Truth, Beauty, Justice, and the Data Scientist’s Dilemma

TL;DR for operators

The useful question is not whether AI will “replace data scientists”. That framing is wonderfully dramatic and operationally lazy.

Timpone and Yang’s paper, AI, Humans, and Data Science: Optimizing Roles Across Workflows and the Workforce, gives a better mechanism: allocate human and AI work by asking what kind of quality each workflow stage needs.¹ Early planning needs creative breadth and problem definition. Execution needs accurate, valid, and ethically defensible data and modelling. Activation needs contextual interpretation, stakeholder judgement, and responsible action.

That leads to a practical division of labour. Generative AI can widen ideation, draft research plans, help structure surveys, generate code, prepare data, run models, and summarise findings. Agentic AI can orchestrate parts of the execution pipeline when goals, data boundaries, methods, and evaluation criteria are explicit. Human data scientists remain responsible for reducing ambiguity before automation begins and evaluating whether the output is true, interpretable, fair, compliant, and useful enough to act on.

For executives, the immediate implication is simple: automate the execution layer, not the accountability layer. Use AI to increase analytic throughput. Do not confuse throughput with judgment. A dashboard generated in seconds can still be wrong in six different ways, now merely at enterprise scale. Progress, apparently.

The paper’s most underappreciated warning is about the talent pipeline. If AI absorbs the junior work where analysts normally learn data wrangling, modelling assumptions, measurement quality, and stakeholder communication, companies may later discover that they have fewer senior people capable of supervising the machines. The spreadsheet did not eliminate financial judgment. It did, however, create many confident errors. Agentic analytics may repeat that story with better syntax.

The machine can run the workflow; it cannot own the purpose

A familiar business scene: a manager wants answers by Friday. The data team has messy inputs, a vague objective, and three stakeholders who mean different things by “customer value”. An AI agent can now ingest files, write Python, run models, generate charts, and produce a polished summary. It may even do this before the second meeting invite is accepted.

This is exactly why the danger is no longer that AI cannot do enough. The danger is that it can do a great deal before anyone has clarified what should be done.

Timpone and Yang’s central contribution is a role-allocation framework for AI and human data scientists across the data science workflow. They distinguish analytic AI, generative AI, and agentic AI, then map these capabilities against planning, execution, and activation. The paper is not an empirical benchmark. There are no model leaderboards, ablations, or performance curves to worship. Its figures and tables are organising devices: Figure 1 frames human-machine collaboration; Figure 3 defines the workflow; Tables 1 and 5 summarise recommended actor balance; Tables 2 to 4 provide evaluation checklists for sample representativeness, measurement quality, and modelling quality.

That matters because the paper’s evidence is conceptual and synthetic rather than experimental. Its value is not “AI achieves X% on task Y”. Its value is a mechanism for deciding when automation is appropriate, when it needs supervision, and when human leadership remains structurally necessary.

The mechanism has three moving parts:

This is the paper’s real move. It does not ask whether AI is “smart”. It asks what kind of quality the task requires and whether the task environment gives AI enough structure to proceed safely.

Planning is not just prompt generation with nicer stationery

The paper divides the workflow into three phases: planning, execution, and activation. Planning includes “ideate and define” and “design and plan”. This is where many organisations will be tempted to let generative AI take over early because it can produce research questions, survey drafts, hypotheses, and project plans at satisfying speed.

The authors do not reject that use. In fact, they argue that generative AI can be valuable in ideation because the relevant quality criterion is largely Beauty: breadth, richness, fertility, surprise, and interpretability. AI can broaden the option set. It can put strange concepts together. It can produce drafts that humans would not have written first. This is useful.

But planning is also where the wrong question quietly poisons the entire project. A beautifully articulated wrong objective is still wrong, only now easier to approve.

The authors’ recommended balance is therefore “AI complements humans”. Humans should lead the framing, while AI helps expand and formalise the possibilities. Once the work moves from ideation to design, the primary criterion shifts toward Truth. The plan must specify the data, methods, assumptions, intended audience, and ethical boundaries. At that point, a generative or agentic system may help write documentation, structure an analysis plan, or draft survey instruments, but it should not be treated as the owner of the research design.

The business lesson is sharper than “keep a human in the loop”. That phrase has been used so often it now functions as a scented candle for governance decks. The practical version is: humans must define the loop before AI enters it.

A useful planning review should ask:

This is not bureaucracy. It is pre-automation risk reduction.

Execution is where AI looks strongest—and where the trap is easiest to miss

The execution phase includes gathering and processing data, then conducting analysis. This is the part of the workflow where AI agents are most visibly useful. They can write code, call tools, clean datasets, run multiple models, generate synthetic data, extract information from text, compare outputs, and produce charts.

The paper agrees that AI can lead much of the concrete execution work once decision criteria are clear. This is the section where the authors are most automation-friendly. They explicitly argue that AI agents can perform the bulk of execution activities when humans have already clarified goals, context, output expectations, method boundaries, compliance constraints, and ethical rules.

The word “already” is doing a lot of work.

Execution often looks structured because it has files, variables, models, and code. That appearance is misleading. The authors use VUCA to show why data and modelling decisions are rarely as tidy as the interface suggests.

Volatility appears when data quality changes across time, groups, collection modes, or business conditions. Uncertainty appears when multiple imperfect designs or modelling choices are plausible. Complexity appears when variables, constructs, confounds, and indirect relationships are entangled. Ambiguity appears when the same phrase—say, “predictive model”—could mean a forecast, recommender system, classification model, or LLM-based inference workflow.

An AI agent can process ambiguity as a task. A human data scientist has to recognise ambiguity as a risk.

This is where the paper’s checklists matter. Tables 2 to 4 are not empirical tests; they are implementation safeguards. They ask whether the research team is diverse, whether the unit of analysis is correct, whether the sample represents the population of interest, whether meaningful heterogeneity exists, whether measures match theoretical concepts, whether proxies introduce measurement error, whether training data is adequate, whether the chosen method matches the theorised process, and whether the model remains valid as circumstances change.

Translated into business language: before letting an agent run the analysis, make sure the agent is not optimising inside a badly specified box.

The old point-and-click problem has acquired a natural-language interface

One of the paper’s more useful analogies is to the earlier spread of statistical software. SAS, SPSS, Stata, and similar tools made analysis faster and more accessible. They also made it easier for users to run methods they did not understand. Point, click, export table, misinterpret interaction term, present confidently. A classic.

The authors call this risk “blindness-by-design”: the tool hides the assumptions well enough that the user stops noticing they exist. Agentic AI intensifies the problem because the interface is no longer a menu of statistical options. It is a conversational box that accepts vague ambition and returns executable artefacts.

That makes misuse smoother. Instead of selecting the wrong model from a dropdown, the user can ask a broad question and let the agent decide what “analysis” means. The agent may generate code, choose variables, transform data, handle missingness, interpret coefficients, and summarise implications. Each step may be plausible. The whole pipeline may still be wrong.

The real risk is not that AI produces nonsense. Nonsense is often detectable. The higher risk is competent-looking analysis built on unstated assumptions.

For operators, the control point is not merely output review. It is workflow review:

This is a better operating model than the generic “human approves final answer”. By the time the final answer exists, many invisible choices have already shaped it.

Activation is where dashboards become consequences

A data science project does not end when the model runs. It ends when someone acts on the insight.

The paper’s activation phase covers “create insights” and “activate insights”. Here the recommended actor balance shifts back toward humans. AI can help synthesise findings, explore heterogeneity, draft narratives, and generate communication artefacts. But the authors argue that final interpretation and real-world implementation require human leadership because Truth, Justice, and Beauty all become active at once.

This is where many analytics transformations fail quietly. The model may be technically sound, but the action attached to it may be ethically careless, politically naive, or simply impractical. A churn model can identify high-risk customers. It cannot decide whether the retention intervention is fair, brand-safe, legally defensible, or operationally possible. A hiring model can rank candidates. It cannot by itself determine whether the ranking embeds historical exclusion. A pricing model can segment willingness to pay. It cannot determine whether the business wants to be the sort of company that exploits every vulnerability it detects. Annoying, yes. Also called management.

The paper’s activation argument is especially relevant for AI agents because agents are built to act. That is their selling point. But the more directly an AI system can trigger downstream actions—send offers, change prices, generate compliance summaries, prioritise customers, alter recommendations—the more important it becomes to separate analytic confidence from decision authority.

A narrow, low-stakes action may be fully automated. The paper gives the kind of example where a granular website design adjustment might be safe to automate. But broader business decisions involve stakeholder interpretation, organisational politics, change management, fairness, and accountability. AI can support that work. It should not own it.

The workforce problem is not replacement; it is apprenticeship collapse

The paper’s most strategically uncomfortable argument concerns the data science workforce.

If AI automates junior tasks—ETL, data wrangling, standard modelling, templated dashboards, routine reporting—companies may reduce entry-level hiring. In the short term, that looks efficient. In the long term, it can damage the pipeline of senior data scientists. Senior judgment does not appear magically after someone has read three governance memos and watched an LLM generate SQL. It is built through repeated exposure to messy data, bad assumptions, stakeholder confusion, failed models, ambiguous results, and the slow humiliation of discovering that the first answer was wrong.

This is a classic automation trap. Remove the “easy” work, and you may remove the training ground. Then, several years later, the organisation wonders why it has plenty of tools but too few people who understand when not to trust them.

The authors connect this to the changing nature of the profession. Demand may decline for routine, structured, predictable, automatable work. Demand should rise for problem framing, analytical strategy, conceptual understanding, critical evaluation of AI output, domain expertise, ethical reasoning, and communication. In other words, the job becomes less about manually producing every analytic artefact and more about designing, supervising, validating, and activating AI-augmented workflows.

That shift sounds elegant. It is also hard to staff if nobody trains the next generation.

Businesses therefore need an apprenticeship model for AI-era data science. Junior analysts should not simply be removed from the workflow. They should be moved into supervised AI-augmented work where they learn to inspect assumptions, challenge outputs, compare methods, audit data quality, and explain results. The company that automates junior learning to save money may eventually pay senior contractor rates to repair the consequences. Markets do enjoy comedy.

What the paper directly shows, and what Cognaptus infers

Because this paper is a conceptual framework rather than an experimental study, it is important not to overclaim. Its figures and tables organise existing theory, prior literature, and professional judgment. They do not prove that one exact staffing model outperforms another.

Here is the clean separation:

This distinction matters. The paper should not be read as a veto on AI agents in analytics. It is closer to an organisational design memo: automate where structure is high, judgment burden is low, and validation is possible; keep humans responsible where ambiguity, ethics, context, and action matter.

A practical operating model for AI-augmented analytics

The paper’s mechanism can be turned into a simple operating model.

First, classify the workflow stage. Is the task planning, execution, or activation? A request such as “analyse customer churn” is not one task. It includes defining churn, selecting the population, identifying relevant signals, choosing methods, modelling, interpreting drivers, designing interventions, and deciding what action is acceptable.

Second, classify the quality criterion. Does success mainly require Truth, Beauty, Justice, or some combination? Data cleaning and model comparison lean heavily toward Truth. Ideation and synthesis require Beauty. Sampling, privacy, bias, and deployment require Justice. Most important projects require all three, which is inconvenient but traditional.

Third, assess VUCA. If the data is unstable, the objective contested, the variables entangled, or the interpretation ambiguous, increase human control. If the task is routine, low-stakes, well-documented, and easy to validate, AI can take more initiative.

Fourth, specify the actor balance before execution. Do not allow the agent to infer its own authority from a vague prompt. Define what it may do autonomously, what it must ask before doing, and what humans must approve.

A lightweight governance pattern might look like this:

The goal is not to slow analytics down. The goal is to avoid accelerating the wrong thing.

The boundary: this is a design lens, not a benchmark

The paper’s limitation is also its strength. It is broad, integrative, and practical, but it does not provide a new dataset, experimental comparison, causal estimate, or quantified performance gain. It synthesises frameworks: human-machine collaboration, Truth-Beauty-Justice, VUCA, and labour-economics thinking about task automation.

That means organisations should not quote it as proof that “humans must always approve all AI analytics” or, conversely, that “AI agents can safely lead execution everywhere”. The more precise reading is conditional: AI leadership is most defensible when humans have reduced VUCA, defined evaluation criteria, set ethical and compliance boundaries, and preserved review points that inspect assumptions rather than just outputs.

Another boundary is that the framework is easiest to apply in analytic functions mature enough to name their methods, data sources, stakeholders, and decision rights. In immature organisations, the first use of the framework may reveal that the data science workflow was never properly governed even before AI arrived. Awkward, but useful.

The data scientist’s dilemma is really a management dilemma

The paper’s title foregrounds AI, humans, and data science. But its deepest implication is managerial.

Data scientists are not valuable merely because they can write code, run regressions, tune models, or produce slides. AI will increasingly perform much of that work faster. Their value lies in connecting analytic machinery to valid questions, meaningful data, interpretable models, ethical boundaries, and decisions that survive contact with reality.

That also means executives cannot treat AI agents as a shortcut around analytic maturity. A company that lacks clear problem framing, data governance, measurement discipline, model validation, and accountability will not become analytically sophisticated by adding an agentic interface. It will simply automate its confusion.

The better future is not human versus machine. It is a workflow where machines transact, iterate, predict, draft, explore, and synthesise; humans define, judge, contextualise, validate, explain, and take responsibility. That division will shift as tools improve. It should shift. But it should shift deliberately, not because a demo looked impressive and the budget meeting was approaching.

Truth, Beauty, and Justice may sound a little grand for daily analytics work. Still, the labels are useful. Truth asks whether the answer is valid. Beauty asks whether the answer is intelligible and rich enough to matter. Justice asks whether acting on the answer is fair and responsible.

An AI agent can help produce all three. It cannot be accountable for all three. That remains, inconveniently, a human job.

Cognaptus: Automate the Present, Incubate the Future.