Oracle and OpenAI announced an expanded partnership on 27 October 2026 at Oracle AI World in Las Vegas, bringing ChatGPT, OpenAI models and Codex into Oracle business software. Employees will be able to ask plain-language questions about company data and complete authorised workflows inside Oracle applications.

The brand names are not the point. The pattern is an AI agent over a company's own systems, with permissions and approval chains intact — and it works without an Oracle stack, through custom agents over your own CRM, ERP and data.

Oracle + OpenAI: ChatGPT Inside Your Business Software

What Oracle and OpenAI actually announced

The partnership combines OpenAI models, ChatGPT and Codex with Oracle applications, data and cloud infrastructure, so staff can find company information and complete authorised workflows by asking in plain language. Oracle says the offerings follow existing permissions, approval chains and access controls, with human review where required — the detail that matters most.

Two cautions first. Oracle says development schedules and pricing may change, so treat commercial details as provisional; this post invents no pricing. The buying route is US-specific: in the US, Oracle customers can buy API, Codex and ChatGPT Work access through Oracle Marketplace using eligible Oracle Universal Credits. Do not assume the same route applies in the United Kingdom, the UAE, Saudi Arabia or Pakistan without checking.

The four offerings, in plain English

1. Oracle AI Workspace: ChatGPT plugins over Oracle systems

Workspace is a set of ChatGPT plugins connecting Fusion Applications, NetSuite, Oracle AI Database and Oracle Cloud Infrastructure. An employee could ask for an order status or invoice figure, then complete an authorised workflow in the same conversation — but only within what they are already allowed to see and do. Workspace as a whole is still in development, although individual plugins are listed.

2. Oracle App Composer: build apps with coding tools including Codex

App Composer lets teams use coding tools, including Codex, to build applications connected to approved company data, deployed on Oracle cloud infrastructure — faster internal tools, built against approved data, not a shadow dataset. As in any serious custom software development project, data source and access rules must be agreed first.

3. Oracle AI Security Mesh: security testing and probing tools

Security Mesh covers security testing and probing tools. That matters because an agent on live data is a new attack surface: prompt manipulation, over-broad access and leaked context do not exist with a standalone chatbot. Ask what was probed, what it found, and how often it is repeated.

4. Oracle AI SMB Bundle: a packaged offer for smaller businesses

The SMB Bundle is a packaged offer for smaller businesses on Oracle Cloud Infrastructure, which can turn a project that never starts into a workable pilot. The same logic applies outside Oracle: a scoped agent over one workflow beats an enterprise-wide rollout as a first step. When weighing build versus buy, compare SaaS development options against a custom build.

Why this matters even if you do not run Oracle

Most readers do not run Fusion or NetSuite; many run spreadsheets, a CRM, an accounting package and WhatsApp. The announcement still matters: valuable AI is no longer a separate chat window, but an agent on top of the systems where work already happens. The question changes from "which chatbot?" to "which workflows may an agent touch, and under what controls?" See our guide on chatbots versus AI agents and retrieval-augmented chatbot guide for grounded answers over your own documents.

The pattern to copy: an agent over your own systems, with permissions intact

The pattern has four parts.

• One conversational front door. Staff ask in plain language via a chatbot, a WhatsApp bot, or a voice system such as the AI calling agent.
• Live connections, not exports. The agent reads the CRM, ERP or order system through controlled connections, so answers reflect current data. That is the core of serious AI automation.
• Permissions at the data layer. The agent inherits the user's role. A warehouse operative and a finance manager asking the same question should get different answers. If a vendor cannot show that, stop the pilot.
• Actions gated by approval. Reading is low-risk; acting is not. Orders, credit notes and customer messages should pass the same approval chain a human would — the principle behind well-built CRM development and restaurant ERP workflows.

See AI automation works for logistics companies, and guide to GPT-6.1 Terra for business automation for cheaper agent models.

Governance questions to ask before letting an agent act

An agent that only reads is a search tool; one that acts is a colleague with system access. Ask these before a pilot touches live data:

Who can it impersonate, and what can it change?

Demand a written list of permitted actions per role and every system it can write to. "Anything the user can do" is delegation without limits. Start read-only, add one write action at a time, and log every action.

Where does the data go, and what is retained?

Confirm which models process your data, whether prompts are used for training, and how long logs are kept. Businesses building AI agents should settle this in writing before the first live query.

What happens when it is wrong?

Every agent will eventually misread a figure. Does a human catch it before it reaches a customer or ledger, can it be reversed, and is there an audit trail? Keep human review on payments and price changes.

How is it tested, and how often?

Ask for the equivalent of Oracle's Security Mesh for your build: adversarial and prompt-injection tests, repeated after every significant change.

How a custom agent delivers the same benefit without an Oracle stack

A custom agent need not replicate Oracle's whole platform. Most businesses need one or two workflows done well: pick the job eating staff hours, connect the agent to its existing systems, enforce existing roles, and put human approval where money or customers are involved.

That suits mixed systems: a retailer might connect an agent to store and stock through ecommerce development; a service firm to CRM and scheduling; a mobile team can take it to the floor through Flutter app development. A scoped pilot can be measured, then extended or stopped on evidence. As a rough guide only — an estimate, not a quoted price — a single-workflow pilot is typically measured in weeks.

A practical pilot plan

• Week 1 — pick one workflow and one role. Choose a high-volume, low-risk job. Record current time and error rate, and list the permissions needed.
• Week 2 — run read-only, with a human reviewer. The agent answers and drafts but takes no action. Review every output and check role-based permissions are respected.
• Week 3 — add one gated action. Allow one reversible action behind the approval chain — drafting, not sending, a customer reply. Measure time saved.
• Week 4 — decide on evidence. Scale it, tighten it, or stop it. Do not expand until the first passes.

Coverage appeared in Unite.AI and OkayNews. Remember: Workspace as a whole is still in development, and Oracle says development schedules and pricing may change.

Frequently asked questions

What did Oracle and OpenAI announce on 27 October 2026?

At Oracle AI World in Las Vegas, an expanded partnership combining OpenAI models, ChatGPT and Codex with Oracle applications, data and cloud infrastructure, through Workspace, App Composer, Security Mesh and the SMB Bundle.

Is Oracle AI Workspace available now?

Not as a finished whole. Workspace as a whole is still in development, although individual plugins are listed.

Can my business buy ChatGPT access through Oracle?

In the US, Oracle customers can buy API, Codex and ChatGPT Work access through Oracle Marketplace using eligible Oracle Universal Credits. That route is US-specific.

Do I need to run Oracle to get this kind of AI agent?

No. The pattern — an agent over your own CRM, ERP and data, with permissions intact — can be built as a custom agent over your existing systems, starting with a single-workflow, read-only pilot.

What should I check before letting an AI agent act on live data?

Check permitted actions per role, where data goes and retention, what happens when it is wrong and whether actions can be reversed, and how it is security-tested. Keep human approval on payments and anything irreversible.