify the phenomenon, context, unit of analysis and intended type of claim. Descriptive questions may ask how patterns vary over time; explanatory questions ask why or under what conditions outcomes occur; interpretive questions ask how actors understand digital practices; and predictive questions ask how accurately an outcome can be estimated for new cases. These purposes require different designs.

Researchers should distinguish the unit of observation from the unit of analysis. Individual clicks may be observations, while the unit of analysis is a customer, organisation or market. Multiple records may belong to one actor, creating dependence and a risk of overstating the effective sample size. Similarly, a network edge between two accounts is an observation, but the research may analyse communities, organisations or the network as a whole.

A useful design sequence is:

• formulate the theoretical or practical problem;
• define constructs and plausible mechanisms;
• identify the population and boundaries of the digital system;
• establish data access, provenance and ethical permissions;
• assess coverage, missingness, duplication and measurement error;
• choose analytic methods appropriate to the question;
• pre-specify important decisions where feasible;
• analyse, validate and interpret results in context; and
• document data, code, transformations and limitations.

Digital studies may use mixed methods. A researcher might first identify patterns in service records, then interview employees to explain how those patterns arise. Alternatively, qualitative platform research may inform the construction of a computational dictionary or the interpretation of a predictive model. Integration should occur at the level of research design and inference, not merely by placing quantitative and qualitative sections side by side.

Analytical Approaches for Digital Business Research

Text and Language Analysis

Text analysis may be qualitative, computational or mixed. Qualitative approaches examine meaning, discourse, framing, identity and power through close interpretation. Computational approaches may classify documents, identify topics, estimate sentiment, extract entities or compare language across groups and periods.

A text-mining workflow commonly includes corpus definition, cleaning, tokenisation, representation, modelling and validation. Older representations such as term-frequency matrices can be useful for transparent comparisons. Embedding-based methods represent words or documents in high-dimensional spaces and can capture semantic similarity, but their representations inherit patterns from training data. Large language models can classify or summarise text, yet their outputs may vary with prompts and may introduce fabricated or unsupported content.

Sentiment analysis illustrates the need for caution. A negative word does not necessarily indicate negative evaluation; sarcasm, negation, industry-specific language and cultural differences can change meaning. A model trained on product reviews may perform poorly on employee grievances. Validation should therefore include domain-relevant labelled examples, error analysis and, where possible, comparison with human coding. Researchers should report the dictionary, model, prompt, version, thresholds and treatment of ambiguous cases.

Network Analysis

Network analysis represents entities as nodes and relationships as ties. In business research, nodes may be firms, employees, customers or technologies, while ties may represent collaboration, ownership, communication, supply relationships or co-investment. Network data can reveal centrality, brokerage, density, community structure and diffusion patterns.

Centrality measures must be interpreted in relation to the network and theory. Degree centrality indicates the number of connections; betweenness captures a position on paths between others; eigenvector-style measures reward connections to influential nodes. None automatically means power, innovation or performance. A highly connected employee may be a coordinator, a bottleneck or simply a recipient of automated messages. Network boundaries and tie definitions are consequential: excluding informal relationships can alter conclusions about structure.

Network dependence also challenges conventional statistical assumptions. Observations are not necessarily independent, and processes may be reciprocal. A finding that connected firms adopt similar technology could reflect diffusion, shared suppliers, common industry conditions or homophily. Longitudinal designs, temporal ordering and suitable network models are needed before claiming contagion or causal influence.

Machine Learning and Statistical Modelling

Machine learning refers to computational methods that learn patterns from data for prediction, classification, ranking or representation. Supervised learning uses labelled outcomes; unsupervised learning identifies structure without a specified outcome; and reinforcement learning learns through feedback from actions. These categories are useful, but they do not replace research design.

Predictive performance should be evaluated on data not used to fit the model. A training set supports estimation, a validation set assists model selection, and a test set provides an assessment of generalisation. Cross-validation is useful when data are limited, although time-dependent data require temporal rather than random splitting. Accuracy alone may be inappropriate for imbalanced outcomes; precision, recall, calibration, area-based measures and cost-sensitive metrics may be more informative.

Prediction and explanation are related but distinct. A model can predict employee attrition effectively without identifying a causal intervention that would reduce attrition. Correlation, feature importance and model attention do not by themselves establish causal mechanisms. Causal claims require a design and assumptions suited to causality, such as experiments, natural experiments, longitudinal controls or credible identification strategies.

Overfitting occurs when a model learns idiosyncrasies of the sample rather than general patterns. Data leakage occurs when information unavailable at the time of prediction enters the model. Both problems are common in organisational datasets. For example, using a later resignation code to predict earlier resignation produces impressive but invalid performance. Researchers should maintain a clear temporal boundary and conduct robustness checks.

Provenance, Validity and Reproducibility

Data provenance is the documented history of a dataset: its origin, collection process, transformations, custodians, access conditions and subsequent uses. A provenance record should identify when and how data were collected, the software and versions used, changes to variables, exclusions, missing-data decisions and links between raw and analytical files. Provenance is not clerical detail. It permits researchers to evaluate whether a result is traceable and whether another researcher can understand the basis of a claim.

Digital data require several forms of validity assessment. Construct validity asks whether indicators represent the intended concept. Coverage validity asks which members or events are included and excluded. Criterion validity examines relationships with an appropriate external measure. Ecological validity asks whether findings travel beyond the digital setting. Temporal validity is especially important when platforms, interfaces, policies and user populations change.

Reproducibility has at least two dimensions. Computational reproducibility means that the same data and documented code produce the reported result. Conceptual or scientific reproducibility concerns whether a study’s conclusion is supported when the question is examined with new data, a different context or an alternative operationalisation. Commercial and privacy restrictions may prevent full data sharing, but researchers can still provide synthetic data, code, data dictionaries, restricted-access procedures and detailed methodological documentation.

Open practices should be proportionate to risk. Releasing raw employee messages or identifiable location trails would be irresponsible merely because openness is desirable. De-identification is not always sufficient, especially when multiple datasets can be linked. Governance must balance transparency, confidentiality, security and the legitimate interests of participants.

Bias, Explainability and Responsible Algorithmic Inquiry

Algorithmic bias can enter at multiple stages: problem formulation, sampling, labelling, feature construction, model training, deployment and interpretation. Historical data may encode unequal treatment. A model may appear accurate overall while producing systematically different error rates for groups. Proxy variables can reproduce protected characteristics even when those characteristics are removed. Bias is therefore not eliminated by deleting demographic fields.

Fairness is multidimensional. Equal selection rates, equal error rates, calibration and individual similarity can conflict, particularly when groups have different base rates. Researchers must state which fairness principle is relevant and why. A technical fairness metric cannot decide the normative question of what treatment is justified.

Explainability refers to the ability to understand how a system produces or supports an output. A transparent linear model may be easier to interpret globally, while a complex model may require local explanations for individual predictions. Explanations can be useful for debugging, contestability and managerial review, but post-hoc explanations are not always faithful representations of model reasoning. They should be treated as evidence about model behaviour, not as automatic proof of causal logic.

Responsible research requires attention to privacy, consent, confidentiality, security, autonomy and potential harm. Public availability does not necessarily imply ethical permission for unrestricted analysis. Researchers should consider whether users reasonably expected their data to be collected or repurposed, whether vulnerable groups are involved, and whether publication could enable identification or discrimination. Data minimisation, access controls, retention limits, impact assessments and human review are practical safeguards.

Generative AI in the Research Workflow

Generative AI systems can assist with literature discovery, transcription support, code suggestions, data documentation, translation, synthetic examples, preliminary classification and language editing. Their use does not transfer scholarly responsibility to the system. Outputs may be inaccurate, fabricated, biased, non-reproducible or based on undisclosed training sources.

A responsible workflow treats generative AI as an auditable assistant rather than an author or evidence source. Researchers should:

• avoid entering confidential, personally identifiable or proprietary data into unapproved systems;
• verify every substantive claim against authoritative sources or the underlying data;
• retain prompts, model names, dates, settings and edited outputs when they materially affect the research;
• distinguish machine assistance from human judgement in methods and acknowledgements;
• inspect generated code for security, logic and licensing problems; and
• ensure that participants, co-authors, supervisors and publishers are informed according to applicable policies.

Generative AI is particularly risky in literature reviews because it may invent citations or misrepresent findings. It can also amplify majority-language assumptions and flatten theoretical disagreement. In qualitative analysis, it may assist with candidate coding, but researchers must preserve interpretive accountability and assess whether sensitive narratives are being exposed. In quantitative work, generated code must be tested against known examples and independently reviewed.

Digital Transformation as a Socio-Technical Phenomenon

Digital transformation should not be reduced to the purchase of software or the adoption of artificial intelligence. It involves changes in work, authority, routines, capabilities, business models, customer relationships and institutional expectations. Socio-technical perspectives emphasise that technologies and social arrangements mutually shape one another. The same system can produce different outcomes in organisations with different skills, incentives, leadership practices and governance structures.

Several theoretical lenses are useful. The technology–organisation–environment perspective directs attention to technological characteristics, organisational readiness and external pressures. The resource-based view asks whether digital capabilities are valuable, rare, difficult to imitate and organisationally embedded. Institutional theory highlights regulatory, professional and normative pressures. Sociomaterial approaches examine how human and material elements become entangled in practice. Dynamic capability perspectives focus on sensing opportunities, seizing them and reconfiguring resources.

These lenses should not be treated as interchangeable labels. Each highlights different mechanisms and may imply different data requirements. A study of digital transformation should specify whether it examines adoption, implementation, use, capability development, performance, identity or power. It should also consider unintended consequences: work intensification, surveillance, exclusion, deskilling, new dependencies on vendors and the redistribution of discretion.

Illustrative Case: Predicting Customer Churn in a Digital Bank

Background

The following is a fictional illustrative case. Meridian, a rapidly growing digital bank, wished to reduce customer churn. It possessed transaction histories, mobile-application events, support-chat transcripts, marketing responses and account-closure records. Executives proposed an AI system that would rank customers by likelihood of leaving and trigger retention offers.

Situation and Research Problem

An initial analysis found that customers with fewer logins and more support contacts were more likely to close accounts. A highly accurate model was developed using several hundred variables. Management planned to deploy it immediately. A doctoral research team was asked to evaluate whether the system would improve retention fairly and sustainably.

Analysis

The team first examined provenance. The outcome variable, “churn,” represented account closure, but did not include customers who became inactive without closing accounts. Support contacts were recorded more completely for customers using a particular channel, creating coverage differences. Some variables were generated after a customer had already expressed an intention to leave, producing temporal leakage.

The team then investigated representation. Customers with limited connectivity and those using accessibility services generated fewer app events, so apparent low engagement partly reflected measurement differences. Historical retention campaigns had targeted particular neighbourhoods; campaign-response variables therefore acted as proxies for prior managerial decisions. A sentiment model trained on general consumer reviews performed poorly on banking complaints, especially where customers used formal or multilingual language.

The research design combined temporal model validation, subgroup error analysis, interviews with service employees and an examination of retention-offer consequences. Rather than asking only whether the model predicted closure, the team asked whether acting on predictions improved outcomes compared with a clearly defined intervention and whether harms were distributed unevenly.

Approach and Solution

The researchers removed post-outcome variables, created a time-based test set and compared the complex model with interpretable baselines. They reported calibration and false-negative rates by relevant customer groups, while avoiding unnecessary publication of identifying attributes. A human review process was designed for high-impact decisions, and customers were given a channel to challenge erroneous information. Retention offers were tested through a controlled pilot, with safeguards against disadvantaging customers who were not targeted.

The team also reframed the organisational problem. Interviews showed that many complaints concerned confusing fees and delayed support, issues that incentives to individual customers would not solve. The bank therefore combined targeted assistance with process improvements and clearer communication.

Outcome

The complex model’s predictive advantage over the baseline was modest after leakage was removed. The pilot improved retention among some customers, but indiscriminate offers increased costs and did not address the main service failures. The final programme used the model as one input into service prioritisation rather than as an autonomous decision-maker. Monitoring continued after deployment because customer behaviour and platform features changed.

Learning Points

The case demonstrates that predictive accuracy is not the same as organisational value. Provenance, temporal validity, fairness, intervention design and institutional context all affected the usefulness of the system. It also shows why digital transformation should be studied as a process of organisational learning and redesign, not as the installation of an algorithm.

Challenges and Common Errors

A frequent error is equating scale with quality. Researchers may also mistake accessibility for representativeness: data available through an API are shaped by the platform’s permissions and commercial interests. Another error is changing the research question after inspecting thousands of variables without documenting exploratory decisions. This increases the risk of false discoveries and retrospective storytelling.

Other recurring problems include ignoring dependence among observations, using random train–test splits for time series, reporting only average performance, treating missingness as harmless, deleting outliers without theoretical justification, and interpreting feature importance as causation. In digital ethnography, researchers may overlook the relationship between public visibility and participant vulnerability. In network studies, conclusions may change when boundaries or tie definitions change.

Doctoral work should also resist technological determinism—the assumption that a technology inevitably produces a particular organisational outcome. Adoption is mediated by power, resources, routines, professional judgement, regulation and resistance. A strong study makes these mechanisms visible and considers alternative explanations.

Current Developments and Best Practices

Research is increasingly shaped by foundation models, multimodal systems, automated experimentation, privacy-enhancing technologies, federated analysis, synthetic data and real-time organisational dashboards. These developments create opportunities but also raise questions about model drift, energy use, intellectual property, data sovereignty and concentration of infrastructure among a small number of providers.

Good practice includes aligning the dataset with the claim, maintaining a provenance and decision log, pre-registering confirmatory analyses where feasible, separating exploratory from confirmatory work, validating models in the deployment context, reporting subgroup and temporal performance, and involving domain experts and affected stakeholders. Researchers should preserve a humanly interpretable account of the research process even when sophisticated tools are used.

The most defensible digital research is neither technologically celebratory nor technologically sceptical. It treats computational methods as valuable instruments whose meaning depends on theory, design, evidence quality and governance. Such research can reveal patterns that would otherwise remain invisible while remaining honest about what digital traces cannot show.

Chapter Summary

Digital research uses traces generated through platforms, transactions, communications and organisational systems, but traces become evidence only through valid conceptualisation and contextual interpretation. Big data offers scale and variety while creating problems of coverage, provenance, measurement and privacy. Text analysis, network analysis and machine learning are useful when matched to appropriate questions and validated against the realities of the domain. Prediction should not be confused with causal explanation, and explainability should support scrutiny rather than provide a superficial justification.

Algorithmic bias can arise throughout the data and modelling pipeline. Fairness requires explicit normative choices, subgroup evaluation and governance. Reproducibility involves documented data transformations, code, assumptions and limitations, while responsible openness must protect confidentiality. Generative AI can support research tasks but requires verification, disclosure, secure handling and continuing human accountability. Finally, digital transformation is socio-technical: organisational outcomes emerge from interactions among technologies, people, routines, institutions and power. The doctoral researcher’s task is to connect computational capability with rigorous design and defensible inference.

Key Terms

Digital trace: A record of activity produced through interaction with a digital system.

Big data: Data characterised by combinations of volume, velocity, variety, veracity and value.

Platform data: Data generated and structured by a digital platform’s users, interfaces and governance rules.

Data provenance: The documented origin, custody, transformation and use history of data.

Unit of analysis: The entity about which a research claim is made.

Data leakage: The inclusion of information in model development that would not be available at the time of prediction.

Overfitting: Learning sample-specific noise rather than patterns that generalise.

Text mining: Computational processing of textual material to identify patterns or classifications.

Sentiment analysis: Estimation of evaluative or emotional orientation in text.

Network analysis: Study of entities and the relationships connecting them.

Centrality: A family of measures describing the structural position of a node in a network.

Machine learning: Methods that learn patterns from data for prediction, classification or representation.

Explainability: The extent to which a system’s outputs or behaviour can be understood and scrutinised.

Algorithmic bias: Systematic and consequential distortion arising in data, models, decisions or deployment.

Reproducibility: The ability to trace and recreate analyses or assess conclusions using new evidence.

Generative AI: AI capable of producing text, code, images or other content in response to instructions.

Socio-technical system: An arrangement in which technologies and social practices mutually shape outcomes.

Technological determinism: The assumption that technology independently and inevitably determines social or organisational outcomes.

Data minimisation: Collecting and retaining only data necessary for a legitimate purpose.

Model drift: Decline or change in model performance as data, behaviour or context changes.

Review Questions

Short Answer Questions

1. Why does a large dataset not automatically provide valid evidence?

2. Distinguish the unit of observation from the unit of analysis.

3. What is data provenance, and why is it important?

4. Explain the difference between prediction and causal explanation.

5. Why can public digital data still raise ethical concerns?

Long Answer Questions

1. Explain how platform governance shapes the digital traces available to researchers.

2. Compare the methodological uses and limitations of text analysis, network analysis and machine learning in business research.

3. Discuss the main sources of algorithmic bias and the implications of competing fairness criteria.

4. Develop a responsible protocol for using generative AI in a doctoral research project.

5. Explain why digital transformation should be analysed as a socio-technical process rather than a technological event.

Analytical / Application Questions

1. A researcher uses employee login frequency as a measure of engagement. Critically assess the construct validity, possible biases and additional evidence required.

2. Design a validation strategy for a model predicting customer churn when the organisation’s data and customer behaviour change over time.

3. Using the Meridian case, explain how a technically accurate model could still produce poor organisational and ethical outcomes.

22. Multiple Choice Questions with Answers

Question 1

Which statement best explains why digital traces should not automatically be treated as valid evidence?

A. Digital traces are always too small for statistical analysis.

B. They may reflect platform rules and recording practices rather than the underlying phenomenon.

C. They cannot be combined with qualitative evidence.

D. They are never produced without researcher intervention.

Correct Answer: B. They may reflect platform rules and recording practices rather than the underlying phenomenon.

Question 2

A researcher uses customer records from 2022 to predict account closure in 2023, but includes a variable recorded only after customers requested closure. What problem is most clearly present?

A. Network dependence

B. Construct validity

C. Data leakage

D. Underfitting

Correct Answer: C. Data leakage

Question 3

Which practice most directly supports computational reproducibility?

A. Reporting only the final model accuracy

B. Providing documented code, transformations and software versions

C. Removing all missing observations without explanation

D. Using the largest available dataset

Correct Answer: B. Providing documented code, transformations and software versions

Question 4

Why should sentiment analysis of employee complaints be validated with domain-relevant examples?

A. Sentiment models are incapable of processing written language.

B. Meaning may depend on sarcasm, specialised language and context.

C. Validation eliminates the need for human interpretation.

D. Employee complaints are always statistically independent.

Correct Answer: B. Meaning may depend on sarcasm, specialised language and context.

Question 5

In a network study, a highly central employee should be interpreted as powerful only when:

A. the employee has the largest number of messages in every dataset.

B. centrality is connected to a theoretically justified meaning of influence or coordination.

C. the network contains at least one hundred nodes.

D. the employee belongs to the oldest department.

Correct Answer: B. centrality is connected to a theoretically justified meaning of influence or coordination.

Question 6

Which statement correctly distinguishes prediction from causal explanation?

A. A predictive model always identifies the intervention that will change an outcome.

B. Causal explanation is unnecessary if predictive accuracy is high.

C. Prediction estimates outcomes, whereas causal explanation requires assumptions and design suited to causal inference.

D. Prediction can be conducted only with qualitative data.

Correct Answer: C. Prediction estimates outcomes, whereas causal explanation requires assumptions and design suited to causal inference.

Question 7

Deleting demographic variables from an algorithm necessarily removes algorithmic bias. Which assessment is most accurate?

A. True, because demographic variables are the only source of bias.

B. True, provided the model is complex enough.

C. False, because proxy variables and historical inequalities may remain.

D. False, because algorithms cannot be evaluated for fairness.

Correct Answer: C. False, because proxy variables and historical inequalities may remain.

Question 8

Which is the most responsible use of generative AI in a literature review?

A. Accepting generated citations without verification

B. Uploading confidential participant transcripts to an unapproved system

C. Using it for preliminary support while verifying claims against authoritative sources

D. Allowing it to determine the final theoretical framework without disclosure

Correct Answer: C. Using it for preliminary support while verifying claims against authoritative sources

Question 9

What does a socio-technical perspective on digital transformation emphasise?

A. Technology inevitably determines organisational outcomes.

B. Outcomes emerge from interactions among technology, people, routines and institutions.

C. Organisational culture becomes irrelevant after implementation.

D. Only technical performance should be measured.

Correct Answer: B. Outcomes emerge from interactions among technology, people, routines and institutions.

Question 10

In the illustrative banking case, why did the research team use a time-based test set?

A. To ensure that future information did not enter evaluation of past prediction.

B. To increase the number of variables available to the model.

C. To eliminate the need for subgroup analysis.

D. To guarantee that the model would be causal.

Correct Answer: A. To ensure that future information did not enter evaluation of past prediction.