Data handling
Data handling and quality control
How we handle the documents you give us, and what we check before going into production.
Documents
How we handle your documents
This is how we handle data in contract development. For our own products (Pentra, Mensetsu Kanpe AI and others), see each service's terms and privacy policy.
| Before we receive anything | We sign a non-disclosure agreement (NDA) before receiving documents. Your own template is fine. |
|---|---|
| Where data goes (AI) | By default we use the OpenAI and Anthropic APIs (both US providers). If you need processing in Japan, we build on your own Azure OpenAI or Amazon Bedrock. A region name alone does not guarantee in-country processing, so before the proof of concept we agree in writing on the model, the processing method (including settings that stop processing from being routed to other regions) and the storage location. |
| Use for training | Content sent through the API is not used for training under each provider's default settings (OpenAI, Anthropic). We do not turn on any setting that opts in to training, and we do not send feedback that contains your information. "Not used for training" and "not stored" are different things: providers may keep data for a set period per feature, for example to monitor abuse. |
| AI in our development work | We also use AI (Claude, OpenAI Codex and others) in our own development work. We discuss with you beforehand whether your actual documents are given to a development AI, and if you prefer, we build using only fictional data or redacted documents. |
| Storage | Documents are kept only in a place set up for your project and are never mixed with other projects or our own services. |
| Deletion | Documents in our storage and in email attachments are deleted within 30 days after the contract ends, and we confirm the deletion in writing. Copies in automatic backups disappear within 30 days of deletion at the latest. Data held by AI providers is deleted according to their own policies. |
| Subcontracting | We do not subcontract work to other people. We use only the AI providers named above and any cloud provider we agree on with you. |
| If there is an incident | If we find a leak or a risk of one, we send a first report within 24 hours, then report the cause and our response after investigating. |
| Personal data | If documents contain personal data, we agree with you before the proof of concept on the purpose and scope of use, and on how data is handled when it is entrusted to us (Estonia) and passed on to AI providers (US). If there is a way to avoid personal data, we recommend it. |
Quality control
What we check before going into production
Machine checks, a second AI's review and human judgment, layered together.
Set the pass criteria first
We put the criteria for accuracy, speed and cost in writing before we measure. We do not loosen them after seeing results.
Automated tests
Automated tests run on every change. Pentra has more than 1,400 tests and our internal systems have more than 2,800.
Independent review by a second AI
An AI from a different company (OpenAI Codex) reviews each change independently. When it is unavailable, a separately configured AI follows the same procedure. We fix every serious finding before release. An AI review is no substitute for a person, so the CEO makes the final check.
Run it in a staging environment
We confirm behavior in a staging environment built like production before going live.
Make rollback instant
Every production release records where to roll back to. If something goes wrong, we return to the previous version at once.
A person approves risky actions
High-impact changes such as payments, authentication, sending volume and database structure are never left to AI. A person approves each one before it goes ahead.