A lot of AI career advice ends with two words: judgment and taste.

Most advice stops there, leaving data professionals with nothing concrete to change.

Assume you know your domain and can do strong analytical work. The practical version comes down to three things: (a) choose work tied to a real decision (b) turn useful work into something others can reuse and (c) build trust with the people who act on your recommendations.

1. Choose work that changes a decision

AI makes it cheap to answer more questions, which can fill your backlog with interesting requests that nobody will act on.

My bet is that half of your analysis requests disappear if you force each one to name two things:

Decision: the choice someone needs to make.

Action: what that person may do after seeing the answer.

If you cannot fill in both lines, put the request on hold until someone can explain why it matters.

Use your knowledge of the business to identify who will make the call and what is at stake. Run the analysis only when the answer could change a decision.

Once the problem deserves attention, let AI help with execution and stay responsible for what happens next.

2. Turn useful work into something others can reuse

Save useful analysis as something your teammates can run without you.

Keep the definition and logic, then package them in a dashboard, workflow, or AI agent. AI can use that context to answer the next version of the same question without forcing someone to wait for you or rebuild the analysis.

Your teammates can build on your work while you spend time on a new decision.

3. Build trust with decision makers

Trust is the primary currency of a data team.

Decision makers act on your recommendations when they trust your judgment and know you understand their business.

Use one-on-ones and staff meetings to learn what they need to accomplish. Explain your reasoning in plain language. Be honest about what you do not know. After a decision, ask what happened.

Spend some of the time AI saves on these relationships. Decision makers who trust you involve you before a request reaches the backlog. You can shape the question with context that never makes it into the ticket.

Open your current backlog and add a decision and action to each request. Put work with no business consequence on hold. Turn one repeated analysis into a self-service workflow, then spend the saved time learning what a decision maker needs to accomplish this quarter.

One problem runs through all three: the agent has to know which table matters, what the metric means, whether the result makes sense, and what evidence a human needs before acting.

Note: Every figure below is a self-reported internal result. The companies used different users, questions, and evaluation methods. Treat the figures as clues, not a leaderboard. Each system combines several kinds of context. The labels above show the clearest difference in emphasis.

Same problem, different starting point

OpenAI reads the code

OpenAI’s data platform serves more than 3,500 internal users and spans more than 600 petabytes across 70,000 datasets. Its data agent combines six context layers: table usage, human annotations, code enrichment, institutional knowledge, memory, and live warehouse inspection. OpenAI’s write-up explains each one.

Code enrichment is the interesting part. Codex reads the code that produces a table and recovers assumptions, freshness logic, and business intent that may never appear in the schema.

The goal is to answer two very analyst questions: “what’s in here” and “when can I use it?”

Anthropic governs the meaning

Anthropic starts with a semantic layer owned by humans, then uses skills to steer Claude toward the right source and analysis process.

In Anthropic’s internal evaluations, accuracy did not exceed 21% without skills. With skills, aggregate accuracy stayed above 95%.

Then the documents went stale.

Offline accuracy fell from roughly 95% to roughly 65% over a month. Anthropic now updates the skill document in the same pull request that changes the data model.

The definitions have to ship with the data.

Meta starts with the analyst

Meta’s warehouse has millions of tables, but an individual analyst usually works within a few dozen.

Its agent uses personal query history and offline summaries to bound the working domain. Shared knowledge comes through Cookbooks, Recipes, and Ingredients, including semantic models, documentation, snippets, and memories.

Meta’s starting question is personal: what does this analyst already work on?

Old SQL can point. It cannot decide.

Meta found that 88% of data-scientist queries used only tables that person had queried in the preceding 90 days.

Recent history can narrow the search space.

Anthropic tested something else. It gave Claude raw access to thousands of old dashboard, transformation, and notebook queries. Accuracy moved by less than one percentage point. For about 80% of the missed questions, the answer was somewhere in that corpus.

The findings answer different questions.

Meta measured how well recent table use describes an analyst’s working set. Anthropic measured whether a large raw query archive improved answer accuracy.

OpenAI uses query history as one of six context layers, then combines it with code and human annotations.

These are separate internal findings, not a benchmark comparison.

Old SQL can show the neighborhood. It cannot choose the official metric. Someone still has to turn repeated patterns into approved definitions and trusted sources.

The shared loop

My synthesis of the three systems is simple:

QUESTION → CONTEXT → QUERY → EVIDENCE → CORRECTION

OpenAI and Meta describe agents that run SQL, inspect the result, and adjust. Anthropic puts more weight on governed sources and verification before the answer reaches a user.

All three keep the evidence close:

• Meta places SQL front and center with each data point.

• OpenAI summarizes assumptions and execution steps, then links the underlying results.

• Anthropic shows the source tier, freshness, and owner in a provenance footer.

They also preserve corrections.

OpenAI saves memories. Meta adds reusable team knowledge through Cookbooks. Anthropic turns stakeholder corrections into documentation changes and evaluation cases.

A correction should change the next run.

A useful place to start

Pick one question that returns every week: “Why did conversion drop?”

Give the agent four things:

• an approved metric definition

• a trusted source and query path

• the evidence required beside the answer

• a place to save the correction

Run it again next week. If someone repeats the same filter, caveat, or source choice, the system did not learn.

This is the design I am building into ChatData: approved metric context, proof receipts, review state, and reusable answer paths.

A trusted answer should make the next answer faster to produce and easier to defend.

AI can do the analysis. You own the answer.

Most companies are not behind because they lack artificial intelligence tools.

They are behind because their learning does not compound.

A data analyst asks ChatGPT for help writing a database query.
A product manager asks Claude to summarize a dashboard.
A data lead uses artificial intelligence to clean up metric docs.

Useful.

But the company still starts over every time someone asks:

“Why did this number change?”

That is the gap.

An AI-native data team is not a team where everyone uses chatbots.

It is a team where every trusted answer makes the next answer faster, safer, and easier to reuse.

There are four levels.

Level 1: one person gets faster

This is where most teams start.

People use artificial intelligence for:

writing database queries
summarizing dashboards
cleaning up metric definitions
drafting analysis
checking logic before a meeting

This is worth doing.

But the knowledge is private.

It lives in one chat, one notebook, one analyst’s head, or one meeting thread that disappears after the decision.

The test is simple:

Can someone else run the same task next week and get the same standard of answer?

If not, the team is still at level 1.

Level 2: repeated work becomes an agent

At level 2, the team stops treating artificial intelligence like a helper for random tasks.

It gives agents repeatable jobs.

A query review agent checks whether the logic makes sense.
A metric definition agent checks the owner, grain, filters, and caveats.
A dashboard quality agent checks whether the chart matches the trusted source.
An experiment readout agent checks whether the conclusion follows from the setup and the data.

This is the move from prompt to workflow.

The agent has a job.
It has inputs.
It has tools.
It has a review standard.

This is where teams start to feel fast.

It is also where they start to create new problems.

Because a fast agent with weak context can produce confident answers that disagree with each other.

One agent uses the product team’s definition of activation.
Another uses the sales team’s definition.
One checks the dashboard.
Another checks the raw database table.
One remembers the caveat.
Another misses it.

Now you have more output, but less trust.

That is not leverage.

That is cleanup work with better branding.

Level 3: the company stops re-teaching the same facts

At level 3, agents stop working from isolated prompts.

They pull from shared company context:

approved metric definitions
trusted data sources
dashboard lineage
business rules
known caveats
owner notes
prior accepted answers

This is where artificial intelligence analytics starts to compound.

An analyst clarifies a metric once. Future agents can reuse it.

A data owner approves a caveat once. It stops getting lost in chat.

A recurring business question gets answered with proof once. The next answer starts from that approved path instead of rediscovering the logic.

This layer is not glamorous.

It is also the difference between “AI helped me” and “the company learned.”

Without it, you do not have an AI-native data team.

You have faster guessing.

Level 4: answers improve the system

Level 4 is the real shift.

The answer is no longer the finish line.

A business question creates an answer.
The answer includes evidence.
A human reviews or corrects it.
The correction updates the definition, dashboard, checklist, or owner note.
The next agent starts from the improved version.

That is a closed loop.

The data team stops acting like a request queue.

It becomes decision infrastructure.

Not because it has more dashboards.

Because the company gets better at asking and answering the same important questions.

The mature system tracks things like:

Was the right definition used?
Was the right source checked?
Was the evidence strong enough?
Was the answer accepted, corrected, saved, or reused?
Did the correction improve the next answer?
Did the system ask for clarification instead of guessing?

That is the new scoreboard.

Not how many charts were viewed.

How many trusted answers were created.
How many were corrected before they caused damage.
How many became reusable.
How many repeated questions stopped interrupting the data team.

The hidden failure

Many companies will think they are at level 3 because they have a metric catalog, a semantic layer, or a folder full of documentation.

That is not enough.

A definition sitting in a doc is not company memory.

A dashboard nobody trusts is not a source of truth.

A chat answer that solved the problem once but never became reusable is not leverage.

It is residue.

The compounding asset is the reviewed answer path.

For every recurring question, the system should know:

what the metric means
which source to trust
which filters matter
which caveats change the answer
who owns the definition
what evidence makes the answer reusable
when the agent should ask a clarifying question
when the agent should send the answer to a human for review
when the agent should refuse to answer

That is how a data team scales without turning every AI answer into another review burden.

The 30-day move

Do not start with the whole data warehouse.

Start with one recurring question.

Something like:

“Why did self-serve conversion drop last week?”

Then build the loop around it.

Week 1: write the exact question, the metric, the decision it supports, and what would make the answer wrong.

Week 2: build one narrow agent with a clear job and review standard.

Week 3: create the context packet: definition, grain, source, owner, caveats, and examples of accepted answers.

Week 4: close the loop. Save the evidence. Save the answer path. Turn corrections into future checks. Update the docs and dashboards.

That is enough to show whether the company is serious.

The test is not:

“How many agents do we have?”

The test is:

Can the business ask better questions, get trusted answers faster, and make the next answer better because of what it learned this time?

If yes, artificial intelligence is compounding.

If no, it is just a prettier dashboard help desk.

What level is your company at right now?

And what is the biggest blocker between where you are and level 4?

This post was originally written in Sep 2025 (here), and I am cross-posting here since I had a conversation just yesterday which made it obvious that this transformation is accelerating.

I’ve been watching this mess unfold for years now.

Data Engineer, Analytics Engineer, Business Intelligence Developer, Data Analyst, Business Intelligence Analyst, Decision Scientist, Applied Scientist, Research Scientist, ML Engineer, MLOps Engineer, Data Scientist, Quantitative Analyst, Product Analyst, Growth Analyst, Marketing Analyst...

Stop.

We’ve created a job title circus that serves nobody. Not companies. Not professionals. Not the industry.

And it’s about to collapse.

I’m calling it now: By 2027, this chaos consolidates into 2 clear job families. The companies that figure this out first will dominate. The professionals who adapt early will thrive. Everyone else is going to get left behind in the confusion.

Why My LinkedIn Comment Exploded (And What It Tells Us)

last week, I commented that we need to go from ~15 job families to 2 in data. That comment exploded (in a good way).

Why? Because everyone knows it’s true, but nobody wants to say it out loud.

Here’s what’s really happening:

Companies are drowning. They post jobs for “Senior Data Analysts with ML experience and Python skills who can build data pipelines and present to executives.” That’s not a data analyst. That’s three different roles mashed together because nobody knows how to define what they actually need.

Professionals are stuck in title limbo. I see resumes with 5 different job titles for the same work. “Data Scientist” at one company equals “Business Analyst” at another. It’s complete madness.

Career paths are broken. How do you grow from “Analyst” to “Principal Data Scientist”? Nobody knows because we made up half these roles in the last 5 years.

Salaries are all over the place. Same work, different titles, $50K gaps. The market can’t price what it can’t define.

The chaos is unsustainable. And market forces are already fixing it.

The AI Reality Check

Everyone’s debating whether AI will replace data analysts. Wrong question.

The right question: Which data roles add unique human value when AI handles the routine stuff?

Here’s what I’ve learned watching AI transform workflows at multiple companies:

What AI eliminated:

• Manual data cleaning and prep

• Basic visualization creation

• Standard reporting and dashboards

• Simple statistical analysis

• Code debugging and optimization

What AI amplified:

• Strategic problem solving with business context

• Complex decision maker management and communication

• Creative experimentation and hypothesis generation

• Systems thinking and architectural decisions

• Ethical AI governance and bias detection

The result? Most data roles are converging into two categories based on what humans uniquely provide.

The 2 Job Families That Will Survive

After analyzing hundreds of job descriptions and actual work responsibilities, the consolidation is obvious:

Data Builders 🔧

Implementation & Data platform Specialists

What they do:

• Build and maintain data systems that work at scale

• Implement ML models and automation workflows

• Ensure data quality, security, and performance

• Create tools that enable everyone else to succeed, including AI

Current job titles merging here: Data Engineer, Analytics Engineer, MLOps Engineer, Platform Engineer

Why they matter: These are the people who make data work. They’re the builders.

Data Strategists 🧠

The Insight and Decision Specialists

What they do:

• Solve complex business problems using data and AI

• Drive strategic decisions through analysis and experimentation

• Translate between technical capabilities and business needs

• Design experiments that move key metrics

Current job titles merging here: Data Scientist, Business Analyst, Product Analyst, Decision Scientist, Research Scientist

Why they matter: These are the people who turn data into business impact. They’re the thinkers.

Why This Consolidation Is Inevitable

Follow the money.

Companies are tired of managing 15 different data roles with overlapping responsibilities. They want clear hiring criteria, predictable career paths, efficient team structures, and measurable impact.

The consolidation isn’t coming. It’s already here.

Smart companies are quietly restructuring their data teams around these two core functions. The job boards just haven’t caught up yet.

I’m seeing startups hire “Data Engineers” and “Data Scientists” with genuinely distinct responsibilities. Tech companies organize around “Data Platform” and “Data Science” teams with clear interfaces. Traditional enterprises finally invest in both infrastructure and strategy capabilities separately.

What This Means for Your Career Right Now

If you’re currently infrastructure focused:

• Double down on systems thinking and architecture

• Master AI/ML deployment and monitoring

• Focus on automation and scalability

• Understand business impact of technical decisions

If you’re currently strategy focused:

• Develop deeper business and industry knowledge

• Master experimental design and causal inference

• Become exceptional at communication and influence

• Learn to leverage AI tools for analysis

If you’re somewhere in between:

Pick a side as an IC. The generalist “data analyst who does everything” role is disappearing fastest.

Learn both as a data leader.

The Uncomfortable Truth

Most data professionals cling to outdated role definitions because change is scary.

But here’s the thing about market consolidation: You can be part of it, or you can be disrupted by it.

The professionals who recognize this shift early will have their pick of opportunities. The ones who keep debating whether they’re a “Senior Business Intelligence Analyst” or a “Principal Marketing Data Scientist” will get left behind.

What Happens Next

2025: Early adopters start restructuring teams

2026: Consulting firms publish best practices (catching up to reality)

2027: The consolidation becomes industry standard

By 2030: We’ll look back at our current job title chaos like we look back at having 47 different types of “webmaster” in the late 90s.

The Bottom Line

This consolidation isn’t a prediction. It’s already happening.

The question isn’t whether this will occur. The question is whether you’ll adapt proactively or get dragged along reactively.

The future belongs to data professionals who understand that impact matters more than job titles.

Impact comes from doing one of two things exceptionally well: Building data systems that scale, or using data to drive business results.

Everything else is noise.

Originally written in Sep 2025 on Paras’s linkedin newsletter.

Every few weeks, someone says semantic models are dead.

The argument usually sounds something like this: AI can query raw data, figure out joins, and answer questions in plain English, so why spend time modeling metrics at all?

I get why that sounds appealing.

But at least from where I sit right now, it is the wrong conclusion.

AI is absolutely making analytics faster. What it is not doing is magically creating shared business meaning.

That is still the hard part.

A human analyst can look at messy data and bring judgment to it. They know which revenue definition leadership actually uses. They know which table is technically available but operationally wrong. They know which segment should be excluded, which timezone matters, and which “right” answer will still cause confusion in the room.

AI does not automatically inherit that judgment.

So if you put AI on top of raw tables, fragmented dashboards, and inconsistent KPI logic, you do not eliminate confusion. You scale it.

That is why I do not think semantic models are dead.

If anything, AI makes them more important.

Because the bottleneck is no longer query writing. The bottleneck is trust.

A semantic model is not just BI plumbing. It is the layer that tells both humans and machines what a metric officially means. It creates one governed definition, reusable logic, and a more stable path between a business question and a trustworthy answer.

Without that layer, AI can still answer quickly. It just cannot answer consistently.

And in analytics, inconsistency is what kills trust.

The real risk is not only hallucination. It is metric drift.

One prompt pulls from one source. Another prompt pulls from another. A dashboard says one thing. A spreadsheet says another. An executive asks a simple question, and suddenly the meeting turns into a debate about definitions instead of a decision about the business.

That is not an AI problem in isolation.

That is a definition problem.

Semantic models help reduce that ambiguity. They give the business, the data team, and now the AI layer a shared contract for what counts.

To be clear, I do not think a semantic model solves everything.

You still need business context, caveats, approved answer paths, and operating judgment around how metrics should be used. But that is exactly the point: semantic models are not the whole solution, yet they are still a foundational part of the solution.

So my current POV is simple: AI makes natural-language analytics easier. It does not make semantic discipline optional.

The teams that win will not be the ones with the cleverest prompts. They will be the ones with the clearest definitions, the strongest trust layer, and the discipline to keep AI grounded in governed business meaning.

POV is mine; AI helped shape the final draft.

After interviewing hundreds of data professionals in past 12 months, I’ve noticed clear patterns separating candidates we move forward with from those who don’t get the call. It’s not always about credentials, it’s about how you show up in the process itself.

Here are the 9 tactics that candidates we actually hire demonstrate, organized around three core qualities: agency, accountability, and adaptability.

AGENCY: Take Ownership of Your Search

Tip #1: Build Something Real

Don’t polish your portfolio forever. Candidates who move forward build a functional project this week that solves a real problem, even if it’s rough. Use a public dataset, ship a GitHub repo, deploy a live dashboard. Employers want evidence you can execute, not just interview well.

Tip #2: Show Up With Opinions

Generic interview answers don’t stick. The candidates we hire come prepared with specific perspectives: “Here’s how I’d approach your churn problem using causal inference rather than standard cohort analysis” or “I’d restructure your metrics layer using dbt semantic models.” You don’t need to be right, you need to demonstrate you think.

Tip #3: Make Yourself Visible Before the Interview

Engage publicly. Comment thoughtfully on LinkedIn posts from data leaders. Write one substantive post about an industry trend. When I interview someone whose thinking I’ve already seen online, they’re immediately credible. You’re not building fame, you’re building proof that you know what you’re talking about.

ACCOUNTABILITY: Prove Impact With Numbers

Tip #4: Quantify Every Win in Your Resume

Replace “Led analytics initiatives” with “Built attribution model that optimized $75M in annual marketing spend, reducing CAC by 18% and accelerating payback period from 16 to 9 months.” Candidates I hire share >5 specific, quantified impact stories.

Tip #5: Own Your Failures

In interviews, when asked about mistakes, candidates we advance take real accountability: “I missed the forecast by 15%. Here’s the root cause, what I should have done differently, and the monitoring system I built to prevent it from happening again.” We’re hiring for judgment and accountability, not perfection.

Tip #6: Prepare Your Impact Narrative

Before any interview, write a one-page summary of 3 major projects with: (1) the business problem, (2) your specific contribution, (3) the measurable outcome, (4) what you learned. Reference this throughout the interview. This shows you think in outcomes, not just outputs.

ADAPTABILITY: Show You’re Built for Tomorrow

Tip #7: Demonstrate You Can Learn Fast

The candidates we hire don’t say “I’m learning AI”, they say “I shipped a multi-agent system using Claude.” Spend this week building something with a modern AI tool. Deploy it. Show you can move quickly in an evolving ecosystem.

Tip #8: Highlight Your Pivots

If you’ve evolved your role, own it confidently. “I was hired as an analytics engineer but recognized the team needed data infrastructure, so I taught myself Airflow and restructured our pipeline architecture” shows adaptability and initiative. Companies in 2026 need people who can wear multiple hats.

Tip #9: Know Where the Industry is Heading

Candidates we hire reference new frameworks, tools, and industry trends in conversations. They mention semantic layers, causal inference, or agent-based architectures casually, not as jargon, but as genuine familiarity. Show you’re reading, listening to podcasts, and thinking forward.

Put It Into Practice

Before your next interview:

✓ Agency: Build something tangible. Ship a real project.

✓ Accountability: Quantify your three biggest wins. Know your numbers.

✓ Adaptability: Learn one new AI tool. Show proof you built with it.

The difference between “qualified” and “we’re hiring this person” often comes down to these signals, especially in an AI-native environment where the rules are still being written.

Live 1:1 Practive

For a limited time, I am offering live 1:1 mock interviews (all proceeds are donated). https://topmate.io/parasdoshi

Disclaimer: This post was edited by AI, but the core ideas, framework, and thinking were fed to the AI by Paras

Which of these 9 resonates most with your job search right now? DM me your biggest challenge. I’m here to help.

Session Overview:

Data isn’t the hard part. Sustaining its impact is.

Too many data teams succeed in shipping dashboards but stall when it comes to lasting influence. Why? Because impact doesn’t scale without intention.

At the next Inner Join, I’m hosting Paras Doshi, Head of Data at Opendoor, to unpack a field-tested framework for building data-led organizations that last.

He calls it the 3 Ps framework:

1. People: How to centralize data culture while empowering teams
2. Platform: Why discoverability, trust, and governance need to be built in—not bolted on
3. Process: Moving from service desk to product mindset (with examples)

 Plus: How Paras’ team turns ad-hoc requests into repeatable assets—and measures long-term data value.

—-

Additionally, I have written about centralized vs de-centralized data teams here if you want to double click into structuring data org’s.