03 June 2026

Google Data Analyst Shares Her $300k/year Codex Workflow

1Frameworks

10Insights

Frameworks in this episode

Strategy6 steps

Human-Governed Agentic Analytics Workflow

Direct AI analysis with secure data, clear questions, and human validation

Read →

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Myth Buster· 2

Myth Buster14:00

AI Will Analyze Incomplete Data Without Asking for More

An AI system may assume that the files it receives define the full analytical universe. It can confidently form conclusions from incomplete evidence, so the human analyst must identify what information is absent and obtain the correct datasets.

The model may not recognize that additional data exists
Incomplete inputs can still produce confident conclusions
Selecting the right evidence remains a human responsibility

“You give it like the missing data, it's going to make conclusion based off that, too.”
Sundus · 14:30

“It just assumes that it doesn't have the option to get additional data set from you.”
Sundus · 14:30

#data quality#missing data#ai limitations

Myth Buster27:30

AI Does Not Make Dirty Data Safe to Analyze

Codex initially behaved as though the supplied file was already clean. Real-world datasets rarely are, so missing or incorrect values must be found, cleaned, and normalized before the resulting analysis can be trusted.

AI may assume the input is analysis-ready
Missing values can distort conclusions
Cleaning and normalization remain essential
Real-world data is rarely perfect

“if your data has missing values or if it's incorrect, then your the answers that you're going to get is incorrect.”
Sundus · 28:00

“the data is perfect, which in real world scenario it's never really perfect.”
Sundus · 28:00

#data cleaning#normalization#data quality

Hot Take· 1

Hot Take28:30

Agentic Analytics Will Become a Core Workplace Skill

Sundus expects AI-assisted analytics to spread across organizations over the next five to ten years. Practitioners who learn it early can increase their productivity, teach colleagues, and become internal leaders while retaining the analytical fundamentals their roles require.

AI-assisted analysis is expected to become mainstream
Early practitioners can differentiate themselves
Teaching colleagues can create internal leadership opportunities
Analytics roles will evolve rather than simply disappear

“5 to 10 years from now, we're going to see more and more of this where this is going to become like a core skill…”
Sundus · 28:30

“your jobs are safe. They're not going away.”
Sundus · 30:00

#careers#agentic analytics#future of work#data skills

Explainer· 2

Explainer03:00

Why Customer Retention Is More Than a Binary Metric

Retention can change through spending, product mix, and partial churn rather than a simple retained-or-lost outcome. That complexity makes retention a useful demonstration case for AI-assisted analysis because the analyst must distinguish several possible behavioral changes.

Customers may remain while spending less or more
A customer can churn from one product while retaining another
Retention analysis requires multiple behavioral dimensions

“Retention is like this complex topic, right? Because it's not just like binary, like I was a customer or was not a customer.”
Host · 03:00

#retention#customer behavior#metrics

Explainer17:30

The Best Analysts Discover the Question Behind the Question

Stakeholders frequently request an analysis that does not correspond to the decision they actually need to make. Skilled analysts probe the request, reframe the underlying problem, and ask creative questions that reveal more useful interpretations of the same information.

Initial stakeholder requests can misstate the real problem
Creative questions distinguish expert analysis
Problem framing can matter more than computation
Analysts should clarify the decision behind the request

“they will ask a question, but when you like dig in a little bit deeper, it turns out like the problem that they're trying to…”
Sundus · 18:00

#problem framing#stakeholders#critical thinking

Story· 1

Story12:00

A Mobile Release Emerges as the Retention-Drop Culprit

The demonstration dataset showed weekly retention falling sharply while mobile crash exposure increased. Codex linked the pattern to a newly launched mobile version, suggesting that crashes interrupted returning customers before they could complete their sessions.

Weekly retention fell substantially
Mobile crash exposure rose during the same period
The suspected root cause was a new mobile app release
The scenario illustrates how operational events can explain metric outliers

“Customer weekly level retention drops from 72% to 46%, which is a drastic drop.”
Sundus · 13:00

“And it's related to new mobile version launch. So, that's our root cause.”
Sundus · 13:00

#mobile#retention#root cause#product analytics

Tool· 2

Tool15:00

An Enterprise AI Layer Can Locate Tables and Draft SQL

Sundus describes using AI at Google to locate data distributed across a large internal environment. By describing the desired dataset and query, she could have the system identify relevant tables and write SQL, addressing one of enterprise analytics’ most time-consuming discovery problems.

Large companies often spread related data across many systems
Internal retrieval context can help AI locate relevant tables
AI can translate a dataset description into SQL
The analyst still decides whether the discovered data is appropriate

“It would automatically figure out which tables the data lives in.”
Sundus · 15:00

“It would automatically write the sequel and finding the data set is one of the hardest challenge when you work at a bigger company where…”
Sundus · 15:00

#sql#enterprise data#rag#data discovery

Tool19:00

Codex Turns a CSV into Excel Analysis and a Leadership Deck

The demonstration produced an Excel workbook with an executive summary, weekly trends, cohorts, and plots, plus a seven-slide leadership presentation. These artifacts provide a useful starting point, but the generated deck still requires visual and factual editing.

Codex created multiple cohort-analysis worksheets
The workbook included trends, metrics, and charts
The agent generated a presentation alongside the analysis
Generated slides were useful but visually imperfect

“One is the Excel file with the retention drop cohort analysis, and the second one is the leadership deck.”
Sundus · 20:00

“Codex doesn't do a great job with decks like they don't look the prettiest”
Sundus · 26:00

#codex#excel#powerpoint#deliverables

Takeaway· 2

Takeaway16:30

Treat the AI Like an Intern, Not an Autonomous Analyst

Sundus compares AI with an intern who can perform coding and numerical work quickly after receiving a defined assignment. A senior practitioner must still review what makes sense, request revisions, and decide which findings are ready to present.

Delegate coding and calculation rather than accountability
Review results collaboratively
Revise unsupported or confusing findings
Human judgment determines presentation readiness

“I'll just give the crunching part and the coding part to my intern.”
Sundus · 17:00

“I need to like sit down with the intern and figure out what exactly makes sense, what doesn't make sense, what we need to revise,…”
Sundus · 17:00

#human oversight#ai collaboration#analytics careers

Takeaway24:00

Positive Outliers Can Reveal Repeatable Growth Opportunities

The cohort output did more than expose a bad week; it also surfaced a week with unusually high retention. Sundus recommends investigating what happened during that period and determining whether a campaign, launch, or other intervention can be repeated.

Look for unusually strong periods as well as declines
Connect metric outliers to operational events
Turn successful anomalies into new hypotheses
Test whether the winning condition can be repeated

“what exactly happened in April 20th week? Why is our retention so high?”
Sundus · 24:00

“Can we investigate what we did in April 20th, and can we repeat it more often?”
Sundus · 24:30

#outliers#growth#hypotheses#retention