Why an analyst needs an AI agent, not just ChatGPT

Analysts adopt AI faster than almost any other profession. And yet almost none of them get real value out of it. It sounds strange, but that is how it is. Almost every analyst already keeps a chat window open to drop in some SQL, ask it to fix a query, make sense of someone else’s code or sketch out a formula, and by various measures most of them do exactly this.

But that is precisely why it is so easy to miss a far more important shift. Using a chat and genuinely changing the way you work are two very different things today, and only a few have mastered the second. Even with ChatGPT open at hand, the analyst still pulls the data, cleans it, checks it against the metric definitions and assembles the report by hand. The chat suggests a query, and the person does all the rest of the work manually. Whereas someone who has moved from a chat to an agent that goes into the databases itself and closes the task end to end is already working at a different pace.

In this piece we will look at where the profession is heading according to major research, what an agent actually is and how it differs from an ordinary chat, and which specific analyst tasks it can already handle.

Where the market is heading

By recent 2026 labour-market surveys, data analytics is one of the most AI-saturated professions there is. Analysts pick these tools up faster than almost any other non-technical specialists. You would think, there they are, the leaders of the transition. But the most interesting part is that mastering a chat and genuinely changing the way you work are, once again, not at all the same thing.

The point is that “using AI” today means two different things, and the difference between them is enormous. The first format is conversational. You open a chat, paste in a chunk of code or a data pull, explain the context, get an answer and carry it back to yourself by hand — into your notebook, into a query, into a report. The second format is agentic. You simply set the task, and the system goes into your data itself, finds the right table, writes and runs the query, calculates by your metric definitions, builds a chart in your style and hands over a finished result that all you have to do is check. The first is a slightly smarter search engine and prompter at hand; it barely changes your real speed, because you still do all the switching between systems yourself. The second changes it radically. Routine that used to eat up half a day starts taking minutes. And this is where the real picture of the market becomes visible.

A recent major study run by the best American universities together with OpenAI, “The Shift to Agentic AI: Evidence from Codex” (June 2026), measured this directly. Data & Analytics is one of its fastest-growing functions; in the pace of agent adoption analysts keep level with engineers. But the key thing is visible there too. Even among analysts, one of the most AI-advanced professions, true agents that do the work themselves account for only about 15% of all AI work on average. The remaining roughly 85% is that same back-and-forth with a chat. In other words, the mass “yeah, I’ve been using AI for ages” and the real gain are two different camps today. The gap is visible in output too. Among researchers and analysts inside OpenAI who moved to agents, the volume done per month grew tens of times over what it had been six months earlier, and at the same amount of time spent. And that means you can do more and take on more complex and ambitious tasks, grow faster both in money and in the level of your work, and, if you want, look for a stronger position.

Employers are already building this shift into their requirements. The ability to work with AI is increasingly spelled out directly in analyst job postings, and specialists with this skill on average cost noticeably more than colleagues in the same role without it. At the same time, few have truly mastered it at the level of an agent rather than back-and-forth with a chat. And it is exactly in that gap that the opportunity lies. Figuring it out now means both securing yourself for the future, when it becomes the norm for everyone, and already today gaining an edge in speed and in money.

It comes down to a simple thing. The market already needs analysts who can do more than just ask a chat — who can work with an AI assistant and tune it to themselves. There are still few of them. Whoever learns this stops shovelling routine by hand and grows into the one who sets the tasks and checks the result. So the whole question is who moves into this camp first. And the good news is that getting there is far simpler than it seems. Exactly how, we will look at next.

What an agent actually is and how it differs from ChatGPT

If you strip away the complicated words, an agent is a program that does not answer a question but performs a task. The difference is roughly like the one between a consultant and a doer. The consultant tells you what needs to be done, and you still do it all yourself. The doer takes it and does it, and you check.

An ordinary chat waits for you to bring it all the context yourself. You copy the code or the data pull, explain what the table is and what needs to be calculated, and then carry the answer back to yourself by hand. An agent is granted access to the necessary tools once, and from then on it opens the right table itself, writes the query, runs it, builds the chart and hands over the result. All that is left for you is to check and confirm.

It works fairly simply. The agent runs in a loop. It got the task, figured out what to do as the first step — say, read a table or run a query — did it, looked at the result, moved on to the next step, and so on until the task is closed. There is no need to dig into this, exactly as there is no need to understand how a mail server works in order to send an email.

There is one important detail, and it is simpler than it seems. For the agent to calculate not at random but by the rules of your particular work, you do not need to train it by hand and spell everything out in text. Setting it up is more of a joint effort. You hand it the documentation for your tables, the metric definitions, examples of past queries, your report templates and the way you usually lay out charts, and it figures all of this out itself and assembles a set of rules it will follow from then on. You check this and correct it where needed. From then on these rules are reused in every task, and any request is handled exactly the same way, whether you launched it yourself or a colleague standing in for you did.

Which tasks it already handles

If you look at what such agents actually do for analysts today, a few typical scenarios take shape. Below are the most common ones, just as examples, so it is clear what we are talking about. In reality you can automate almost any repetitive piece of work this way; these are taken for illustration.

The most common story is the stream of small requests from everyone around. The product manager asks you to pull the conversion by segment for the week, marketing asks how many came from the latest mailing, the manager wants a figure for yesterday’s sales broken down by region. On their own each such request is five to ten minutes, but there are many of them, they tear the day into pieces and leave no room for a proper task. An agent that has access to your data and remembers how your metrics are calculated closes such questions itself. You write to it in plain language, and it goes into the database, writes and runs the query, checks the figure and builds a finished chart. What is more, it can work ahead. For example, it saw a request from the product manager in the mail itself, pulled the needed data, put together the export and wrote to you that here is an email that came in, the answer is already ready, take a look and send it. All that is left for you is to confirm. And a simple question can be both asked and answered right from the messenger, without opening the laptop.

The second familiar scenario is the first pass over a fresh dataset. Every time it is the same thing. Load the file, check the types, look at where the missing values and outliers are, build a couple of distributions to understand what you are even dealing with. This can be handed to the agent in full. It picks up the file itself, chooses a tool suited to its size — whether ordinary pandas on a small pull or something more powerful on a big one — runs the basic checks and hands over a short summary of what is worth a closer look. And if you or your team already have an established order or a ready playbook for such a review, you can hand it to the agent once, and from then on it will run every new dataset exactly by it. You start not from a blank page but with the data already sorted through.

The third scenario, and probably the most rewarding in terms of time, is the regular reports you have to assemble by hand from several sources. One user described how they turned an agent into an analyst for a weekly report. They handed it the documentation for their tables; the agent worked through it and even asked questions about the columns it did not understand. They gave it a list of typical requests in plain language and a couple of examples of finished SQL from history, so the agent could see how things are written here. They gave it a guide to laying out charts and links to several past reports as a template of the structure. In essence the person handed the agent all the context once, and together they broke down which steps the report consists of — the very ones the analyst would go through by hand every time. Now the agent goes through them itself, assembles the report at the set time and sends it to whoever needs it. The only important thing is that before sending it is still worth looking over the figures yourself, because on complex dashboards the agent might miss something. More on this below.

And the fourth scenario, the most telling, is when the agent hands over not just a number but a finished conclusion. Usually a data pull answers the question of “how much” but not the question of “and what to do about it,” and interpreting it further falls to the analyst. An agent can be set up so that it does not just collect data but holds in mind the whole context of your company, which can be loaded into it. Then it does not just calculate but looks for relationships in the numbers itself, says what it found, and immediately suggests what this might affect in the business and what should be done with this data next. The requester receives not a table that still has to be interpreted but a finished, meaningful answer, and this way you can take part of the work even off the product manager and other people. In essence you get a junior analyst with good skills, both in code and in analytics itself, to whom you can hand a large chunk of repetitive routine, while you plug in more on the checking and the decisions and take on new and interesting things.

And the most valuable part is that these scenarios only get better over time. You set a scenario up once, and from then on it is, in essence, a written procedure of your work that is reused every time and can be gradually improved. For example, in one team the analyst’s process was written into such a scenario, and now any junior who joins runs exactly the same checks and the same steps and delivers the quality a senior analyst would give themselves. The approach to the data stops living in one person’s head and becomes available to the whole team. You figure out once how you work, and from then on it works for you constantly, without having to explain everything from scratch each time.

What the limitations are

Two honest caveats are needed here. The first is that AI sometimes gets it wrong, and confidently at that. It may take the wrong table, mix up similar metrics, or, if access to the data dropped for a second, substitute a plausible but made-up figure just to produce an answer. That is why the agent is a first-pass doer, not the final authority. It takes on the routine and the volume, but the final figures a decision is built on are checked by the analyst. In the setup the agent is told directly to calculate only from the available data, to flag where it is unsure, and to make nothing up.

The second caveat is confidentiality. Data pulls with internal and client data must not leak into someone else’s cloud, so the files and access are more sensibly kept with you. That way you get the agent’s speed and at the same time control where the data sits yourself.

Keep these two rules in mind — a human checks the figures and the data stays with you — and the agent stops being a risk. It is simply a faster way to do the same work with the same responsibility for the result.

How to start using this

You can build such an assistant yourself. But it is not only a matter of spending an evening on the setup. You also have to figure out how it all works, what to connect where, how to describe the scenarios, try it, make a mistake somewhere and redo it. This takes time, and not every analyst has the desire to dive into it, especially since there is usually enough of their own work as it is.

That is why there is a simpler path — take a ready, pre-configured solution. There are solutions like kvelo.dev that already work as agents and assemble a personal workspace for you, where everything needed, including typical analyst scenarios, is already tuned and set up automatically, without long manual fiddling. From then on you simply work. The complex tasks you do at the computer, and the simple questions you solve right from the messenger. You write “calculate the conversion by segment for the week” in Telegram or WhatsApp, and the agent went into the data, calculated by your rules, built a chart and sent the answer. In essence you talk to it like a colleague.

The market, by and large, has already decided that working with AI in analytics is the new norm. The difference between those who stay in back-and-forth with a chat and those who learn to delegate to agents will soon be visible in speed, in the number of tasks closed and in who on the team gets the interesting work and who gets the endless data pulls. That is why it is worth starting to figure this out already now. There is still time, and the chance to be among the first is still large. And to figure it out as simply as possible and immediately build it into your work, you can start with a ready solution like Kvelo and just give it a try.