A sales manager at an automotive parts manufacturer wants to know why a customer’s on-time delivery rate fell this quarter, but has to wait for the IT department to find time. The plant manager wants to compare yield rates for the same part number across three production lines, but someone has to extract the data from the MES. Finance wants to know why the cost of a part increased by two dollars, but first has to ask three departments.
None of these questions is difficult. The hard part is waiting in line for an answer.
On September 10, 2026, OpenAI launched Data agent in ChatGPT Work to address precisely this bottleneck.
What Data agent does
Data agent is a new feature in ChatGPT Work, not a new foundation model. An administrator installs it in the workspace, and users invoke it by entering @Data in a conversation. They can then ask questions about company data in plain language without first learning SQL.
It connects to more than just data tables. According to OpenAI’s announcement, the context needed to understand a question comes from semantic layers and trusted sources, including Databricks Genie Ontology, dbt, GitHub, Snowflake Horizon, and definitions embedded in existing BI dashboards. It can first determine how the company defines revenue and gross profit, then perform the calculation.
For data sources, OpenAI lists approved connections including Amazon Redshift, Datadog, Google BigQuery, ClickHouse, Databricks, MongoDB, Snowflake, and more. Files stored in Google Drive and SharePoint can also be included in the analysis.
The output can be an answer or a live dashboard that users can create and interact with in Omni, Oracle BI, Power BI, Sigma, Tableau, and ThoughtSpot. Companies do not need to rebuild their existing reporting systems.
The permissions model is straightforward: queries inherit the existing permissions of the connected account, including table-, row-, and column-level restrictions. Administrators decide which connections are available and which roles may use them.
OpenAI also shared its own internal adoption figures: nearly every product team and more than two-thirds of its GTM organization use data agents to analyze company data. But it also emphasized an important condition: this works because the data team first established shared business definitions, set access rules, and implemented safeguards for sensitive data.
This article focuses on the second half of that statement: the prerequisite of “what the data team did first.”
Six questions beyond features
In the past, only three or five people in a company knew how to pull the numbers themselves. The queue acted as a filter. Now that 300 people can ask questions at the same time, any order maintained by “nobody knows where that table is” will quickly break down.
Once the barrier to entry falls, six responsibilities return to the enterprise:
- Definitions: Does the entire company use only one formula for this number?
- Boundaries: How will you draw a line that no one has previously defined?
- Accountability: How will responsibility be clarified when something goes wrong?
- Records: Three months from now, will you still be able to determine how the result was calculated?
- Knowledge continuity: What happens when the person who asks the best questions leaves?
- Maintenance: Who will maintain the entire system?
These are the issues that hold things up when an FDE works on-site with a customer to integrate the first group of AI Agent into business processes. None is a technical problem. Each has two parts: people inside the company make the decision, while the FDE asks the questions needed to make that decision possible. Once the decision is made, the system takes over and ensures that it is applied consistently every time.
Question 1: Does the entire company use only one formula for this number?
The gross margin in the sales department’s monthly report differs from the figure calculated by finance by two to three percentage points. Sales subtracts standard cost from the shipment value. Finance subtracts actual cost from recognized revenue. Each department has relied on its own reports for years without incident because the two figures never appeared on the same screen.
Natural-language queries will put them on the same screen. Today, the sales manager asks, “What was the gross profit for Customer A last month?” Tomorrow, finance asks the same question. They receive different answers, and both sides can show their calculations.
In EgentWrX, the company’s approved definitions can be stored in the knowledge base in three forms: policies and operating manuals go into the Wiki, standalone policy provisions are retrieved semantically, and lookup tables requiring exact figures are uploaded as tables for direct querying. External-facing answers can be required to cite their sources. If an answer cannot identify its source, the system rejects it.
Choosing which formula to use is a management decision. Neither is inherently right or wrong. Our recommendation is to bring sales and finance together, put both formulas on the same page, and clarify which decisions each department makes with its number. Management can then decide whether to choose one formula or retain both under distinct names. The decision is written into the knowledge base so that, regardless of who asks, the Agent retrieves the correct result.
Question 2: How will you draw a line that no one has previously defined?
A procurement manager has quotations, lead times, and quality records from three suppliers. Should someone in sales be allowed to ask about the price differences among them?
Most companies have never formally answered this question. Boundaries are maintained through convention: the bid comparison sheet sits in procurement’s shared folder, and more than a dozen people technically have access, but no one thinks to look.
OpenAI’s model allows queries to inherit the permissions already assigned to the connected account. This is a sensible design. By avoiding a separate permission system, it prevents two sets of rules from conflicting. The problem is that inheritance works only if the original permissions are already fully defined. At most manufacturers, folder permissions have accumulated one change at a time over more than a decade, and no one has ever conducted a comprehensive review.
EgentWrX builds boundaries around the organizational structure. Roles have either a “company-wide” or “unit-scoped” view. Department heads can see only their own units and subordinate units. Cross-department knowledge sharing has a separate gate: sharing with people in the same unit takes effect immediately, while sharing with the entire company requires an application and administrator approval.
The gate can hold an application, but management must still decide whether to approve it—for example, by defining supplier bid comparisons as information restricted to procurement and management. Companies often fail to document such decisions. During implementation, we raise these questions one by one, discuss them with procurement, sales, and management, settle them on the spot, and write them as a few clear rules. We then configure those rules as role scopes and define who can access each service endpoint. From that point forward, the boundary is a documented decision.
Question 3: How will accountability be clarified when something goes wrong?
In the past, the cost structure behind a customer quotation was calculated by a cost accountant, reviewed by a section manager, and signed by a manager. If something went wrong, the responsible person could be identified.
Now there is an additional step in the middle: “calculated by AI.” No one’s name is attached to that step. It does not appear in a job description or on the internal control audit flowchart. In practice, this—not the technology—is often what blocks implementation. No one wants to sign a report produced by AI.
This concern should not be argued away. It should be addressed through design. Change the object of approval from “this number” to “this action.” The manager is not signing to certify that “the 23.4% gross margin is correct.” The manager is approving the action: “This cost structure may be sent to the customer.” That is the judgment the manager was already making.
In EgentWrX, approval points are built into the process. When multiple tasks form a workflow, any handoff can be configured to “require human approval before handoff.” The Agent stops at that step and waits for a person to approve it, with default timeout handling if the wait becomes too long.
Internal control must first decide who should grant that approval. Thresholds such as the maximum amount that can be approved and the circumstances that always require manual recalculation are often scattered across approval forms and unwritten conventions. The FDE works with finance and internal control to determine who should sign off on the “calculated by AI” step, then configures that person as the workflow’s approver. The previously unnamed position on the internal control flowchart now has an owner.
Question 4: Three months from now, will you still be able to determine how the result was calculated?
A customer questions the defect rate for a particular batch and asks you to prove how the figure was calculated. Which version of the acceptance criteria did the person use three months ago? Did the data come from the MES or from quality assurance’s own report?
“Seeing the evidence at the time” and “reconstructing the process afterward” are two different things. When AI includes its data sources and calculation method in the original conversation, it helps users trust the answer. An audit requires something else: a record that can still be retrieved three months later and verified step by step.
EgentWrX implements recordkeeping through verification types covering more than 50 resource categories. An administrator can click “Verify Integrity” to recalculate the entire chain on the spot. If the records must be provided to an external auditor, they can be exported as a CSV. The download itself is also logged.
How long records must be retained, whether they need to be provided to external auditors, and which customer contract clause applies are questions answered by your compliance requirements and customer agreements. These decisions should ideally be made before connecting the first data source. The FDE works with compliance, quality assurance, and sales to review customer contracts and audit checklists, turn them into a requirements list, and verify each requirement against the system’s audit and export settings.
Question 5: What happens when the person who asks the best questions leaves?
Three months into implementation, most of the results are concentrated among three people: a production planner, a process engineer, and a younger colleague in plant facilities. They ask good questions because they know which fields cannot be trusted and which machine’s yield figures should be treated with caution.
Then one of them leaves. The data remains, but the person takes away the knowledge of “how to ask.” That knowledge was never stored. Know-how used to live in the heads of senior technicians. Now it lives in someone’s conversation history. The organization is no less fragile.
In EgentWrX, a method discovered by one person can become a skill. The user describes it in text, AI drafts it, and the user first applies it to their own Agent. If it is worth using across the company, the user can propose publishing it. Once an administrator approves it, the skill enters the company’s skill library. When the method changes, only that single copy needs to be updated. When knowledge is uploaded, its scope must be selected. Data scoped to a “unit” belongs to the entire department and does not disappear when one person leaves.
Question 6: Who will maintain the entire system?
At most manufacturers, the IT department consists of one to three people focused on the ERP, network, antivirus software, and machine connectivity. Now they must also answer new questions: Who maintains the data connections? Who reviews requests to add data sources? Who handles complaints that “AI got this number wrong”?
Return to OpenAI’s statement. Its internal adoption rate was possible because the data team first established shared definitions, access rules, and safeguards for sensitive data. That prerequisite matters more than the adoption figure, and most Taiwanese manufacturers do not currently have it.
The question, then, is not “When will we have a data team?” It is “Who will take ownership of this?”
EgentWrX does not require the IT department to maintain another user list. Accounts use SSO, and SCIM can let the company’s existing identity system handle account provisioning and deactivation. When a person is removed from that system, the member list in EgentWrX is synchronized accordingly. Groups in the identity system map directly to platform roles and organizational units, so employees have the correct permissions the first time they sign in.
Implementation can be outsourced. Decisions cannot. The people authorized to add data sources and define metrics must work inside the company. On the first process, the FDE’s role is to complete both tasks once while the designated internal owner observes the entire process. For the second process, the internal owner takes the lead while we observe.
Where EgentWrX is strongest
When the six questions are laid out, it becomes clear that Data agent and EgentWrX operate at different layers. Data agent addresses who can begin analyzing data. EgentWrX addresses what comes next: what action should follow the answer, who is responsible, whether a record is retained, and whether the capability remains when people change.
Building this layer depends on three things.
The enterprise decides where its data resides. EgentWrX can be deployed in the cloud, in the company’s own data center, or in a hybrid architecture where sensitive data remains on-premises while inference runs in the cloud. When connected to internal databases such as ERP and MES systems, the Agent has read-only access. All outbound Agent connections are denied by default; only allowlisted destinations can be reached. In industries where contracts prohibit data from leaving the country, this is a prerequisite for getting started.
Every step leaves a verifiable record. Model calls, knowledge access, system writebacks, checksum chains, and integrity verification are all recorded so that the “calculated by AI” step has a defined place on the internal control flowchart.
Capabilities remain with the organization, not the individual. The company skill library, unit-scoped knowledge, and role-based Agent all support the same objective: when a person leaves, the method remains.
None of these three areas depends on model capabilities. When choosing a platform, the model should be the last question you ask. We have written about this separately: https://www.egentwrx.com/en/perspectives/ai-agent-harness-execution-layer
What if you do not have answers to any of the six questions?
Most companies will reach this point without ready-made answers to all six questions. That is normal. We have all heard the standard advice:
Clean up the data, standardize definitions, and review permissions before discussing AI. But for most Taiwanese manufacturers, that approach is not executable.
The advice itself is not wrong. The problem is that “cleaning up the data” has never been a project at these companies. It is a state with no endpoint. There are thousands of Excel files, seven versions of the same spreadsheet, and critical decision rules stored in the heads of a few senior employees. Organizing everything to the point where it is “ready for AI” would require a dedicated team, two to three years, and a budget with no short-term return. In practice, this advice leaves companies stuck in preparation mode year after year.
We reverse the sequence and proceed in three steps.
Step 1: Choose one process, not one department. Apply three criteria: the scope is clear, the definitions are straightforward, and the consequences of errors are manageable. Examples include procurement notifications for delivery exceptions or customer delivery-date queries from sales. They involve few fields, have clear rules, and someone can immediately judge whether the answer is correct. This is what the FDE does during the first week on-site: interview departments, list their pain-point scenarios, and select the one that best meets the three criteria—not the one causing the most pain.
Step 2: Define small-scale answers to the six questions for that process. You do not need to standardize every metric across the company at once. You only need to decide how “delivery date” is defined in this process, who may query it, who signs off, how long records are retained, who maintains the method, and who owns the rules. For each question, the FDE works with the relevant people until the issue is clear enough for a decision. Once that decision is made, it is configured in the system. The output of this step is not a governance document. It is a working process supported by six signed decisions.
Step 3: Use those decisions as the starting point for the second process. The knowledge, skills, roles, and rules created by the first process already exist in the system as company assets. The second process can reuse them directly. This time, the company’s designated owner takes the lead while the FDE steps back and observes.
This does not mean you should avoid thinking things through. It means some issues become clear only after work begins. Once the first process has run for three months, you will know which fields are creating genuine definition conflicts and which data sources no one is actually maintaining. A paper-based inventory cannot reveal these issues.
For companies without a data team, governance will not emerge before use. It is developed process by process through the combined efforts of the FDE, employees, and the system.
Once the first processes are running smoothly, a new set of questions will emerge. We discuss them here: https://www.egentwrx.com/en/perspectives/ai-agent-governance-after-poc. Until then, start by choosing one process.
FAQ
We already have ChatGPT Enterprise. Do we still need another AI Agent platform?
Start by identifying which part of the process is causing pain. If the problem is that answering a numerical question takes three days and the data already resides in a cloud data warehouse, Data agent directly addresses it. If the problem is that no one takes over after the analysis is complete, the same work must be repeated next month, or the process breaks when the person responsible leaves, then the issue is workflow, authority, and recordkeeping—not query speed. Most companies struggle with both, but to different degrees. Address the more obvious pain point first.
We cannot even extract data from our ERP. Can we still get started?
Yes, but you need a different starting point. For most manufacturers, the first process will not begin with a data warehouse because they do not have one. A more common starting point is integration with existing systems and files used on-site: a connector in the company’s data center can let the Agent read ERP or MES data, or the Agent can directly read Excel files, PDFs, and scanned documents in designated folders. First confirm that the data required by this process can enter the system and that the answer can reach the people who need it. Then decide whether the data should be centralized.
If we start with one process, will we end up with a collection of isolated solutions that are even harder to integrate later?
Yes, if every process defines everything independently. The solution is not to standardize the entire company first. It is to require the second process to reuse what the first process has already defined: the same customer codes, the same definition of delivery dates, and the same permission levels. By the fourth or fifth process, you will have a set of shared definitions derived from real operations. Those definitions will be more accurate than anything produced by a one-time, top-down inventory.
Are there situations where we really should clean up the data before implementation?
Yes. A process should not be the first one selected if the consequences of an error cannot be controlled. Direct customer quotations, regulatory filings, and parameters involving personal safety should not be addressed until the data and lines of responsibility are fully understood. But that is a reason to choose a different first process, not a reason to finish cleaning up all company data before getting started.
References
- Now everyone can put data to work — OpenAI’s official announcement (September 10, 2026). The description of Data agent’s capabilities, examples of data sources and semantic layers, list of BI tools, permission inheritance model, and OpenAI’s internal adoption figures are based on the original information provided on this page.
- ChatGPT Work for data teams — OpenAI’s official product page describing connectors and use cases.