ChatGPT Just Got the Keys to Your Oracle Data. Who Checked It?
Permissions decide what AI agents can reach, not whether it's right.
Most organizations treat AI agent risk as an access problem. Lock down permissions, log every action, approve the workflows, and the agent will behave.
The larger problem is what sits behind the access. On October 27, Oracle and OpenAI announced that employees will be able to use ChatGPT to retrieve information, run approved business workflows and take actions across Oracle Fusion Applications, NetSuite and Oracle AI Database, with Oracle saying the governance, approvals and controls of enterprise operations stay in place. Those controls decide what an agent may touch, not whether what it touches is right.
That leaves data teams with recurring work nobody scoped: confirming which supplier record, which revenue figure and which customer definition an agent should act on when the ERP, the warehouse and the CRM each hold a different answer.
And it gets worse as more of the business moves into a chat window. Every new plugin, and every app a non-developer builds with Codex, adds another consumer of the same unresolved records, and another place where a wrong answer looks official.
As ChatGPT starts acting inside enterprise systems, agents will also need to find, verify and work with agents beyond their own company walls. That is what we are building at OpenMap: The Agent Network, the connective tissue for the world's AI agents, starting with every person and every local business. We are raising a £5M seed. Meet us at openmap.ai/investors.
Why Isn't "Approved" the Same as "Correct"?
An approval workflow checks who is asking and what they are allowed to do. It was never designed to check whether your ERP and your warehouse agree on what a region, a customer or a cost center is.
Picture a finance analyst asking ChatGPT for third-quarter revenue by region for the board pack. The permission check passes. The agent pulls one subsidiary's figures from NetSuite and another's from Fusion, where "region" was set up by a different team, years apart.
The access was governed.
The answer was not.
This is not an access control problem. It is a data context problem.
The Cost Isn't Bad Answers. It's Bad Answers That Look Official.
Finance leaders already sense the gap. In Workiva's 2026 midyear survey of more than 2,200 finance, risk and sustainability professionals, only 11% said their organization's data quality is sufficient for AI. One in four said internal audits had caught AI errors that reached external audiences or board members.
When an answer comes back inside a familiar, approved interface, people stop double-checking. The polish of the channel lends the number a confidence the data underneath has not earned.
Before an agent can safely act inside your ERP, someone has to:
Decide which system is the source of truth for each key entity: customer, supplier, product, cost center
Resolve duplicate and conflicting records across the ERP, CRM and warehouse
Agree on one definition for the metrics agents will be asked about most
Name an owner for each critical dataset
Map what breaks downstream when a field or table changes
None of these tasks are particularly difficult. The problem is that they are usually done once, for a migration or an audit, and quietly go stale long before the next agent arrives.
How to Make Your ERP Data Agent-Ready in 30 Days?
You do not need to redesign your data platform before employees start asking ChatGPT about your Oracle data.
Within the first month, focus on the handful of questions people will ask first: revenue, open payables, headcount, inventory, top customers.
Start by tracing each of those answers back to the tables and systems that feed it, and resolve duplicate and conflicting records for the entities involved. Then attach an owner and a definition to each one.
Next, expose this context through the same channels your agents and engineers use to find data, so the agent reads a governed definition, not whichever table it reached first.
DataManagement.AI accelerates this process by automatically discovering metadata, mapping lineage across the modern data stack, identifying downstream dependencies, and creating a centralized knowledge layer that stays continuously updated as your environment evolves. It resolves duplicate and conflicting records, and ownership and policy travel with every entity, so the data behind each approved agent action is known, owned and current.
Instead of spending months documenting assets manually, your teams can see what every agent-facing answer depends on and who stands behind it.
Agent Risk Is Actually a Data Context Problem
Oracle and OpenAI are taking on the hard work of connecting agents to enterprise systems. Identity, permissions and audit come with it.
What does not come with it is agreement on what your data means, where it came from and who owns it.
Organizations that scale agents successfully do not stop at controlling access. They make sure the data behind every approved action is defined, traced and owned.
Warm regards,
Shen and Team