Many business decisions are delayed not because the data is unavailable, but because only a few people can access and read it quickly. Sales teams want to know why conversions are down, finance is looking for sources of waste, while product teams need to understand which features users are starting to abandon. Typically, all these questions end up in the data analyst queue.
Recent developments indicate a different direction. On September 10, 2026, OpenAI introduced Data agents in ChatGPT Work, which are agents that can connect to company data sources, investigate changes, answer questions, and create interactive dashboards through natural language instructions. Google Cloud is also promoting a similar approach through Conversational Analytics and various data agents for BigQuery and its data ecosystem. OpenAI's announcement and Google Cloud's explanation show that data analysis is moving from specialized tools to capabilities that are closer to everyday workflows.
From Waiting for Reports to Asking Questions
The most easily understood change is in the interface. Until now, someone wanting to answer a business question might have to open a dashboard, select filters, ask for analyst help, or write SQL—the common language used to retrieve data from databases.
With data agents, the questions can look like this:
“Why did the average order value drop in January, while monthly revenue remained relatively stable?”
The agent can then outline the analysis steps: checking order value trends, comparing transaction counts, looking at changes in purchased products, and then linking the results to promotions or marketing channels. This does not mean the agent always finds the correct answer. The difference is that the initial exploration process becomes faster and can be done by more people.
In its announcement, OpenAI stated that Data agents can connect to several sources such as Amazon Redshift, Google BigQuery, Datadog, ClickHouse, Databricks, MongoDB, and Snowflake. Agents can also use documents from Google Drive or SharePoint as additional context. Google Cloud, with a more ecosystem-oriented approach, explains the use of agents for conversational analysis, root cause investigation, and data engineering tasks.
What Changes Is Not Just the Chatbot, But the Way Analysis Works
Regular chatbots primarily help explain or summarize information given to them. Data agents operate on a more operational layer: they need to understand data sources, run queries, read results, select the next steps, and present evidence behind their conclusions.
This is where the term semantic layer becomes important. The semantic layer is the layer that explains the business meaning of data—for example, that “net revenue” is not the same as total transactions, or that “active customers” has a specific definition. Without these definitions, two people could ask the same question but receive different numbers because they are using different formulas or time ranges.
OpenAI mentions that Data agents can use metric definitions, specific calculations, and inter-data relationships established by the organization. Google Cloud also emphasizes the importance of business context and governance so that agents not only produce valid SQL but also answer questions with the correct meaning.
Analysis: in practice, the quality of data agents is likely to be more determined by the tidiness of the data foundation than by the model's capabilities alone. Highly sophisticated models can still produce misleading answers if column names are confusing, data is updated late, or metric definitions change without documentation.
New Risks: Answers Look Convincing
The biggest advantage of data agents—ease—also becomes its risk. When someone can obtain a graph in minutes, there is a temptation to assume that the results are ready to be used as a basis for decisions.
However, some things still need to be checked:
- Is the data source accurate? Sales figures from the transaction system may differ from figures that have already accounted for cancellations or returns.
- Is the time range correct? Monthly comparisons can change if one period uses local time and another period uses UTC.
- Is the cause-and-effect relationship really proven? A drop in conversions occurring simultaneously with a campaign change does not automatically mean that the campaign is the cause.
- Is sensitive data protected? Access to salaries, customer information, or health data should not be broadened just because an agent can run queries.
OpenAI states that administrators can define data connections and roles that are allowed to use them. Queries also follow the permissions of the connected accounts, including restrictions on tables, rows, and columns. This is an important foundation, but not a substitute for audits and testing. Technical permissions answer who can see the data; organizations still need to determine who can make decisions based on that data.
What Does This Mean for Us?
For non-technical users, data agents can reduce reliance on rigid dashboards. Follow-up questions can be asked directly, and results can be developed from the same conversation. For analysts, this does not mean their roles are disappearing. Their work can shift from answering routine requests to designing metrics, checking data quality, and handling more complex questions.
For developers and data teams, the challenge is to create systems that agents can understand without sacrificing control. Documentation of schemas, data catalogs, metric definitions, environment separation, and activity logging will become increasingly important. Agents that can read data without an audit trail are difficult to trust, even if the results appear neat.
What Can Be Done Now
- Start with low-risk questions. Use data agents to summarize trends, search for anomalies, or draft analyses—not to directly change prices, delete data, or execute irreversible actions.
- Document key metric definitions. Write down the meanings of active customers, net revenue, churn, acquisition costs, and other commonly used terms.
- Request evidence behind answers. Get into the habit of checking table sources, data periods, filters, and queries used.
- Separate access based on roles. Do not grant access to the entire data warehouse if the user's needs only cover one domain.
- Test with questions that have known answers. This helps identify definition discrepancies, outdated data, or misinterpretations before agents are used for important decisions.
Data agents are not just a new way to talk to databases. They shorten the distance between business questions and analysis, but they also make data discipline increasingly important. If the foundation is tidy, more people can work with information without always waiting for reports. If the foundation is messy, organizations will only obtain faster answers—not better decisions.
Sources & Further Reading
- OpenAI — Now everyone can put data to work
- Google Cloud — New data agents across the Agentic Data Cloud
- Google Cloud — Agentic analytics with the Data Agent Kit
– Rio Yotto @rioyotto