Data Scientists
15-2051.00Median wage $120,230/yr262,440 employed (US)Rank #38 of 923 scored · top 4% by substitution
Develop and implement a set of techniques or analytics applications to transform raw data into meaningful information using data-oriented programming languages and visualization software. Apply data mining, data modeling, natural language processing, and machine learning to extract and analyze information from large structured and unstructured datasets. Visualize, interpret, and report data findings. May create dynamic data reports.
Sub-scores
0–100 · band = confidence interval from rater disagreement
Substitution60
Exposure56
Augmentation88
Substitution — the headline: capability discounted by cost, barriers and adoption.
Exposure — technical capability alone, regardless of whether anyone deploys it.
Augmentation — how much AI assists without replacing. High here + moderate substitution = a changing job, not a disappearing one.
Tasks on the substitution scale
16 rated tasks, binned by substitution score.
Position among all scored occupations
Distribution of 923 occupation scores; the marker is this occupation.
Tasks with substitution ≥ 70
38%
Run 1.0.0-draft.1 · computed 2026-08-05 · rater panel: claude-sonnet-5, claude-haiku-4-5-20251001 · intervals span rater disagreement.
Why this score
The five weighted dimensions of the composite, averaged across this occupation's tasks (importance-weighted, panel mean). Exact weights and formulas: /api/v1/methodology.
Task automatabilityw 35%55
panel mean rating 3.2/5 → substitution pressure 55/100
Technical feasibility todayw 20%57
panel mean rating 3.3/5 → substitution pressure 57/100
Cost vs. human wagew 15%61
panel mean rating 3.4/5 → substitution pressure 61/100
Adoption barriersw 20%inverted — strong barriers lower the score70
panel mean rating 2.2/5 (barrier strength) → substitution pressure 70/100
Sector adoption velocityw 10%65
panel mean rating 3.6/5 → substitution pressure 65/100
Task breakdown (16 tasks)
Substitution pressure per task, weighted by O*NET importance in the composite. Expand a task for the full rater audit trail — every rating, every model, every rationale.
Compare models using statistical performance metrics, such as loss functions or proportion of explained variance.
90CI 80–100 · exposure 87 · augmentation 75 · click for rater detail
Compare models using statistical performance metrics, such as loss functions or proportion of explained variance.
90| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 5/5 | Data science teams, tech companies, finance, and professional services (high-digitization sectors) have already integrated automated metric computation into standard ML pipelines and experiment-tracking tools (MLflow, Weights & Biases, etc.). Adoption is deep and fast. |
| Sector adoption velocity | claude-sonnet-5 | 4/5 | Data science and ML engineering teams have rapidly adopted automated experiment tracking and AutoML tools (e.g., MLflow, AutoML platforms, Copilot-assisted notebooks) as standard practice in tech and analytics sectors. |
| Augmentation potential | claude-haiku-4-5-20251001 | 3/5 | While the core metric computation is fully automatable, AI can assist by recommending which metrics to use or visualizing comparisons, but the task itself is mechanistic and does not require ongoing human judgment once metrics are chosen. |
| Augmentation potential | claude-sonnet-5 | 5/5 | AI coding assistants and AutoML tools substantially speed up metric computation, visualization, and comparison, letting data scientists focus on interpretation and strategic model selection while staying in the loop. |
| Task automatability | claude-haiku-4-5-20251001 | 5/5 | Comparing models using statistical performance metrics is highly automatable. Current AI systems and standard software can compute loss functions, R², RMSE, and other metrics end-to-end with 100% reproducibility, vastly exceeding the 50% time-saving threshold. A data scientist typically spends minutes on this task; automation reduces it to seconds. |
| Task automatability | claude-sonnet-5 | 4/5 | Computing and comparing metrics like AUC, RMSE, or explained variance across models is a well-defined, computational task that AI code assistants and automated ML tools can execute end-to-end with proper setup, though interpretation and decision-making retain some human involvement. |
| Adoption barriers | claude-haiku-4-5-20251001 | 1/5 | No licensing, regulatory, or legal barriers prevent automation of metric computation. Organizations face no mandate to have a human manually calculate loss or variance metrics; it is purely a technical, commoditized task. |
| Adoption barriers | claude-sonnet-5 | 1/5 | There is no licensing, regulatory, or liability barrier to using automated tools for statistical model comparison; it's a standard technical practice with no legal gatekeeping. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 5/5 | The computational cost of running metric comparisons is negligible—often cents or free per model evaluation. The human loaded wage for a data scientist ($80–150k annually) means even brief time per metric comparison exceeds the AI cost by orders of magnitude. |
| Cost vs. human wage | claude-sonnet-5 | 4/5 | Running metric computations via automated pipelines or AI-assisted scripts costs a fraction of a data scientist's hourly wage, though initial pipeline setup and validation retain some human cost. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 5/5 | Mature, deployed products (scikit-learn, TensorFlow, PyTorch, statistical software) perform this task reliably at scale in production environments across thousands of organizations. Metric computation is deterministic and well-established. |
| Technical feasibility today | claude-sonnet-5 | 4/5 | AutoML platforms and code-generation tools (e.g., AI coding assistants integrated with notebooks) already reliably compute and tabulate performance metrics across candidate models in production ML workflows today. |
Create graphs, charts, or other visualizations to convey the results of data analysis using specialized software.
86CI 75–97 · exposure 87 · augmentation 100 · click for rater detail
Create graphs, charts, or other visualizations to convey the results of data analysis using specialized software.
86| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 4/5 | Data science teams and analytics-heavy organizations (tech, finance, consulting) are rapidly adopting AI-assisted visualization tools in production; adoption is deep in digitized sectors and accelerating. |
| Sector adoption velocity | claude-sonnet-5 | 4/5 | Data science and analytics functions are in fast-adopting professional/tech sectors, with AI-assisted visualization tools already embedded in mainstream BI and notebook platforms. |
| Augmentation potential | claude-haiku-4-5-20251001 | 5/5 | AI dramatically augments human productivity by instantly generating multiple visualization options, iterating on design, and auto-formatting data; humans remain in control of interpretation and strategic choices while working much faster. |
| Augmentation potential | claude-sonnet-5 | 5/5 | AI dramatically speeds up chart creation, suggests visualization types, and auto-generates code, letting data scientists iterate much faster while retaining control over final analytical framing. |
| Task automatability | claude-haiku-4-5-20251001 | 5/5 | Current AI systems can generate visualizations end-to-end from data and specifications using tools like Python (matplotlib, seaborn, plotly) with LLM guidance, or dedicated visualization APIs. This routinely achieves >50% time savings compared to manual chart construction. |
| Task automatability | claude-sonnet-5 | 4/5 | Modern AI tools (code interpreters, BI copilots, notebook assistants) can generate charts and visualizations from data with minimal prompting, saving significant time on routine visualization work, though customization and storytelling judgment still require human input. |
| Adoption barriers | claude-haiku-4-5-20251001 | 2/5 | Minimal legal or regulatory barriers exist; visualization generation is not a licensed activity. However, some organizational friction remains around data governance, chart validation, and preference for human review before stakeholder presentation. |
| Adoption barriers | claude-sonnet-5 | 1/5 | No licensing, regulatory, or liability barriers restrict AI-assisted visualization creation; it's a purely technical task with no legal sign-off requirement. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 5/5 | API-based visualization generation and LLM-assisted code costs are negligible per chart compared to the loaded hourly wage of a data scientist ($60–150/hour), making AI orders of magnitude cheaper for this task. |
| Cost vs. human wage | claude-sonnet-5 | 4/5 | Generating a chart via AI costs fractions of a cent in compute versus substantial analyst time, though integration and review add some overhead reducing the full order-of-magnitude gain. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 5/5 | Mature products including ChatGPT with code execution, GitHub Copilot, and specialized data visualization platforms reliably generate charts and graphs in production environments for data scientists daily. |
| Technical feasibility today | claude-sonnet-5 | 4/5 | Products like ChatGPT Code Interpreter, GitHub Copilot, Tableau's AI features, and Power BI Copilot reliably generate visualizations in production today, though edge cases and highly customized/publication-quality charts still need human refinement. |
Apply feature selection algorithms to models predicting outcomes of interest, such as sales, attrition, and healthcare use.
82CI 75–89 · exposure 80 · augmentation 100 · click for rater detail
Apply feature selection algorithms to models predicting outcomes of interest, such as sales, attrition, and healthcare use.
82| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 4/5 | Cloud ML platforms, AutoML vendors, and internal ML teams in tech, finance, and enterprise software have rapidly adopted automated feature selection as a standard preprocessing step. Production use is common in digitized sectors, though smaller and non-tech organizations lag. |
| Sector adoption velocity | claude-sonnet-5 | 4/5 | Data science and analytics teams in tech, finance, and healthcare have rapidly adopted AutoML and automated feature engineering tools as standard practice in production pipelines. |
| Augmentation potential | claude-haiku-4-5-20251001 | 5/5 | Feature selection tools significantly augment data scientists by rapidly exploring high-dimensional spaces, surfacing candidate features, and automating comparative evaluation. This frees the scientist to focus on domain interpretation, business validation, and model refinement—a classic productivity multiplier while keeping human judgment central. |
| Augmentation potential | claude-sonnet-5 | 5/5 | AI substantially speeds up and improves feature selection by testing many algorithms/combinations quickly, letting data scientists focus on interpretation, validation, and business context. |
| Task automatability | claude-haiku-4-5-20251001 | 4/5 | Feature selection algorithms are already automated in scikit-learn, pandas, and specialized ML libraries. An AI system can run filter, wrapper, and embedded methods end-to-end, evaluate them against held-out validation sets, and produce ranked feature lists with minimal human setup. The remaining manual steps (domain validation, business interpretation) prevent a full 5, but the technical core easily achieves >50% time savings at equal quality. |
| Task automatability | claude-sonnet-5 | 4/5 | Feature selection (e.g., LASSO, recursive feature elimination, tree-based importance, automated feature engineering) is well-supported by libraries and AutoML tools that can run these algorithms with minimal human coding, saving substantial time versus manual implementation. |
| Adoption barriers | claude-haiku-4-5-20251001 | 2/5 | Few hard regulatory or licensing barriers exist for automated feature selection itself. Organizations do impose internal governance (model validation, stakeholder sign-off), but these are procedural oversight rather than legal prohibition. A data scientist's judgment on business relevance remains expected, creating some organizational friction but no hard blocker. |
| Adoption barriers | claude-sonnet-5 | 1/5 | There is no licensing, regulatory, or human-sign-off requirement specifically for feature selection steps within a modeling pipeline. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 5/5 | Running feature selection algorithms via cloud ML platforms or open-source libraries costs pennies per model iteration. The marginal cost of inference and feature evaluation is orders of magnitude below the fully-loaded wage of a data scientist performing manual feature engineering and validation. |
| Cost vs. human wage | claude-sonnet-5 | 4/5 | Running automated feature selection algorithms is computationally cheap compared to a data scientist manually iterating through feature subsets, though some compute and engineering setup cost remains. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 5/5 | Mature, deployed products (H2O AutoML, Amazon SageMaker, Dataiku, DataRobot) routinely perform automated feature selection in production. These systems integrate feature importance estimation, cross-validated selection, and reporting at scale in real organizations. |
| Technical feasibility today | claude-sonnet-5 | 4/5 | Production tools like scikit-learn pipelines, AutoML platforms (DataRobot, H2O, Azure AutoML) reliably perform automated feature selection at scale across industries today, though final validation and domain interpretation still involve humans. |
Clean and manipulate raw data using statistical software.
81CI 75–86 · exposure 75 · augmentation 100 · click for rater detail
Clean and manipulate raw data using statistical software.
81| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 5/5 | Tech-forward industries (finance, tech, e-commerce, analytics firms) are rapidly adopting AI-assisted and automated data cleaning. Production deployments of AI agents for ETL and data preparation are common in high-digitization sectors. |
| Sector adoption velocity | claude-sonnet-5 | 4/5 | Data science and analytics teams in tech, finance, and professional services have rapidly adopted AI coding assistants and automated data pipeline tools in production workflows. |
| Augmentation potential | claude-haiku-4-5-20251001 | 5/5 | AI excels at suggesting transformations, generating cleaning code, and identifying patterns; data scientists using generative AI assistants report substantial productivity gains while remaining in control of validation and quality decisions. |
| Augmentation potential | claude-sonnet-5 | 5/5 | AI coding assistants substantially speed up writing and debugging cleaning scripts, suggest transformations, and catch anomalies, making this one of the clearest cases of productivity augmentation for data scientists. |
| Task automatability | claude-haiku-4-5-20251001 | 4/5 | AI tools can automate most data cleaning and manipulation tasks—handling missing values, outlier detection, format standardization, and basic transformations—with significant time savings. However, domain-specific anomalies and nuanced data quality decisions often require human judgment, preventing a perfect 5. |
| Task automatability | claude-sonnet-5 | 4/5 | Data cleaning and manipulation (handling missing values, transformations, merges, formatting) is highly automatable with LLM-based coding assistants and libraries like pandas, especially for structured, well-defined datasets, though edge cases and domain-specific judgment still require oversight. |
| Adoption barriers | claude-haiku-4-5-20251001 | 2/5 | Few regulatory or legal barriers protect this task; it is typically an internal, non-customer-facing operation. Minimal licensing or liability constraints exist, though some organizations may impose internal governance on data handling. |
| Adoption barriers | claude-sonnet-5 | 1/5 | No licensing, regulatory, or liability barriers restrict automating data cleaning; it's a routine technical task with no legal requirement for human sign-off. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 5/5 | The cost of cloud-based data cleaning infrastructure and AI-assisted code generation is orders of magnitude cheaper than human labor, especially at scale, even accounting for oversight and integration overhead. |
| Cost vs. human wage | claude-sonnet-5 | 4/5 | AI coding assistants cost a small fraction of a data scientist's hourly wage and can draft/execute cleaning code in seconds, though human validation time slightly reduces the savings. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 4/5 | Deployed products (pandas libraries, dbt, data preparation tools like Trifacta, and generative AI code assistants) reliably perform these tasks in production environments. Most have material limitations on complex, domain-specific logic, keeping the rating below 5. |
| Technical feasibility today | claude-sonnet-5 | 4/5 | Production tools like GitHub Copilot, code interpreters, and AI-augmented notebooks (e.g., Jupyter AI, DataRobot) reliably generate cleaning scripts and pipelines today, though complex or messy real-world data still needs human review. |
Write new functions or applications in programming languages to conduct analyses.
81CI 75–86 · exposure 75 · augmentation 100 · click for rater detail
Write new functions or applications in programming languages to conduct analyses.
81| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 4/5 | Tech-forward sectors (finance, SaaS, big tech, consulting) are rapidly deploying AI coding assistants and have normalized their use in production; adoption among data science and engineering teams is already substantial and accelerating, though smaller firms and non-tech sectors lag. |
| Sector adoption velocity | claude-sonnet-5 | 5/5 | Software/data science is among the fastest-adopting domains for AI coding tools, with widespread production use of copilots and agentic coding assistants. |
| Augmentation potential | claude-haiku-4-5-20251001 | 5/5 | AI code generation dramatically amplifies data scientist productivity by handling boilerplate, standard transformations, and common algorithm implementations, freeing humans to focus on problem formulation, validation, and domain interpretation. This is one of the most effective human-AI collaboration patterns in professional work today. |
| Augmentation potential | claude-sonnet-5 | 5/5 | AI coding assistants dramatically speed up writing, debugging, and refactoring functions while the data scientist retains control over design and validation. |
| Task automatability | claude-haiku-4-5-20251001 | 4/5 | Current AI coding assistants (GitHub Copilot, Claude, ChatGPT) can generate substantial portions of analysis functions in Python, R, and SQL with high quality, and can often produce complete, working code for standard statistical or data manipulation tasks. However, nuanced domain logic, integration with complex existing codebases, and requirement translation still frequently require human review and iteration, preventing a full 5 rating. |
| Task automatability | claude-sonnet-5 | 4/5 | Coding assistants can generate substantial portions of analytical scripts and functions with significant time savings, though complex domain-specific logic still needs human review and iteration. |
| Adoption barriers | claude-haiku-4-5-20251001 | 2/5 | Data scientists must often validate code quality and correctness before deployment, and some organizations have policies around code review. However, no legal licensing requirement, liability cap, or regulatory mandate mandates human authorship of code; adoption is primarily a workplace and technical-debt concern, not a hard barrier. |
| Adoption barriers | claude-sonnet-5 | 1/5 | No licensing, liability, or regulatory requirements restrict who writes analytical code; organizations freely adopt AI-assisted coding tools. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 5/5 | API inference costs for code generation are negligible per function (pennies to fractions of cents), while developer time to write equivalent code from scratch costs $50–150+ per hour loaded; even accounting for review overhead, AI-assisted code generation is one to two orders of magnitude cheaper than unassisted human coding. |
| Cost vs. human wage | claude-sonnet-5 | 4/5 | AI coding subscriptions cost a small fraction of a data scientist's hourly wage, though human oversight and debugging time add to the effective cost. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 4/5 | Deployed products like GitHub Copilot and LLM-based code assistants are actively used in production by data scientists and engineering teams, with measurable time savings on code generation. Error rates remain notable (syntax, logic flaws), and oversight is standard practice, but the tooling is mature enough for broad organizational use. |
| Technical feasibility today | claude-sonnet-5 | 4/5 | Tools like GitHub Copilot, Cursor, and ChatGPT are deployed at scale in production coding workflows for data analysis tasks, though non-trivial error rates persist for complex or novel functions. |
Analyze, manipulate, or process large sets of data using statistical software.
75CI 75–75 · exposure 75 · augmentation 100 · click for rater detail
Analyze, manipulate, or process large sets of data using statistical software.
75| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 4/5 | Information, finance, and tech sectors are rapidly adopting AI-assisted and AI-driven statistical workflows; ChatGPT and GitHub Copilot adoption among data scientists is high and accelerating. Adoption is measured and deep in digitized industries, though slower in highly regulated or non-tech sectors. |
| Sector adoption velocity | claude-sonnet-5 | 4/5 | Tech, finance, and professional services sectors employing data scientists show fast, deep adoption of AI coding assistants and copilots as part of standard workflows. |
| Augmentation potential | claude-haiku-4-5-20251001 | 5/5 | AI robustly augments data science productivity by generating code templates, suggesting transformations, automating routine statistical tasks, and accelerating exploratory analysis while the human validates, interprets, and directs. This is one of the highest-impact augmentation use cases in current practice. |
| Augmentation potential | claude-sonnet-5 | 5/5 | AI dramatically speeds up writing, debugging, and documenting statistical code, and helps interpret outputs, while the data scientist retains responsibility for methodology and interpretation. |
| Task automatability | claude-haiku-4-5-20251001 | 4/5 | AI can automate significant portions of data manipulation, statistical analysis, and exploratory workflows using tools like automated machine learning platforms and code-generation systems (e.g., ChatGPT, GitHub Copilot for Python/SQL), achieving substantial time savings. However, domain expertise in interpreting results, handling edge cases, and ensuring statistical validity typically still requires human oversight, preventing a full 5. |
| Task automatability | claude-sonnet-5 | 4/5 | LLM-based coding assistants and agents can write and execute statistical analysis code (R/Python/SQL) achieving significant time savings on data cleaning, transformation, and standard statistical procedures, though complex custom analyses still need human framing. |
| Adoption barriers | claude-haiku-4-5-20251001 | 2/5 | Few hard barriers exist: no licensing requirement mandates human data scientists, and liability is typically absorbed by organizations employing the AI. However, organizational inertia, preference for explainability/oversight by humans, and sector-specific governance (finance, healthcare) create moderate friction against full automation. |
| Adoption barriers | claude-sonnet-5 | 2/5 | No licensing requirement and no mandated human sign-off for internal data analysis, though organizations often require review of results before decisions are made on them, creating mild friction. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 4/5 | AI inference for code generation and statistical scripting costs (via APIs or subscriptions) are now a fraction of hourly data scientist wages, with minimal overhead; integration into IDE and analysis workflows adds modest costs. The cost advantage is substantial though not quite order-of-magnitude due to oversight and validation requirements. |
| Cost vs. human wage | claude-sonnet-5 | 4/5 | AI-assisted coding and analysis tools cost a small fraction of a data scientist's hourly wage for equivalent code generation and data processing, though human oversight and validation still add cost. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 4/5 | Mature AI products (ChatGPT, Claude, GitHub Copilot, automated ML platforms) demonstrably perform statistical coding, data transformation, and exploratory analysis in production environments at scale. Limitations remain in complex multi-step analyses and custom statistical designs, but core capabilities are well-established and widely deployed. |
| Technical feasibility today | claude-sonnet-5 | 4/5 | Products like GitHub Copilot, ChatGPT Code Interpreter/Advanced Data Analysis, and various AI-augmented notebooks are deployed and reliably used in production for data manipulation and statistical scripting tasks. |
Apply sampling techniques to determine groups to be surveyed or use complete enumeration methods.
67CI 55–79 · exposure 62 · augmentation 75 · click for rater detail
Apply sampling techniques to determine groups to be surveyed or use complete enumeration methods.
67| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 4/5 | Data science and analytics roles are in high-digitization sectors (tech, finance, professional services) with rapid AI adoption. Automated statistical pipelines and survey design tools are already deployed in production analytics teams, showing measurable displacement of routine sampling work. |
| Sector adoption velocity | claude-sonnet-5 | 3/5 | Data science and analytics functions are moderately fast adopters of AI-assisted coding and statistical tools, though full survey design automation remains a niche pilot use case. |
| Augmentation potential | claude-haiku-4-5-20251001 | 4/5 | AI significantly assists data scientists by automating sampling methodology selection, calculating optimal sample sizes, and generating enumeration strategies, allowing humans to focus on business context and validation of survey design rather than computational execution. |
| Augmentation potential | claude-sonnet-5 | 4/5 | AI substantially speeds up writing sampling code, checking assumptions, and suggesting methodology options, meaningfully boosting a data scientist's productivity while they retain final design authority. |
| Task automatability | claude-haiku-4-5-20251001 | 4/5 | AI systems can readily apply standard sampling methodologies (stratified, random, cluster sampling) and complete enumeration logic given dataset parameters. This involves computational decisions and statistical technique selection where current tools (Python/R automation, AI-assisted statistical packages) can execute 50%+ time savings at equal or better quality for well-defined survey design problems. |
| Task automatability | claude-sonnet-5 | 3/5 | AI can recommend and implement standard sampling schemes (stratified, cluster, random) given clear objectives and data, but determining appropriate population definitions and survey design still requires human judgment and domain context. |
| Adoption barriers | claude-haiku-4-5-20251001 | 2/5 | Few hard barriers exist; sampling technique application is computational and algorithmic rather than requiring licensure or legal sign-off. Organizational friction around autonomy and methodological oversight is the primary friction, not regulatory or liability constraints. |
| Adoption barriers | claude-sonnet-5 | 2/5 | No licensing requirement typically governs this task, though some survey work (e.g., official statistics, government surveys) has methodological standards requiring sign-off by qualified statisticians. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 5/5 | Computational sampling and enumeration can be executed at near-zero marginal cost via automated pipelines once configured, making AI orders of magnitude cheaper than paying a data scientist labor for routine sampling design and application. |
| Cost vs. human wage | claude-sonnet-5 | 3/5 | AI can cheaply generate sampling code and plans, but oversight, validation against survey goals, and domain expertise checks keep total cost roughly comparable to a skilled analyst for non-trivial designs. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 4/5 | Multiple deployed products (statistical software with AI-assisted features, automated analytics platforms, specialized survey design tools) reliably perform sampling technique selection and execution in production. Error rates are low for standard scenarios, though edge cases with complex stratification requirements may still require human review. |
| Technical feasibility today | claude-sonnet-5 | 3/5 | Code-generation and statistical assistant tools can implement sampling methods reliably for well-specified problems, but they are not yet deployed as autonomous survey-design decision-makers in production. |
Read scientific articles, conference papers, or other sources of research to identify emerging analytic trends and technologies.
67CI 50–84 · exposure 55 · augmentation 88 · click for rater detail
Read scientific articles, conference papers, or other sources of research to identify emerging analytic trends and technologies.
67| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 3/5 | Tech and biotech firms increasingly pilot AI literature review tools; adoption is growing but remains concentrated in high-digitization sectors. Many data science teams still rely primarily on manual exploration, so deployment is not yet deep or universal. |
| Sector adoption velocity | claude-sonnet-5 | 4/5 | Data science and tech-adjacent professional fields show fast adoption of AI tools for research synthesis, literature review, and trend-spotting. |
| Augmentation potential | claude-haiku-4-5-20251001 | 4/5 | AI assistants substantially boost productivity for literature review by rapidly scanning, summarizing, and clustering papers, surfacing candidate trends for human evaluation. Data scientists using these tools can cover far more ground while maintaining critical judgment on what is truly emerging. |
| Augmentation potential | claude-sonnet-5 | 5/5 | AI dramatically accelerates literature discovery, summarization, and trend identification while the human retains judgment over relevance and application, making it a strong augmentation case. |
| Task automatability | claude-haiku-4-5-20251001 | 2/5 | Current AI can extract key information from papers and identify explicit topics/keywords, but identifying genuinely emerging trends requires synthesis across dispersed literature, contextual judgment about significance, and domain expertise to distinguish signal from noise. This demands human interpretation. |
| Task automatability | claude-sonnet-5 | 4/5 | AI can summarize and synthesize large volumes of research literature, extract key trends, and surface relevant papers, saving substantial time over manual reading, though nuanced trend judgment still benefits from human review. |
| Adoption barriers | claude-haiku-4-5-20251001 | 2/5 | No licensing or legal barriers; organizations face low friction in piloting AI-assisted literature review. Adoption is primarily constrained by organizational inertia and preference for human domain expertise rather than regulatory or liability concerns. |
| Adoption barriers | claude-sonnet-5 | 1/5 | There are no licensing, regulatory, or liability barriers to using AI for reading and summarizing research literature. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 4/5 | AI-powered literature review and summarization tools have low per-article costs compared to a data scientist's hourly rate. Infrastructure is inexpensive, though domain experts still oversee results, so the all-in cost is favorable for AI compared to manual reading. |
| Cost vs. human wage | claude-sonnet-5 | 5/5 | AI-based literature review and summarization tools cost a fraction of a data scientist's hourly wage for equivalent reading and synthesis throughput. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 3/5 | Products exist (literature review assistants, paper summarization tools, semantic search) and perform reliably on narrow subtasks like extraction and clustering. However, reliable end-to-end trend identification remains inconsistent; systems often miss subtle patterns or generate false positives that require human verification. |
| Technical feasibility today | claude-sonnet-5 | 4/5 | Products like AI research assistants, literature summarization tools, and semantic search engines (e.g., Elicit, Consensus, Perplexity) are deployed and used reliably for literature scanning and synthesis today. |
Identify relationships and trends or any factors that could affect the results of research.
62CI 57–66 · exposure 55 · augmentation 100 · click for rater detail
Identify relationships and trends or any factors that could affect the results of research.
62| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 4/5 | Tech, finance, and enterprise analytics sectors are rapidly embedding AI-assisted EDA tools and agent-based analysis workflows; pilot and early production adoption is widespread. |
| Sector adoption velocity | claude-sonnet-5 | 4/5 | Data science is among the most AI-forward professional fields, with widespread adoption of ML-assisted analytics, automated feature discovery, and LLM-based exploratory tools in production pipelines. |
| Augmentation potential | claude-haiku-4-5-20251001 | 5/5 | AI dashboards, automated statistical summaries, and LLM-driven hypothesis generation substantially accelerate trend discovery and pattern surfacing, allowing data scientists to focus on validation and interpretation rather than manual scanning. |
| Augmentation potential | claude-sonnet-5 | 5/5 | AI substantially accelerates hypothesis generation, pattern detection, and trend identification, letting data scientists explore far more relationships and confounders than manual analysis would allow. |
| Task automatability | claude-haiku-4-5-20251001 | 3/5 | AI systems can identify statistical correlations, detect temporal trends, and flag potential confounding variables in datasets with strong signal, but typically require human validation of causal interpretation, domain context, and research quality—meeting perhaps 40–50% of the analyst's cognitive workload at equal quality. |
| Task automatability | claude-sonnet-5 | 3/5 | AI can surface correlations, trends, and confounders in structured datasets quickly, but identifying meaningful causal relationships and contextually relevant factors still requires domain judgment that current systems only partially replicate. |
| Adoption barriers | claude-haiku-4-5-20251001 | 2/5 | No licensing or legal requirement mandates human data scientists for this exploratory phase; institutional inertia and QA expectations provide modest friction but no hard regulatory barrier. |
| Adoption barriers | claude-sonnet-5 | 2/5 | No licensing requirement, but research integrity norms, peer review, and organizational accountability for conclusions create moderate friction against fully delegating interpretive judgment to AI. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 4/5 | AI-assisted exploratory analysis (notebooks, cloud-based tools, LLM APIs) costs a small fraction of a data scientist's fully loaded wage, though integration and oversight remain non-trivial per task. |
| Cost vs. human wage | claude-sonnet-5 | 3/5 | AI tools reduce time spent on exploratory analysis substantially, but the need for human oversight, domain framing, and validation keeps blended costs only moderately below a skilled data scientist's fully loaded cost. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 4/5 | Multiple mature products (statistical libraries, automated EDA tools, LLM-assisted analysis pipelines) reliably detect associations and trends in structured data; however, error rates remain material for subtle confounders or domain-specific artifacts that require expert judgment. |
| Technical feasibility today | claude-sonnet-5 | 3/5 | Deployed tools (AutoML, statistical analysis assistants, LLM-based data exploration agents) can find patterns and flag anomalies, but their outputs on nuanced research questions still require expert validation and often miss subtle confounds. |
Test, validate, and reformulate models to ensure accurate prediction of outcomes of interest.
57CI 57–57 · exposure 50 · augmentation 100 · click for rater detail
Test, validate, and reformulate models to ensure accurate prediction of outcomes of interest.
57| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 4/5 | Tech-forward and finance companies have rapidly adopted automated ML pipelines with integrated validation; MLOps practices and continuous integration/continuous deployment (CI/CD) for models are now standard in information and finance sectors, showing deep, fast adoption of testing automation. |
| Sector adoption velocity | claude-sonnet-5 | 4/5 | Data science and ML engineering in tech, finance, and other digitized sectors have rapidly adopted AutoML, MLOps, and AI-assisted coding tools for model testing and iteration. |
| Augmentation potential | claude-haiku-4-5-20251001 | 5/5 | AI augmentation is particularly strong here: automated validation reports, anomaly detection, hyperparameter suggestions, and visualization of model behavior substantially amplify a data scientist's ability to diagnose and iterate on models while retaining full human control over reformulation decisions. |
| Augmentation potential | claude-sonnet-5 | 5/5 | AI tools substantially speed up hyperparameter tuning, code generation for validation scripts, and diagnostic reporting, greatly boosting data scientist productivity while humans retain final judgment. |
| Task automatability | claude-haiku-4-5-20251001 | 3/5 | Current AI can partially automate statistical validation (hyperparameter tuning, cross-validation, automated testing frameworks), but the conceptual work of model reformulation based on validation failures—understanding *why* a model failed and what structural changes are needed—requires significant human judgment that existing systems cannot reliably do end-to-end. |
| Task automatability | claude-sonnet-5 | 3/5 | AI can assist with running validation metrics, cross-validation, and even suggest reformulations, but the judgment of whether results are meaningful, choosing appropriate validation strategy, and interpreting business context still require substantial human oversight. |
| Adoption barriers | claude-haiku-4-5-20251001 | 2/5 | Few hard regulatory barriers exist for automating validation itself, though organizational governance around model release often requires human sign-off. The primary barriers are soft: organizations prefer human accountability for model decisions and reformulation requires domain context that teams are reluctant to fully automate. |
| Adoption barriers | claude-sonnet-5 | 2/5 | No licensing requirement, but organizational risk tolerance and need for domain expertise to validate model correctness create moderate friction against full automation. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 3/5 | Automated validation and testing tools significantly reduce the compute and labor cost of the mechanical validation phase, but integration overhead, human oversight, and the need for domain expertise in reformulation keep total cost comparable to a mid-level data scientist's time. |
| Cost vs. human wage | claude-sonnet-5 | 3/5 | Automated validation pipelines reduce compute and some labor cost, but still require skilled data scientist oversight and infrastructure, making the cost roughly comparable rather than an order of magnitude cheaper. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 3/5 | Production tools exist for automated model testing and validation (MLflow, Vertex AI, H2O AutoML), but they operate within narrow scopes (e.g., standard metrics, predefined test suites). The broader task of interpreting results and reformulating models based on domain knowledge remains largely manual in deployed workflows. |
| Technical feasibility today | claude-sonnet-5 | 3/5 | AutoML and code-assistant tools (e.g., Copilot, AutoML platforms) exist and are used in production for model testing/validation, but they still have narrow scope and require human review for edge cases, data leakage, and domain-specific validity. |
Deliver oral or written presentations of the results of mathematical modeling and data analysis to management or other end users.
52CI 46–57 · exposure 50 · augmentation 88 · click for rater detail
Deliver oral or written presentations of the results of mathematical modeling and data analysis to management or other end users.
52| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 3/5 | Many firms are piloting AI-assisted report generation and presentation drafting (e.g., auto-generated dashboards, LLM summaries), but production replacement of the entire presentation task remains rare; adoption is at the augmentation stage rather than full automation. |
| Sector adoption velocity | claude-sonnet-5 | 4/5 | Data science and analytics teams in tech, finance, and professional services are fast adopters of AI writing/summarization tools for reporting tasks. |
| Augmentation potential | claude-haiku-4-5-20251001 | 4/5 | AI tools meaningfully assist data scientists by auto-generating draft narratives, creating visualizations, and summarizing findings—this substantially raises their productivity in assembling and communicating results while the scientist retains judgment over accuracy and framing. |
| Augmentation potential | claude-sonnet-5 | 5/5 | AI substantially speeds up drafting slides, summarizing statistical findings, and generating narrative explanations, letting the human focus on delivery and interpretation. |
| Task automatability | claude-haiku-4-5-20251001 | 3/5 | AI can auto-generate written summaries and visualizations of analysis results with moderate quality, but delivering tailored presentations to management requires domain understanding, audience calibration, and real-time adjustment that current AI systems handle unevenly—roughly half the work could be automated with significant setup. |
| Task automatability | claude-sonnet-5 | 3/5 | AI can draft written reports, slides, and summaries of analysis results with significant time savings, but live oral presentation to management and adaptive Q&A still requires human delivery and judgment. |
| Adoption barriers | claude-haiku-4-5-20251001 | 3/5 | Organizational and reputational barriers exist: management expects human credibility and accountability for high-stakes insights, and liability for misreported results typically requires a named human expert; there is no legal licensing requirement but strong informal pressure for human sign-off. |
| Adoption barriers | claude-sonnet-5 | 2/5 | No licensing requirement, but organizational norms often expect a human to personally present findings to management for credibility, trust, and accountability. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 2/5 | Current AI tools (API calls, integration, oversight) cost significantly less than a data scientist's hourly rate, but human review and customization overhead is high; for full end-to-end replacement with equal quality, the all-in cost remains comparable to or exceeds hiring a junior analyst. |
| Cost vs. human wage | claude-sonnet-5 | 3/5 | AI drastically cuts drafting time for written summaries, but human review, customization, and actual delivery still require substantial paid time, keeping overall cost comparable rather than order-of-magnitude cheaper. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 3/5 | Deployed tools (LLMs for report generation, automated BI dashboards, presentation software) exist and function in production, but error rates remain material: misinterpretations of model output, tone-deafness to audience, and inaccurate summarization occur regularly enough to require human review. |
| Technical feasibility today | claude-sonnet-5 | 3/5 | Tools like ChatGPT, Copilot, and BI assistants routinely generate presentation drafts and narrative summaries in production, but they don't autonomously deliver oral presentations or handle stakeholder interaction reliably. |
Design surveys, opinion polls, or other instruments to collect data.
40CI 21–59 · exposure 38 · augmentation 75 · click for rater detail
Design surveys, opinion polls, or other instruments to collect data.
40| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 2/5 | While data science adopts AI tooling broadly, survey design automation specifically is not seeing production-scale deployment. Organizations remain cautious about algorithmic survey design due to methodological and compliance risks, limiting observed velocity. |
| Sector adoption velocity | claude-sonnet-5 | 3/5 | Data science and market research teams are adopting AI drafting tools for surveys at a moderate pace, with pilots common but full automation of instrument design still uncommon in production workflows. |
| Augmentation potential | claude-haiku-4-5-20251001 | 4/5 | AI can meaningfully assist survey designers by generating question suggestions, identifying potential biases, proposing response scales, and offering templates—substantially raising designer productivity while the human retains control over methodological choices and validation. |
| Augmentation potential | claude-sonnet-5 | 4/5 | AI significantly speeds up brainstorming, wording refinement, and structuring of survey questions, letting data scientists focus on sampling design and validation while staying in the loop. |
| Task automatability | claude-haiku-4-5-20251001 | 2/5 | AI can assist with survey drafting and template generation, but designing statistically sound instruments requires domain expertise, understanding of bias, and alignment with specific research questions that demand human judgment. End-to-end automation would not reliably meet the 50% time-saving threshold for equal quality output. |
| Task automatability | claude-sonnet-5 | 3/5 | AI can draft survey questions, suggest scales, and check for bias/wording issues, but designing a valid instrument requires domain judgment, sampling strategy, and stakeholder alignment that still needs human oversight for at least half the workflow. |
| Adoption barriers | claude-haiku-4-5-20251001 | 4/5 | Survey design for research, market research, and policy contexts often requires sign-off by research ethics boards, compliance with data protection regulations, and accountability for methodological rigor. Professional standards and organizational liability create significant adoption friction. |
| Adoption barriers | claude-sonnet-5 | 2/5 | No licensing requirement exists for survey design, though quality/validity concerns and organizational review create some friction before AI-drafted instruments are deployed in high-stakes research. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 1/5 | The loaded cost of a data scientist's time for survey design is substantial, but AI systems would require significant human review, iteration, and domain expertise oversight, making the all-in cost comparable or higher than having a human do it directly. |
| Cost vs. human wage | claude-sonnet-5 | 4/5 | Generating draft survey items and structures via LLMs costs a small fraction of a data scientist's time compared to manually drafting and iterating on instruments from scratch. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 2/5 | While generative AI can produce survey templates and suggest question types, no deployed product reliably designs methodologically sound survey instruments end-to-end. Tools exist for survey administration, not design; actual design requires human expertise and faces high stakes for methodological errors. |
| Technical feasibility today | claude-sonnet-5 | 3/5 | Products like ChatGPT, Qualtrics AI, and SurveyMonkey Genius exist to draft and refine survey items today, but they operate narrowly (question generation, basic bias checks) rather than end-to-end instrument design including sampling and validation. |
Identify solutions to business problems, such as budgeting, staffing, and marketing decisions, using the results of data analysis.
33CI 28–39 · exposure 25 · augmentation 88 · click for rater detail
Identify solutions to business problems, such as budgeting, staffing, and marketing decisions, using the results of data analysis.
33| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 3/5 | Data-driven organizations are actively piloting AI-assisted analytics and insight generation, but adoption of AI as the primary agent for *identifying solutions* to business problems remains limited to augmentation and recommendation scenarios. Full displacement remains uncommon in production. |
| Sector adoption velocity | claude-sonnet-5 | 4/5 | Data science and business analytics roles are in a fast-adopting sector (professional/tech services) where AI-assisted analytics tools are already widely deployed in production pipelines. |
| Augmentation potential | claude-haiku-4-5-20251001 | 4/5 | AI substantially augments data scientists on this task by rapidly generating candidate solutions, surfacing patterns in data, drafting option summaries, and stress-testing assumptions, while the human retains critical evaluation and decision authority. This is a strong augmentation use case. |
| Augmentation potential | claude-sonnet-5 | 5/5 | AI significantly enhances data scientists' ability to analyze datasets, generate insights, and draft recommendations, greatly increasing productivity while humans retain final decision authority. |
| Task automatability | claude-haiku-4-5-20251001 | 2/5 | While AI can perform the data analysis portion and even suggest patterns, identifying *solutions* to business problems requires domain expertise, stakeholder input, and judgment about business constraints and risk tolerance that current AI systems cannot reliably provide end-to-end. The task involves translating technical findings into actionable decisions, which remains substantially human-dependent. |
| Task automatability | claude-sonnet-5 | 2/5 | AI can analyze data and surface patterns, but translating results into concrete business decisions like staffing or budgeting requires contextual judgment, stakeholder knowledge, and accountability that current systems cannot fully replicate end-to-end.' |
| Adoption barriers | claude-haiku-4-5-20251001 | 4/5 | This task sits within a managerial and strategic decision-making context where organizational friction, liability concerns (poor decisions harm the business), and implicit requirements for human accountability and judgment substantially protect against full automation. Organizations are cautious about delegating solution identification to AI without human sign-off. |
| Adoption barriers | claude-sonnet-5 | 3/5 | No formal licensing is required, but organizational accountability, liability for bad business decisions, and the need for stakeholder buy-in create meaningful friction against full automation. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 2/5 | Current AI systems (data analysis tools, dashboards, LLMs) still require significant expert human oversight to translate results into valid business solutions, meaning the all-in cost (inference + human review + integration) remains comparable to or higher than employing a data scientist for this synthesis work. |
| Cost vs. human wage | claude-sonnet-5 | 3/5 | AI tools reduce analysis time and cost, but the decision-making component still requires costly human oversight, integration, and validation, keeping overall cost roughly comparable rather than dramatically cheaper. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 2/5 | No production systems today autonomously identify and present solutions to complex business problems from data analysis. Deployed tools can assist with analysis or suggest options, but the integration of data results into vetted business solutions requires human validation and executive sign-off in practice. |
| Technical feasibility today | claude-sonnet-5 | 2/5 | Deployed analytics and BI copilots can generate insights and recommendations, but production systems rarely make final business decisions autonomously; human review is standard practice today. |
Identify business problems or management objectives that can be addressed through data analysis.
31CI 25–38 · exposure 25 · augmentation 63 · click for rater detail
Identify business problems or management objectives that can be addressed through data analysis.
31| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 2/5 | Adoption of AI for problem identification is slow because the task requires deep organizational context, executive trust, and accountability—sectors like finance and tech are experimenting with AI-assisted discovery tools, but full automation remains rare and limited to narrow, pre-scoped use cases. |
| Sector adoption velocity | claude-sonnet-5 | 3/5 | Data science and analytics functions in tech, finance, and professional services are adopting AI copilots quickly for exploratory and coding tasks, though strategic problem framing lags behind these adoption trends. |
| Augmentation potential | claude-haiku-4-5-20251001 | 3/5 | AI can assist by surfacing data anomalies, suggesting potential analytical approaches, or synthesizing prior case studies, helping data scientists brainstorm and validate problem scope faster. However, the human data scientist remains essential for translating business language into analytical questions. |
| Augmentation potential | claude-sonnet-5 | 4/5 | AI can help brainstorm potential use cases, summarize data patterns, and surface analogous problems from other domains, meaningfully accelerating a data scientist's ideation and scoping process. |
| Task automatability | claude-haiku-4-5-20251001 | 2/5 | Identifying business problems requires understanding organizational context, strategy, and stakeholder needs—tasks that demand human judgment and domain expertise. While AI can surface patterns in existing data or suggest hypotheses, current systems cannot reliably elicit and frame business objectives without significant human guidance, making end-to-end automation with 50% time savings infeasible. |
| Task automatability | claude-sonnet-5 | 2/5 | This requires understanding organizational context, stakeholder priorities, and tacit business knowledge that current AI cannot independently gather or synthesize into strategic framing; AI can support but not replace this end-to-end. |
| Adoption barriers | claude-haiku-4-5-20251001 | 4/5 | Strong organizational and domain barriers protect this task: business problem identification is a core strategic and consultative function that stakeholders expect humans to lead, and misalignment on objectives carries high business risk, creating friction against full automation. |
| Adoption barriers | claude-sonnet-5 | 2/5 | No licensing or regulatory requirement exists, but organizational trust, stakeholder relationships, and accountability for strategic direction create moderate friction against pure AI substitution. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 2/5 | AI systems today offer limited assistance with problem identification and still require expensive human data scientists to validate, refine, and contextualize findings. The overhead of oversight and iteration means current AI is not cheaper than direct human engagement for this task. |
| Cost vs. human wage | claude-sonnet-5 | 2/5 | Since AI cannot independently perform this well, the effective cost includes substantial human oversight and business judgment, keeping costs comparable to or higher than a skilled human doing it alone. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 2/5 | No deployed product reliably identifies business problems or management objectives independently. Tools like analytics platforms and business intelligence systems support discovery but require human analysts to interpret organizational needs and validate problem statements; they do not perform this task end-to-end in production. |
| Technical feasibility today | claude-sonnet-5 | 2/5 | Some AI copilots can suggest analysis opportunities from data or documents, but no deployed product reliably identifies novel business problems worth solving without heavy human framing and validation. |
Recommend data-driven solutions to key stakeholders.
31CI 30–32 · exposure 25 · augmentation 75 · click for rater detail
Recommend data-driven solutions to key stakeholders.
31| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 2/5 | Although data science is a digital-forward field, actual deployment of AI systems to autonomously generate stakeholder recommendations remains limited; most organizations use AI for analysis and drafting assistance, but the final recommendation still comes from a human data scientist. |
| Sector adoption velocity | claude-sonnet-5 | 3/5 | Data science and analytics functions are moderately fast adopters of AI tools for drafting and analysis, but the final recommendation-and-persuasion step remains largely human-led in most organizations today. |
| Augmentation potential | claude-haiku-4-5-20251001 | 4/5 | Modern generative AI significantly assists data scientists in synthesizing findings, creating visualizations, drafting explanations, and structuring recommendations, substantially accelerating the time from analysis to presentation while the human retains final authority and framing. |
| Augmentation potential | claude-sonnet-5 | 4/5 | AI significantly boosts productivity in generating data summaries, visualizations, and framing options, which data scientists can then refine into stakeholder recommendations. |
| Task automatability | claude-haiku-4-5-20251001 | 2/5 | Data scientists typically synthesize analysis into narrative recommendations tailored to stakeholder context, priorities, and constraints—requiring judgment and persuasion that current AI systems cannot reliably replicate end-to-end. While AI can generate draft insights and candidate solutions from data, translating these into actionable, politically-informed recommendations for specific decision-makers remains a predominantly human task. |
| Task automatability | claude-sonnet-5 | 2/5 | Recommending solutions to stakeholders requires synthesizing business context, organizational politics, and judgment calls that current AI cannot fully replicate end-to-end, though it can draft analysis summaries. |
| Adoption barriers | claude-haiku-4-5-20251001 | 3/5 | While there is no hard licensing requirement, stakeholder expectation of human accountability and the reputational risk of automated recommendations creates material organizational friction; customers and internal teams often prefer recommendations to be authored and owned by a human. |
| Adoption barriers | claude-sonnet-5 | 3/5 | No formal licensing requirement, but organizational trust, accountability for business outcomes, and stakeholder preference for human judgment create moderate friction against full automation. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 2/5 | Using AI to draft or assist with recommendations still requires substantial human oversight and refinement to ensure accuracy and alignment with business context, making total cost (inference + integration + validation) comparable to or exceeding the cost of a data scientist performing the task directly. |
| Cost vs. human wage | claude-sonnet-5 | 2/5 | While AI can cheaply produce draft analyses, the human cost of validating, contextualizing, and presenting recommendations to stakeholders remains substantial, keeping overall cost comparable to a human-led process. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 2/5 | No mature product reliably performs stakeholder recommendation generation at production scale; generative AI can assist in drafting talking points or summarizing findings, but lacks the contextual understanding, organizational knowledge, and accountability that real stakeholders expect from a recommendation. |
| Technical feasibility today | claude-sonnet-5 | 2/5 | AI tools can generate data summaries and suggest options, but no deployed product reliably formulates and delivers stakeholder-ready strategic recommendations without heavy human curation. |
Propose solutions in engineering, the sciences, and other fields using mathematical theories and techniques.
29CI 25–32 · exposure 25 · augmentation 75 · click for rater detail
Propose solutions in engineering, the sciences, and other fields using mathematical theories and techniques.
29| Dimension | Model | Rating | Rationale |
|---|---|---|---|
| Sector adoption velocity | claude-haiku-4-5-20251001 | 2/5 | Scientific and engineering firms are experimenting with AI assistance (e.g., generative models in drug discovery), but production deployment of autonomous solution proposal remains limited and typically in narrow, well-scoped domains. Adoption is slower than in information-processing tasks. |
| Sector adoption velocity | claude-sonnet-5 | 3/5 | Data science and adjacent technical fields show moderate AI tool adoption for coding and analysis, but using AI to autonomously propose scientific/engineering solutions remains mostly pilot-stage. |
| Augmentation potential | claude-haiku-4-5-20251001 | 4/5 | AI excels at accelerating literature review, generating candidate mathematical formulations, and exploring parameter spaces, enabling human scientists to iterate faster and explore a wider solution space. These are high-value augmentations while the human remains responsible for judgment and validation. |
| Augmentation potential | claude-sonnet-5 | 4/5 | AI substantially helps generate hypotheses, run exploratory calculations, and draft technical write-ups, meaningfully speeding up the ideation and drafting phase while the scientist validates and refines. |
| Task automatability | claude-haiku-4-5-20251001 | 2/5 | Proposing novel solutions requires domain expertise, creativity, and judgment that current AI struggles with at parity quality. While AI can generate candidate solutions and assist in mathematical formulation, the creative synthesis and evaluation of trade-offs remains human-dependent, and independent end-to-end automation would likely miss critical contextual constraints. |
| Task automatability | claude-sonnet-5 | 2/5 | AI can suggest mathematical approaches or draft partial solutions but proposing well-grounded solutions across engineering and scientific domains requires deep contextual judgment, validation, and creativity that current systems cannot reliably deliver end-to-end. |
| Adoption barriers | claude-haiku-4-5-20251001 | 4/5 | Professional and legal liability for engineering/scientific proposals is high; clients and regulators expect a qualified human expert to be accountable. Organizational culture in R&D and engineering heavily favors human sign-off on novel solutions, and many sectors require licensed engineers to certify proposals. |
| Adoption barriers | claude-sonnet-5 | 3/5 | No licensing requirement blocks AI-generated proposals, but liability for incorrect engineering/scientific solutions and organizational need for expert sign-off create meaningful friction. |
| Cost vs. human wage | claude-haiku-4-5-20251001 | 2/5 | Computational cost for advanced reasoning and oversight by domain experts to validate AI-generated solutions is substantial, while the error cost of false or suboptimal proposals is high. The all-in cost approaches or exceeds hiring a junior scientist for exploratory work. |
| Cost vs. human wage | claude-sonnet-5 | 2/5 | While AI queries are cheap, the human oversight, validation, and domain expertise still required to trust a proposed technical solution keeps the effective cost comparable to or only modestly below a skilled human's. |
| Technical feasibility today | claude-haiku-4-5-20251001 | 2/5 | No mature product reliably proposes novel engineering or scientific solutions independently. LLMs can draft solution sketches and mathematical frameworks, but human experts still filter and validate; deployment in real R&D remains assistive rather than autonomous. |
| Technical feasibility today | claude-sonnet-5 | 2/5 | Products like code assistants and LLM-based reasoning tools can suggest formulas or approaches, but no deployed product reliably proposes validated engineering/scientific solutions at production scale without heavy human vetting. |
Related occupations — Computer & Mathematical
How to read this
A high substitution score does not mean this job disappears — it means a large share of its current tasks face replacement pressure, so the mix of tasks is likely to change. High augmentation alongside substitution typically means the occupation reorganizes around the protected tasks. Wide confidence intervals mean the rater panel disagreed: treat those scores as open questions, not verdicts.
What would change this score
New model capabilities (automatability, feasibility), falling inference costs (cost ratio), regulation and licensing shifts (barriers), and measured sector adoption (velocity) all re-enter at every index release. Each release is recomputed, versioned and kept queryable — scores are claims with a date on them, not permanent labels.