Where Your Oracle EPM Numbers Go When You Ask an AI
When you ask an AI assistant about your plan, the numbers leave Oracle. They make four stops before you read the answer: your Oracle pod, the connector, the app you type in, and the model.
- Most security reviews I see look at one of the four stops, the model. The other three matter as much.
- Each stop has an operator, and each one can keep a record.
- Three questions cover all four: where does it run, who operates it, and what does it keep, for how long.
The four stops at a glance
Below is each stop, what it can keep, and the question to ask at each one. The numbers travel left to right.
- Stop 1Your Oracle podRun by OracleKeeps a sign-in log for 60 days
- Stop 2The connectorRun by a vendor or by your teamKeeps logs, cache and the sign-in
- Stop 3Your chat appRun by the AI companyKeeps your chat history
- Stop 4The modelRun by the AI company, your cloud, or your own serversKeeps prompts and answers for a time
The answer comes back through your chat app, so that stop sees the numbers twice.
Stop 1Your Oracle pod
The numbers start here, in Oracle's cloud. Oracle checks the roles of the account that signed in, and it returns only what that account may read.
Two setups are common. In the first, the assistant signs in as each person, so each person keeps their own access. In the second, it signs in as one stored account, so everyone who uses it gets that account's access. The second is easier to set up. It's also how a planner who should only see Canada ends up seeing every entity.
Oracle keeps an access log of who signed in to the environment. It retains that log for the last 60 days only (Oracle).
Does the assistant sign in as each person, or as one shared account? Which roles will use it, and what can each one read?
Stop 2The connector
The connector is a small service between the assistant and Oracle. It turns the assistant's request into Oracle REST calls, and it carries the numbers back. You'll often hear it called an MCP server.
It can run on someone's laptop, at a vendor, or in your own cloud account. Wherever it runs, it sees every request and every number that passes through. It also holds the sign-in to your pod.
Request logs, cached results and the Oracle sign-in. Whoever operates it decides how much of that it keeps, and for how long.
Where does it run, and who operates it? What does it keep, and for how long? Does its log hold member names only, or the values too? How is the Oracle sign-in stored?
Stop 3The app you type in
This is Claude, ChatGPT or Copilot. You type the question here and read the answer here. So the numbers pass through it twice: once when the connector hands them over, and again in the answer you read.
Your chat history, in the AI company's cloud. On business plans, your company's admins usually set how long it stays. Check what your plan says.
Where is the chat history stored? Who in your company can see it? How long is it kept, and can you delete it?
Stop 4The model
The model reads the numbers to write the answer. There is no way around that: an assistant can't explain a variance it never saw.
The model can run in the AI company's own cloud. It can also run at your cloud provider and be called through your own cloud account. Claude through Amazon Bedrock or Google Cloud, and OpenAI's models through Microsoft Azure, are two examples. A third option is an open model such as Llama on your own servers. The model still reads the numbers, but it runs in your building, your team operates it, and your team decides what it keeps.
Your prompts and its answers, for a period your plan and provider set.
Is my data used to train models? How long are prompts and answers kept? In which country is the model run?
What each stop keeps
Every stop can keep a record, and a different party runs each one.
| Stop | Who runs it | What it can keep | The question to ask |
|---|---|---|---|
| Your Oracle pod | Oracle | An access log of sign-ins, for 60 days | Does the assistant sign in as each person? |
| The connector | A vendor, or your own team | Request logs, cached results, the Oracle sign-in | Does its log hold names only, or values too? |
| The app you type in | The AI company | Chat history | Where is it stored, and who can see it? |
| The model | The AI company, your cloud provider, or your own team | Prompts and answers, for a set period | Is my data used for training? |
Three questions for your next AI and EPM meeting
Three questions cover all four stops:
- Where does each stop run?
- Who operates it?
- What does it keep, and for how long?
If your team can answer all three for every stop, you're ready to talk about use cases. If not, start there. It's a one-hour conversation, and it's the one your security team will ask for anyway.
You don't have to do it alone. Oracle now runs a program for teams that connect ChatGPT, Codex or Claude to Planning, and it includes a security review of authentication, authorization and audit (Oracle). I covered it in my 26.09 update review.
Go further:
- Security Field Guide, part 1: the four risks of an AI agent on Oracle EPM.
- Security Field Guide, part 2: the ten-question readiness check.
- Questions about your own setup? Book a 30-minute call.
Questions people ask
Do my Oracle EPM numbers go to the AI company?
The numbers the assistant reads go to the model, so it can write the answer. Where that model runs, and what it keeps, depends on your plan and your provider.
Can the connector run in my own cloud?
Yes. A connector can run on a laptop, at a vendor, or in your own cloud account. When it runs in your account, your team decides what it keeps.
How long does Oracle keep access logs for EPM Cloud?
For the last 60 days. Oracle suggests downloading copies if you need a longer history.
What should I check before I connect an AI assistant to Oracle EPM?
For each stop, ask where it runs, who operates it and what it keeps. Then check whether the assistant signs in as each person or as one shared account.
Sources
- Oracle, Using Access Logs to Monitor Usage. Checked October 12, 2026.
- Oracle, Enrolling in the Implementation Success Program for Conversational Interface. Checked October 12, 2026.
I'm an Oracle ACE Pro. That's a community recognition program. I don't work for Oracle, and no vendor paid for this page.
Not sure what each stop keeps in your setup?
Book a free 30 minute call. We go through the four stops for your environment, and you leave knowing which questions are still open and who to ask. No pitch.
Fred Mamadjanov
EPM Solution Architect · Oracle ACE Pro · Vena Certified