Human-Governed Agentic Analytics Workflow
Direct AI analysis with secure data, clear questions, and human validation
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 6
- Confidence
- 98%
The workflow places human judgment around an AI agent’s analytical execution. First, the analyst confirms that both the tool and the proposed data use comply with company security rules. The analyst then inspects data quality and defines a precise business question, the required analytical method, and the desired deliverable. AI handles time-consuming work such as coding, cohort construction, visualization, and drafting, while the analyst checks whether the evidence is complete and the outputs make sense. Independent arithmetic, historical context, domain knowledge, and targeted follow-up questions are used to challenge the result. Finally, the analyst converts the validated diagnosis into an audience-appropriate deliverable that explains both the root cause and the recommended response. The mechanism improves speed without transferring accountability to the model.
Origin
Extracted from Marketing Against The Grain during Sundus's demonstration of a Codex-assisted customer-retention investigation and leadership deck.
Core principles
- 01Use only company-approved AI tools for sensitive data
- 02Define the business problem and desired output before analysis
- 03Treat AI as an analytical intern rather than an autonomous authority
- 04Validate outputs with simple math and domain knowledge
- 05Inspect data completeness and cleanliness before trusting conclusions
- 06Pair every diagnosed problem with a proposed response
How to run it
- 1
Establish the security boundary
Determine whether company data may be processed by the chosen AI system. Use an internally approved enterprise tool with the required security controls.
Pro tip Check the organization’s existing enterprise subscriptions before evaluating outside tools.
Watch out Never upload company data to an unapproved service.
- 2
Audit the input data
Inspect the schema, time coverage, missing values, duplicates, and obvious inconsistencies. Clean and normalize the data before asking the model to explain it.
Pro tip Ask the agent to produce a data-quality report before beginning the main analysis.
Watch out AI may silently assume that supplied data is complete and clean.
- 3
Frame the analytical assignment
Specify the decision to support, the question to answer, the analytical method to use, and the required output. Make the success condition concrete enough that the result can be evaluated.
Pro tip Include the audience and format, such as a cohort analysis and leadership deck.
Watch out A vague request can produce polished work that answers the wrong question.
- 4
Delegate execution to the agent
Let the AI inspect the data, write analysis code, calculate metrics, generate charts, and draft deliverables. Monitor permissions and review the intermediate artifacts it creates.
Pro tip Treat the agent like an intern who can execute quickly but needs direction.
Watch out Do not interpret smooth execution as proof that the analysis is correct.
- 5
Validate the diagnosis
Recalculate headline figures, test whether relationships are plausible, and compare findings with historical and domain knowledge. Determine whether omitted data could produce a different explanation.
Pro tip Begin with simple arithmetic and sanity checks before reviewing complex models.
Watch out Correlation in a chart does not by itself establish causation.
- 6
Turn findings into action
Edit the deliverable for accuracy, clarity, and audience needs. Present the validated root cause together with specific corrective actions and anticipated answers to leadership questions.
Pro tip Add slides that address likely objections before the meeting.
Watch out Do not present a problem without explaining what should happen next.
In the wild
A leader requests an explanation for the previous week’s retention decline by 2:00 p.m. The analyst gives Codex a local customer-retention CSV, requests cohort analysis and a leadership deck, then validates the reported decline and investigates its relationship to mobile crash exposure. The reviewed evidence points to a new mobile release as the likely driver, and the analyst edits the generated files before presenting them.
→ The analyst produces a rapid, evidence-backed diagnosis and an actionable leadership presentation.
At Google, Sundus described the desired dataset and SQL task to an internal AI system. The system identified relevant tables and drafted the query, reducing the effort required to locate information spread across a large organization. Human judgment still determined whether those tables contained the right evidence for the analysis.
→ The analyst found and queried relevant internal data more efficiently without surrendering responsibility for data selection.
Common mistakes
Uploading data to an unapproved tool
Convenience does not override company security policy. Sensitive data should remain inside approved enterprise systems with appropriate controls.
Assuming the supplied dataset is sufficient
The agent may draw conclusions from whatever it receives rather than request missing evidence. The analyst must decide which datasets are necessary.
Presenting generated output without validation
AI can make mathematical, interpretive, and presentation errors. Review calculations, charts, causal claims, and recommendations before attaching your name.
Is it for you?
Best for
It is best for analysts and operators who must turn business data into a defensible decision or leadership deliverable under time pressure.
Not ideal for
It is not ideal when the available data is inaccessible, fundamentally unreliable, or too sparse to support the requested conclusion.
From the transcript
“don't take your company's data and upload it somewhere else.”
“have a clear problem that you're trying to basically solve with it.”
“Always go with what you're going to do about it. Don't just go with like the problem and the root cause.”
From the episode
Google Data Analyst Shares Her $300k/year Codex Workflow