Audience: Learners, analysts, researchers, data scientists, educators, and professionals using AI to support data analysis workflows
Theme: AI can support data analysis, but evidence, code, interpretation, and decisions must remain human-reviewed
A data analysis support agent helps a person move through an analytical workflow more clearly.
It can help frame questions, inspect tables, suggest checks, explain code, draft analysis plans, summarize results, and prepare communication materials.
But it should not replace analytical judgment.
It should not invent results.
It should not make unsupported claims.
It should not hide uncertainty.
A responsible data analysis support agent works inside a human-led system where data, code, assumptions, outputs, and interpretations can be checked.
The goal of this chapter is to define a reusable agent pattern for AI-supported data analysis.
Why Data Analysis Needs Support
Data analysis is rarely just one task.
It usually includes many connected activities:
understanding the question,
identifying the data source,
checking the structure of the data,
cleaning and preparing variables,
choosing an analytical method,
running code,
checking outputs,
interpreting results,
explaining limitations,
and communicating findings.
AI can help with many of these steps.
However, data analysis also carries risk.
A small mistake in data structure, variable meaning, method choice, code logic, or interpretation can change the conclusion.
For that reason, AI-supported analysis must be designed with review points.
The agent should support the analyst, not act as the analyst.
The Role of the Data Analysis Support Agent
The role of this agent is to help a human analyst work more systematically.
It can support:
Question framing
Data understanding
Analysis planning
Code support
Quality checks
Result interpretation
Communication
Documentation
The agent should ask for the necessary context before giving strong recommendations.
It should distinguish between what is known, what is assumed, and what needs checking.
It should help the analyst slow down at important decision points.
A useful data analysis support agent does not only produce code.
It helps protect the quality of the analysis process.
Human–AI–Human Pattern for Data Analysis
The Human–AI–Human pattern is especially important in data work.
Code
flowchart TD A[Human defines analytical question] B[Human provides data context] C[AI assists with workflow planning] D[AI supports code, checks, summaries, or interpretation] E[Human runs or reviews code] F[Human checks outputs against data and methods] G[Human decides what can be claimed] H[Human documents analysis and limitations] A --> B B --> C C --> D D --> E E --> F F --> G G --> H
flowchart TD
A[Human defines analytical question]
B[Human provides data context]
C[AI assists with workflow planning]
D[AI supports code, checks, summaries, or interpretation]
E[Human runs or reviews code]
F[Human checks outputs against data and methods]
G[Human decides what can be claimed]
H[Human documents analysis and limitations]
A --> B
B --> C
C --> D
D --> E
E --> F
F --> G
G --> H
The human begins the process by defining the analytical goal.
The AI assists in the middle.
The human completes the process by checking evidence and deciding what can responsibly be said.
This is different from asking AI to “analyze this data” without a review system.
A responsible system makes the steps visible.
What the Agent Can Help With
A data analysis support agent can be useful across the analysis lifecycle.
1. Clarifying the analytical question
The agent can help turn broad questions into clearer analytical questions.
For example:
Broad question:
Does this training program work?
Clearer analytical question:
Among learners who completed the program, how did assessment scores change from baseline to final evaluation, and what limitations affect interpretation?
This step matters because unclear questions produce unfocused analysis.
The agent should help define:
the population,
the outcome,
the comparison,
the time period,
the unit of analysis,
and the intended decision.
2. Understanding the dataset
The agent can help the analyst inspect the structure of a dataset.
It can suggest checks such as:
Number of rows
Number of columns
Column names
Variable types
Missing values
Duplicate records
Unexpected categories
Date ranges
Outliers
Identifier uniqueness
The agent should not assume that column names are self-explanatory.
It should encourage use of a data dictionary, metadata file, codebook, or project notes.
3. Planning the analysis
The agent can help create a simple analysis plan.
A good plan includes:
Question
Dataset
Variables
Inclusion and exclusion rules
Cleaning steps
Exploratory checks
Main method
Sensitivity checks
Expected outputs
Review points
Limitations
This keeps the analysis from becoming a collection of disconnected code snippets.
4. Supporting code development
The agent can help draft, explain, debug, or improve code.
This can include code in:
R
Python
SQL
Bash
Quarto
Markdown
However, generated code must be reviewed and run by the human analyst.
The agent should not claim that code works unless it has actually been tested in the relevant environment.
When possible, code should be written in small, inspectable steps.
5. Checking outputs
The agent can help create output review checklists.
For example:
Do row counts match expectations?
Are missing values handled consistently?
Are categories labeled correctly?
Do summaries match the analysis question?
Are plots readable and correctly labeled?
Are statistical outputs interpreted cautiously?
Are claims supported by the actual results?
This is one of the most important uses of AI in analysis: helping the analyst remember what to check.
6. Interpreting results
The agent can help translate results into plain language.
But interpretation must remain evidence-based.
The agent should help separate:
What the results show
What the results suggest
What the results do not prove
What needs further analysis
What limitations must be stated
This prevents overclaiming.
7. Communicating findings
The agent can support drafting summaries, reports, slides, briefs, and README sections.
The human must verify that the communication matches the analysis.
Beautiful language should never be allowed to cover weak evidence.
What the Agent Should Not Do
The data analysis support agent should have clear limits.
It should not:
invent datasets,
invent results,
fabricate citations,
hide uncertainty,
choose methods without context,
treat correlation as causation,
ignore missing data,
ignore measurement problems,
make high-stakes decisions alone,
or present untested code as verified.
These limits should be part of the agent instructions.
A useful agent is not only powerful.
It is bounded.
Agent Input Requirements
Before helping with an analysis task, the agent should ask for or identify the required inputs.
A basic input checklist is:
1. Analytical question
2. Dataset or table description
3. Variable definitions
4. Unit of analysis
5. Desired output
6. Tooling environment
7. Known limitations
8. Review requirements
For example:
Question:
What are we trying to learn or decide?
Dataset:
What table, file, database, or result object is being used?
Variables:
What do the important columns mean?
Unit of analysis:
Is each row a person, sample, observation, transaction, gene, protein, document, or event?
Output:
Do we need a table, plot, report, interpretation, or decision memo?
Environment:
Should the workflow use R, Python, SQL, Quarto, Excel, or another tool?
Limitations:
What should not be assumed?
Review:
Who checks the final output before it is used?
Without these inputs, the agent should be cautious.
Reusable Agent Template
A reusable data analysis support agent can be defined like this:
Agent name:
Data Analysis Support Agent
Purpose:
Support a human analyst in planning, checking, coding, interpreting, and communicating data analysis work.
Primary users:
Learners, analysts, researchers, data scientists, educators, and project teams.
Inputs:
Analytical question, dataset description, variable definitions, sample rows or schema, preferred tools, expected output, and known limitations.
Core workflow:
1. Clarify the analytical question.
2. Identify the unit of analysis.
3. Review available data and metadata.
4. Propose an analysis plan.
5. Suggest or draft code in small steps.
6. Identify quality checks.
7. Help interpret outputs cautiously.
8. Draft communication materials.
9. Document assumptions and limitations.
10. Require human review before use.
Output standards:
Outputs should distinguish evidence from interpretation, state assumptions clearly, avoid unsupported claims, and include review notes.
Boundaries:
The agent must not fabricate results, invent data, make unsupported causal claims, or replace human analytical responsibility.
This template can be adapted to different fields.
For example, the same pattern can support public health analysis, omics analysis, education program evaluation, business reporting, or portfolio project review.
Example Prompt for the Agent
A useful starting prompt is:
You are a Data Analysis Support Agent.
Your role is to support, not replace, the human analyst.
Help me clarify the analytical question, identify the required data inputs, plan the analysis, suggest quality checks, support code development, and prepare cautious interpretation.
Do not invent results.
Do not assume variable meanings without evidence.
Do not make causal claims unless the study design supports them.
Separate what is known, what is assumed, and what needs checking.
Ask for missing context when it affects the analysis.
End with a human review checklist.
This prompt gives the agent a role, purpose, boundaries, and expected behavior.
It is better than asking:
Analyze my data.
The better prompt creates a system.
The weaker prompt creates an uncontrolled task.
Example Use Case: Reviewing an Analysis Table
Suppose an analyst has a results table from a model or statistical test.
The agent can help inspect the table by asking:
What does each row represent?
What is the outcome variable?
What is the comparison or predictor?
Which columns contain estimates?
Which columns contain uncertainty measures?
Which columns contain p-values or adjusted p-values?
Were multiple comparisons performed?
What threshold is being used, and why?
What claims are allowed from this design?
For a CDI data workflow, this connects naturally to the broader data science foundation:
Raw or processed data
↓
Tidy results table
↓
Quality checks
↓
Interpretation
↓
Communication
↓
Decision or next action
The agent can support the movement from table to interpretation.
But the analyst must verify that the table is real, correctly produced, and correctly understood.
Example Use Case: Supporting Reproducible Code
The agent can also help with reproducible code.
A responsible code-support interaction might ask the agent to:
Review this R script for readability and reproducibility.
Identify any assumptions.
Suggest checks for missing values and duplicate IDs.
Do not change the analysis method unless you explain why.
Return a revised script and a short review checklist.
This is better than simply asking for code.
It makes review part of the task.
When AI writes code, the human should check:
Does the code run?
Are file paths correct?
Are packages available?
Are objects named consistently?
Are joins performed safely?
Are missing values handled intentionally?
Are outputs saved in the expected location?
Are results reproducible?
The agent can support this checklist, but the human must perform the final validation.
Example Use Case: Drafting a Results Narrative
The agent can help turn results into a written explanation.
A strong request would be:
Using the results table below, draft a cautious results narrative.
Separate findings, interpretation, and limitations.
Do not add numbers that are not present in the table.
Do not claim causality.
Use plain language suitable for a non-technical audience.
This protects against a common risk: fluent overinterpretation.
The output should be checked against the original table before being shared.
Data Analysis Review Checklist
Before using an AI-supported analysis output, the human analyst should review:
Question
[ ] Is the analytical question clearly stated?
[ ] Is the decision or purpose of the analysis clear?
Data
[ ] Is the data source documented?
[ ] Is the unit of analysis clear?
[ ] Are key variables defined?
[ ] Are missing values checked?
[ ] Are duplicates checked where relevant?
Methods
[ ] Is the chosen method appropriate for the question?
[ ] Are assumptions stated?
[ ] Are limitations documented?
Code
[ ] Does the code run in the intended environment?
[ ] Are file paths and outputs correct?
[ ] Are transformations inspectable?
[ ] Are results reproducible?
Outputs
[ ] Are tables and plots labeled clearly?
[ ] Do results match the analysis question?
[ ] Are interpretations supported by evidence?
[ ] Are uncertainty and limitations included?
Communication
[ ] Is the language clear?
[ ] Are claims cautious?
[ ] Is human review complete before sharing?
This checklist turns AI support into accountable analysis practice.
Common Failure Modes
AI-supported data analysis can fail in predictable ways.
Failure mode 1: The agent assumes column meanings
A column named score may refer to many things.
It could be an assessment score, risk score, model score, satisfaction score, or quality score.
The agent should not assume meaning without metadata.
Failure mode 2: The agent writes code that looks right but does not run
Generated code can contain wrong package names, wrong function arguments, or objects that do not exist.
The human must test the code.
Failure mode 3: The agent overstates the result
A descriptive comparison may be written as if it proves impact.
The human must check whether the study design supports the claim.
Failure mode 4: The agent ignores missingness
Missing data can change interpretation.
A responsible analysis should report missingness and explain how it was handled.
Failure mode 5: The agent produces polished but unsupported language
Professional writing is not the same as valid interpretation.
The final narrative must be checked against the evidence.
Documentation Standard
Every AI-supported analysis should leave a small record.
A simple documentation note can include:
AI use record
Project:
Date:
Human analyst:
AI role:
Input materials used:
Outputs generated:
Checks performed:
Known limitations:
Final human decision:
This does not need to be complicated.
The purpose is to make the workflow inspectable.
If someone later asks how the result was produced, the team should be able to explain the process.
How This Connects to CDI Data Science Guides
In the CDI ecosystem, data analysis is not treated as isolated computation.
It is part of a system:
Question
↓
Data
↓
Analysis
↓
Interpretation
↓
Communication
↓
Decision
The data analysis support agent fits into this system as an assistant.
It can help learners and professionals use the Data Science Foundations, Applied Data Science, Omics Systems, Career System, and other CDI guides more effectively.
However, the same principle remains:
AI may support the workflow, but the evidence must carry the claim.
This is especially important in scientific, medical, educational, financial, and policy settings.
Practical Exercise
Choose one small dataset or results table.
Use the data analysis support agent pattern to prepare a review plan.
Your output should include:
1. Analytical question
2. Dataset description
3. Unit of analysis
4. Key variables
5. Proposed checks
6. Suggested analysis steps
7. Expected outputs
8. Interpretation cautions
9. Human review checklist
Do not begin with code.
Begin with the system.
After the plan is clear, then code becomes easier to write, test, and explain.
Chapter Summary
A data analysis support agent helps a human analyst work more clearly and carefully.
It can support question framing, data understanding, analysis planning, code development, quality checks, interpretation, communication, and documentation.
But it must remain bounded.
It should not invent data, fabricate results, make unsupported claims, or replace human responsibility.
The strongest use of this agent is not to speed through analysis blindly.
It is to make analysis more inspectable, reproducible, and explainable.
Looking Ahead
The next chapter introduces the Communication and Reporting Agent.
That agent focuses on turning reviewed work into clear messages, reports, briefs, slides, documentation, and public-facing communication.
The same principle continues:
AI can help shape communication, but human review must protect accuracy, context, tone, and responsibility.