Ian Klosowicz
AI can generate a query, summarize a dataset, and build a chart from a prompt. What it can't do is walk into a meeting where two executives disagree about what a metric should measure and figure out who's actually right. It can't notice that the data looks clean but is missing 3 weeks of records because a pipeline broke quietly in November. It can't read the room, ask the question nobody wrote down, or push back on a scope that will produce a misleading result.
Those aren't edge cases. Those are the moments that define whether an analyst produces work that matters or work that gets ignored.
The conversation about AI and data analysts tends to focus on tasks: can AI write SQL? Yes. Can it build a dashboard? Mostly. Can it summarize a report? Easily. That framing misses the point.
The tasks AI handles well were never where the analytical value lived. They were the cost of getting to the analytical value. Faster query generation means analysts spend less time on mechanics and more time on the work that actually requires a human. That shift doesn't shrink the role. It changes what the role demands.
Understanding which parts of the job still require a person isn't just academic. It tells you what to get good at if you want to build a career that holds up.
AI has no idea what happened in your company last quarter. It doesn't know that the sales team changed its commission structure in March and that the resulting spike in deal closures in April isn't organic growth. It doesn't know that the VP who requested this analysis has a habit of cherry-picking the numbers that support a decision already made. It doesn't know that two departments are using the same metric name to mean different things.
An analyst who has been embedded in a business for 6 months knows all of that. That context doesn't live in a database. It lives in conversations, in Slack threads, in the memory of attending the wrong meeting and realizing the numbers being discussed don't match any report you've ever seen.
Context isn't a soft skill. It's the thing that separates analysis that drives decisions from analysis that describes the past without explaining it.
I work in data engineering now, building pipelines in Snowflake and Coalesce. The single most valuable thing I bring to any conversation about data isn't my ability to write a transformation. It's knowing why a particular table looks the way it does and what business event produced that shape. No tool generates that. It accumulates.
Someone asks: what's the conversion rate on the checkout page?
An AI tool will calculate a number. A good analyst asks three questions first:
That third question is the one that matters most. A checkout conversion rate doesn't tell you whether the checkout experience is the bottleneck. It could be pricing. It could be that the wrong traffic is arriving. A metric that answers the wrong question with precision is just noise with a decimal point.
Reframing the question before running the analysis is a skill that requires understanding what the stakeholder actually needs versus what they asked for. AI can't do that because it optimizes for answering the question it was given, not for whether the question is worth answering.
When I mapped out the 90 days of the Analyst Hive curriculum, deciding what order the skills go in was harder than deciding the content. Question framing goes in Month 2, after someone has enough technical foundation to know what the data can and can't show. You can't frame a better question until you understand the constraints of the data underneath it.
AI tools are good at detecting statistical anomalies in structured, well-documented data. They aren't good at recognizing that the data looks fine but is wrong for a reason that requires domain knowledge to see.
A few examples of what that actually looks like in practice:
None of those problems surface in a standard anomaly detection pass. They surface when an analyst looks at a number that doesn't match their mental model and decides to find out why instead of moving on.
That instinct, that something is off, is the product of having seen a lot of data in a specific domain and knowing what it normally looks like. It doesn't transfer automatically. It builds slowly. AI doesn't have it, and prompting it to check for data quality issues only works if you already know what to prompt for.
AI can write a summary of a dataset. It can't write a summary that accounts for the fact that the person reading it is going to see the conclusion first and work backward to find a reason to reject it. It can't decide that the most important thing to lead with is the one number that changes everything, rather than the 4 numbers that contextualize it. It can't read a room and adjust mid-presentation when the question being asked has shifted.
Communication in an analytics context isn't writing. It's translating. The analyst has to take something that's true in the data and convert it into something that's actionable for a person who doesn't spend their day in SQL.
That translation requires knowing the audience. It requires knowing what they already believe, what they're hoping to see, what they'll do if you confirm their assumption versus challenge it. Those aren't things a language model can infer from a prompt. They come from relationship and observation.
The analysts I've watched succeed earliest after breaking in weren't the ones with the best technical skills. They were the ones who could explain what they found in a way that made someone else want to act on it. That skill is learnable. It doesn't come from a course. It comes from practice and feedback.
There's a version of every data team where people trust the numbers and a version where they don't. The difference isn't the data. It's the analyst.
When a business trusts its analyst, they bring them into the conversation before a decision is made. They share context early. They ask whether a metric makes sense before building a strategy around it. When a business doesn't trust its analyst, they use the data selectively and make decisions the data never informed.
Trust in an analyst is built through consistency, through being right when it was hard to be right, through flagging a problem before it became a crisis, through saying something isn't ready when someone wanted to ship it. That track record is a human asset. It can't be delegated to a tool.
An AI assistant can help you produce faster, cleaner, more thoroughly documented output. It can't show up to the quarterly review and be the person who called the trend 6 weeks ago. That credit belongs to a person.
If you're trying to break into data analytics right now, the pressure to learn every tool and match every job description is real. Most of that pressure is aimed at the wrong target.
The things that will keep your career relevant are the things AI can't replicate:
The technical foundation matters. SQL, Excel, a BI tool. You need those to do the job. But they're the entry ticket, not the career. The career gets built on top of them.
I didn't have a degree in this when I got hired. I wasn't the most technically skilled person who applied. I got hired because I could understand what the team needed and communicate that I could deliver it. That dynamic hasn't changed. If anything, it has become more important as the technical parts of the role get easier to automate.
If you want a structured path through what to build, what to learn, and how to run the job search, the 90-day program at Analyst Hive covers all of it in sequence.
What parts of the data analyst job are hardest for AI to replace?
Business context, question framing, data quality judgment, stakeholder communication, and trust-building are the hardest to replicate. These require domain knowledge, relationship history, and the ability to understand what a person needs versus what they asked for. None of that lives in a prompt.
The broader AI replacement question has a more nuanced answer than either side usually gives it and it matters for how you position yourself in the job market.
Can AI help data analysts or does it only threaten them?
It helps more than it threatens, for analysts who understand what it does well. AI accelerates mechanics: query drafting, report formatting, anomaly flagging on clean data. That frees analysts to spend more time on the judgment-heavy work where the actual value is. The analysts threatened are the ones who only did the mechanics.
Will data analyst jobs still exist in 10 years?
Yes, in some form. The role will look different. Analysts will spend less time on data pulls and more time on interpretation, communication, and decision support. The demand for people who can translate data into action for non-technical leaders is structural. It doesn't go away as tools improve.
What should a new data analyst focus on learning given AI tools?
Learn SQL, Excel, and 1 BI tool well enough to validate AI output and work without it when needed. Then invest heavily in communication, business acumen, and the ability to frame the right analytical question. The second category is where careers are built. The first is just the entry requirement.
Is business communication really that important for data analysts?
It's the highest-value skill in the role. An analyst who can't communicate findings clearly produces work that gets ignored. An analyst who can translate data into decisions that non-technical people act on becomes indispensable. The gap between those 2 outcomes isn't technical ability. It's communication.
How do I develop the judgment skills AI can't replicate?
You build them by doing real analysis on real data with real stakes, getting feedback, and iterating. Portfolio projects help because they give you reps. Working in a domain over time builds the mental model of what clean data looks like versus what broken data looks like. There's no shortcut. The reps are the path.
