Penta and Magnesia
AI and your data
Provided by Sepenta Eğitim Danışmanlık Yazılım Turizm Ticaret Limited Şirketi. Last updated: 26 October 2026.
Draft. Highlighted items still need a value or a check before this page is published.
The short version. AI is used only when someone asks for it, sees only what that job needs, travels encrypted,
is never used to train models, is kept by our AI providers for a short, stated time, and never decides anything on its own: a person reviews
and approves the result. This page says exactly what is sent, to whom, for how long, and what we do not do yet.
What we promise, and how it is true today
| Promise | How it works |
|---|---|
| AI only when someone asks | Nothing is analysed in the background. In Penta, AI runs when a person runs an assessment, a transcription or a story breakdown. In Magnesia, a standup transcript is analysed only when an Admin or Delivery Manager asks for it, and the optional report narrative only when a Delivery Manager drafts it. |
| Only what the job needs | Each request is built from one workspace's data for one task (see the table below). Sign-in details, Jira and GitHub access tokens, the audit log, other people's workspaces and anything not needed for the task are never part of it. The Magnesia report narrative is written from program numbers with no names or email addresses. |
| Never used to train models | Sepenta does not train models on customer content. Our providers' API terms say the same: OpenAI states that data sent to its API is not used to train or improve its models unless a customer opts in, and Anthropic states that by default it does not use inputs or outputs from its commercial products, including its API, for training. We do not opt in. |
| Encrypted on the way | Every call from our servers to an AI provider uses HTTPS (TLS). Provider keys are held in Google Secret Manager, never in code, and are never written to logs. |
| A person decides | AI output is a draft. In Penta, every epic and story is reviewed and approved by a person before anything is created in Jira, and what is created is created as that person. In Magnesia, alerts are prompts for the team to discuss, a transcript can be corrected before it is analysed, and a report narrative is edited before it is published. We make no automated decisions about people with legal or similarly significant effects. |
| Kept briefly by providers | See the retention table below, with the real numbers and the one setting we are still changing. |
| Kept apart between customers | Workspaces are isolated in the database and in file storage, and automated tests check that a request for one workspace cannot read another's data. That includes the data used to build an AI request. |
Exactly what is sent
| Product and task | Sent to | What is sent | When |
|---|---|---|---|
| Penta: transcribe recordings | OpenAI | The audio of a recording a person added | When an assessment is run |
| Penta: describe images | OpenAI | An image a person added | When an assessment is run |
| Penta: assess the business context | OpenAI | The workspace's notes, transcripts and the text extracted from its documents, for that assessment | When a person runs the assessment |
| Penta: draft stories | OpenAI | The approved epic's text | When a person asks for stories |
| Magnesia: transcribe a standup | Google Cloud Speech-to-Text, EU region, in Sepenta's own Google Cloud project | The recording's audio | After a recording is made or uploaded |
| Magnesia: analyse a standup | Anthropic | The transcript (up to 120,000 characters), plus the work in flight from Jira as key, title, status, priority and assignee name, the open actions with their owners, and recent decisions | Only when an Admin or Delivery Manager asks |
| Magnesia: draft the report narrative | Anthropic | Program status, numbers, objectives, risks and decisions needed. No names or email addresses. | Only when a Delivery Manager asks |
Because a standup transcript contains what people said, and the analysis includes Jira assignee names, a customer who records standups should tell its people first. Magnesia shows a consent notice before every recording, and a pasted or uploaded transcript can be used instead of audio.
How long it is kept
| Where | What is kept | How long |
|---|---|---|
| OpenAI (Penta) | API requests, in abuse-monitoring logs | Up to 30 days, then deleted, unless the law requires longer. Not used for training. |
| OpenAI (Penta) | Stored responses (application state) | None. Penta asks OpenAI not to store responses, on every call, and an automated test fails if any call is added without it. |
| Anthropic (Magnesia) | API inputs and outputs | Deleted from Anthropic's systems within 30 days of receipt or generation, with the exceptions Anthropic lists (such as legal compliance and misuse). Not used for training by default. |
| Google Cloud Speech-to-Text (Magnesia) | Audio during transcription | Processed in Sepenta's own Google Cloud project in the EU. The temporary copy made for transcription is deleted when it is done. |
| Sepenta (both) | Your content and its AI results | Until you delete it or ask us to. Automatic deletion after a retention period is not switched on yet, and is planned. Recordings and transcripts in Magnesia are encrypted by the application as well as by the platform. |
How we guard against wrong or hostile AI output
- Text from a recording is treated as data, never as an instruction. Magnesia's analysis is told to ignore anything in a transcript that tells it to do something. Penta's instructions say the same for every document, note, transcript and image, and the model can only return a draft that a person must approve.
- Answers must fit a fixed shape. Penta requires the model's answer to match a strict schema and rejects anything that does not. Magnesia reads the answer as structured data and keeps an alert only if it concerns a rule we asked it to judge and cites a Jira item, pull request or number that really exists in the transcript or the data.
- The model cannot reach anything. It receives text and returns text. It has no access to your Jira, your GitHub or your files, and cannot change anything. Changes happen only when a person approves them.
- Quotes are avoided. Magnesia is told to paraphrase and not quote a transcript.
What you control
- Whether to record at all: a transcript can be pasted or uploaded instead.
- Who can run AI on recordings in Magnesia: only Admins and Delivery Managers.
- Whether anything is sent to Jira: in Penta nothing is, until a person approves it.
- Deletion: ask us at any time and we delete your content, and the AI results with it. See the privacy notice.
What we are doing next
- Apply for OpenAI's Zero Data Retention for Penta, which also removes the 30-day abuse-monitoring copy. It needs OpenAI's approval, so we cannot promise a date.
- EU processing at OpenAI for Penta, where OpenAI approves it. Today a provider may process data in the United States under standard contractual clauses.
- Request a zero data retention arrangement with Anthropic for Magnesia, which Anthropic grants by agreement.
- Automatic deletion of customer content after a set period, and a separate encryption key for each customer.
[CONFIRM: you agree to name these steps publicly, and that signed OpenAI and Anthropic data processing terms are in place]
What we do not claim
- We do not hold ISO 27001 or SOC 2 certification, and we do not claim that AI output is always right.
- Our providers may process data in the United States. We rely on their data processing terms and standard contractual clauses for that, listed on the sub-processors page.
- OpenAI still keeps API requests in abuse-monitoring logs for up to 30 days until Zero Data Retention is approved, as the table shows.
Questions from a security review or a customer questionnaire: write to info@sepenta.io.
Questions: info@sepenta.io. Sepenta Eğitim Danışmanlık Yazılım Turizm Ticaret Limited Şirketi, tax ID 7611208267, Kuşadası, Aydın, Türkiye.