Imagine typing a single sentence into a chat window — “Why did sales drop last quarter?” — and getting back not a vague guess, but a real answer pulled straight from your company’s actual databases, complete with a chart you can click into and explore. That is the promise behind Data agent, a new AI tool from OpenAI built into ChatGPT Work, the company’s enterprise version of ChatGPT designed for internal business use rather than casual chatting.
It sounds like magic. It is also, quietly, one of the more revealing case studies in what “AI agents” actually are, what they can and cannot do, and why the businesses that get the most value from them are not necessarily the ones with the smartest AI — they are the ones that did their homework first.
For anyone trying to build a career, a freelance practice, or a side income around AI tools, this story is worth paying close attention to. It is not really about a chatbot. It is about a widening gap between companies that know how to organize their data and companies that do not — and that gap is exactly where new kinds of paid work are starting to appear.
Meet the AI That Doesn’t Just Answer, It Investigates
To understand why Data agent is different from a normal chatbot, it helps to understand what the word “agent” actually means in AI circles right now. An AI agent is not simply a program that replies to a message. It is a system that takes a goal, breaks it into steps, and carries out those steps on its own — searching, calculating, checking its own work — until the job is finished, without a human walking it through each stage.
Data agent applies that idea to company data. Instead of waiting for the next scheduled report or standing in line for the analytics team to run a query, an employee can open a new chat, type “@Data” followed by a question, and get an answer built from the company’s live information. There is no need to write database queries or learn a new analytics program. Typical questions it is designed to handle include: Why did sales fall? Which costs are rising? Which major client looks likely not to renew? — the kind of questions that used to require a meeting, a spreadsheet, and a few days of waiting.
The One Question That Breaks Every AI Data Tool
Here is the catch, and it is the most important part of the whole story: the quality of the answer has almost nothing to do with how smart the AI is.
Consider a simple example. Ask three departments in the same company what “sales” means, and you may get three different answers. The sales team might count a sale the moment a customer signs a contract. The finance team might only count it once the money has actually landed in the bank. The product team might only count customers who are still actively using the service. If an AI is left to guess which definition applies, it might give an answer that is perfectly correct for one department and completely wrong for another — with no warning that anything went wrong at all.
OpenAI’s solution is to not let the AI guess. Instead, Data agent is built to read pre-agreed definitions that the company itself has set up in advance: business terminology, the exact meaning of each metric, the formulas behind them, and how different datasets relate to one another. This structured layer of definitions is known in the data industry as a Semantic Layer — essentially a translation dictionary that tells any tool exactly which database table a term like “sales” or “active customer” should pull from, and exactly how to calculate it.
Data agent can plug into semantic layers built in several existing systems, including Databricks Genie Ontology, dbt (a widely used tool for organizing and transforming raw data), GitHub, and Snowflake Horizon, as well as dashboards already built in business intelligence software like Power BI or Tableau — two of the most common tools companies use to visualize their data in charts and reports.
The uncomfortable truth here is that Data agent is only ever as smart as the definitions a company gives it. If nobody in the business has ever sat down and agreed on what “sales” actually means, no amount of AI sophistication will fix that. The technical rollout is the easy part; getting humans to agree on shared definitions is the part with no shortcut.
Security First: Why the AI Can’t See More Than You Can
The second thing that determines how trustworthy an answer is comes down to data access — and this is where a common misconception about AI agents gets corrected.
แนะนำ
Data agent can connect directly to databases a company already uses, including Amazon Redshift, Google BigQuery, ClickHouse, Databricks, MongoDB, Snowflake, and Datadog, and it can also pull in documents stored on Google Drive or Microsoft’s SharePoint. But every one of those connections has to be approved by the company’s system administrator first — an ordinary employee cannot simply hook up a new database on their own.
Here is the part worth underlining, because it addresses the fear many people instinctively have about AI agents: the word “agent” tends to make people imagine an AI with its own identity and its own independent access to company data. In reality, Data agent has no special privileges of its own. Every action it takes runs strictly under the permission level of the employee’s own account. If your login can view a customer table but a specific column — say, phone numbers — is locked from your view, Data agent working on your behalf will not be able to see that column either. The system supports permission controls down to the level of individual tables, rows, and columns. In plain terms: the question “could the AI accidentally see something it shouldn’t?” has exactly the same answer as “could the person who asked the question see it themselves?”
From Insight to Action: Dashboards, Approvals, and Handoffs
Getting a first answer is only the starting point. Users can keep asking follow-up questions to dig deeper, and — importantly — they can open up the underlying evidence behind any conclusion the AI reaches. This creates a genuine choice: take the summary at face value and move fast, or spend a little extra time checking exactly which table a number came from, so that when someone in a meeting asks “where did this figure come from?” there is a confident answer ready.
Once an analysis is complete, Data agent can build a dashboard from the results, complete with charts that team members can click into, edit, share, or refresh with new data. Teams that already rely on a business intelligence platform for their morning reports don’t even need to switch to a chat window — Data agent can build or edit dashboards directly inside tools including Omni, Oracle BI, Power BI, Sigma, Tableau, and ThoughtSpot.
From there, the tool can also recommend next steps, flag which team members should be looped in, and send a summary via Slack or email, or trigger an action through a connected tool. Crucially, it will only carry out actions a human has explicitly approved — it does not read data and then quietly start taking action on its own initiative.
The Three-Part Prompt Formula OpenAI Wants You to Steal
Even for readers who will never touch Data agent itself, OpenAI shared a prompting technique worth borrowing for any AI tool. The company offered three sample use cases: diagnosing why a number changed, designing a set of KPIs (key performance indicators — the specific numbers a business tracks to judge success) for a new product, and producing a monthly summary for executives.
What makes these prompts effective isn’t just stating what you want — it’s specifying exactly what elements the answer must contain. A prompt asking why a number changed should also request a comparison against the previous period and a recommendation for what to check next. A prompt asking for an executive summary should request the actual figures, the risks to flag, and a recommended course of action. The formula is simple and transferable to almost any AI chatbot: state the task clearly, then list the specific components the answer must include.
Who’s Already Using It — and the Honest Caveats
According to OpenAI, nearly every product team inside the company already uses Data agent internally, and early enterprise testers in its alpha program include NTT DATA (a major Japanese IT services firm), Thermo Fisher (a global scientific and laboratory equipment company), and ServicePiston (an automotive-industry software provider). It’s worth being clear-eyed here: all of this information currently comes only from OpenAI’s own announcement. There is no independent third-party testing yet, no published pricing, and no confirmed date for general availability.
แนะนำ
Why This Matters Far Beyond Silicon Valley — Including Thailand
It’s tempting to file this under “big enterprise software news” and move on. That would be a mistake for anyone trying to build income around AI skills. The same pattern showing up inside OpenAI’s biggest corporate clients is showing up, at a smaller scale, everywhere digital businesses operate — from e-commerce sellers in Bangkok managing sales across multiple online storefronts, to logistics startups in Ho Chi Minh City, to fast-growing SME clusters across Indonesia. In every one of these markets, the businesses that struggle with AI tools are rarely the ones lacking access to good models. They’re the ones whose sales figures, customer data, and reporting live in three disconnected spreadsheets that don’t agree with each other.
That gap is not a problem AI alone can solve. It is a problem that requires a person who understands both the business and the data — someone who can sit with a small business owner and say, “before we plug in an AI assistant, let’s agree on what ‘a sale’ actually means here.”
The New Income Stream Nobody’s Talking About
This is where the earning opportunity actually lives, and it has nothing to do with waiting for Data agent’s public release. The skill that makes tools like this work — organizing messy business data into clear, agreed-upon definitions — is a service that can be sold today, independent of which AI platform eventually gets used.
Freelancers and consultants who learn the basics of a semantic layer, who can help a small business or agency map out what its core numbers actually mean, and who can then translate that structure into a working dashboard, are positioning themselves for a category of work that barely existed three years ago: AI data readiness consulting. This does not require being a professional data engineer. It requires basic fluency with tools like Power BI or Tableau, an understanding of how businesses typically define metrics like revenue and churn, and the ability to have a structured conversation with a business owner about their numbers. That combination — part translator, part organizer, part AI operator — is exactly the profile that platforms like Upwork and Fiverr are increasingly rewarding with premium freelance rates, and it applies just as directly to local clients in Bangkok or Jakarta as it does to a company in London.
The Takeaway
Data agent is a preview of where enterprise AI is heading: less about clever chatbots, more about connecting AI to the messy, human reality of how a business actually defines success. The lesson for anyone trying to earn a living around AI tools, whether inside a large company or as an independent freelancer serving small businesses, is the same one OpenAI is quietly teaching its own enterprise customers — the AI is never the bottleneck. The shared understanding of what the numbers mean is. Whoever can bridge that gap, for any business, in any market, has found a skill that will stay valuable long after this particular tool has a price tag and a release date.
Key Takeaways
- OpenAI’s new Data agent, built into ChatGPT Work, lets employees ask business questions in plain language and get answers pulled from real company data, plus instant dashboards.
- Its accuracy depends less on AI intelligence and more on whether a company has already agreed on clear definitions for terms like “sales,” stored in a Semantic Layer.
- Data agent has no special access of its own — it only sees what the employee’s own account is already permitted to see, down to individual rows and columns.
- Beyond answering questions, it can build dashboards in tools like Power BI and Tableau, and take approved actions like sending Slack or email summaries.
- The real opportunity for freelancers and consultants isn’t the tool itself — it’s helping businesses organize and define their data before any AI can use it well.
Frequently Asked Questions
Q: What exactly is OpenAI’s Data agent?
A: It is an AI tool inside ChatGPT Work, OpenAI’s enterprise product, that can analyze a company’s real data, answer business questions in plain language, and generate dashboards from the results.
Q: Is Data agent available to the public yet?
A: No. As of this announcement, it is in limited alpha testing with select enterprise customers, and OpenAI has not released pricing or a general availability date.
Q: Can Data agent see all of a company’s data?
A: No. It only has access to whatever the employee’s own account already has permission to view, including any row- or column-level restrictions already in place.
Q: What is a Semantic Layer, and why does it matter?
A: A Semantic Layer is a structured set of business definitions that tells any tool exactly what a term like “revenue” means and where to pull it from. Without one, an AI tool can give technically correct but inconsistent answers.
Q: What databases and platforms does Data agent connect to?
A: It supports connections including Amazon Redshift, Google BigQuery, ClickHouse, Databricks, MongoDB, Snowflake, Datadog, Google Drive, and Microsoft SharePoint, all subject to administrator approval.
Q: Which business intelligence tools can Data agent build dashboards in?
A: It can create or edit dashboards inside Omni, Oracle BI, Power BI, Sigma, Tableau, and ThoughtSpot.
Q: Can Data agent take actions on its own, like sending emails or messages?
A: It can send summaries via Slack or email or trigger connected actions, but only after a human has explicitly approved that specific action.
Q: Which companies are already testing Data agent?
A: OpenAI’s own product teams use it internally, and named alpha customers include NTT DATA, Thermo Fisher, and ServicePiston.
Q: Do I need to know how to code or write database queries to use it?
A: No. The entire design goal is to let people ask questions in normal language without writing queries or learning specialized analytics software.
Q: How is this relevant to freelancers or small businesses outside big tech companies?
A: The underlying skill Data agent depends on — organizing data into clear, agreed definitions — is something freelancers can offer as a service to small and medium businesses right now, independent of any single AI tool.
Q: Is this the same as a general-purpose chatbot like ChatGPT?
A: No. It is a specialized “agent” designed to plan and execute multi-step data analysis tasks tied to a specific company’s real information, rather than a general conversational assistant.
Q: What’s the biggest risk in relying on tools like this?
A: The biggest risk isn’t the AI making things up — it’s a company using the tool before it has agreed internally on what its own key business terms actually mean.
