How we run it

Your deployment is yours

• One deployment per customer: your own application, your own database, your own configuration. There is no shared tenant table anywhere in the platform.
• Sign-in is your Microsoft directory and only your directory. An allow-list of tenants decides who may sign in, and registering a tenant is a deliberate act — nothing is added by a link being clicked.
• Secrets never live in the code. Every deployment's configuration is managed in a secrets manager, injected at run time, and rotated without a rebuild of anything but the container.

Approvals are the brake, and the brake has a record

• Sensitive actions wait for a person: resetting a password or MFA, suspending an account, changing someone's pay, approving time off, letting an agent past a guardrail. The person who decided, when, and under which condition is written down every time.
• Nobody approves their own request. A requester cannot decide their own time off, a proposer cannot approve their own change, and the deploy gates refuse the same person on both sides.
• When two systems could both decide the same thing, one of them is refused. A decision routed to the approvals engine cannot also be taken in the app; there is exactly one place the answer comes from.

AI on your terms

• Bring your own Anthropic key and the AI line item never touches our invoice. Or run human-mode and the AI is off entirely — every step a person would have done is done by a person.
• AI production is opt-in per step, never assumed. The pipelines receive, version and gate what humans make; where the AI does produce, a human signs the step before it counts.
• Agents that hit a guardrail stop and ask. Sensitive or irreversible actions become an approval request, and the agent waits for the answer.
• The AI is not for high-risk uses. It must not be used for, or changed into a system for, anything the EU AI Act lists as high-risk in its Annex III, such as recruiting or screening people, deciding anyone's promotion, pay or dismissal, handing out work by a person's behaviour or traits, monitoring or evaluating how people work, judging creditworthiness, or pricing life or health insurance.

Backups we have restored from

• Hourly backup schedules sweep what is due, with retention rules that prune old copies and record what they pruned. A partial failure is reported as partial — the platform never logs a clean run it did not have.
• We have run the restore drill against a real backup, on the production shape, and fixed what it found. A backup nobody has restored from is a hope, not a backup.

We watch it, and it files its own bugs

• Uptime checks and alerting run on every deployment. Errors in production are collected, deduplicated by fault rather than by event, and filed as tickets in the platform's own service desk, where they have an owner and a lifecycle.
• A daily digest reads the production logs and opens a ticket for any error signature it has not seen before. The platform is its own first customer.

Who else handles data, and why

Every third party the platform can send data to, what reaches them, why, and where they process it. Some apply only when a deployment turns the feature on; each says so.

• AnthropicUnited States

  The content the AI steps work on. In the apps — ticket text, feedback, drafts, questions — it is minimised before it is sent: email addresses and phone numbers are replaced, the people in your directory are named by their role and department instead, and a sign-in identity never goes. Where your platform admin names SharePoint folders for the assistant, the passages it answers from include your own documents, minimised the same way. Where your assistant admin switches on live records, an answer may also look up the asker's own records, run as them: their tickets, tasks, learning, time off, pay, customers, channel messages and figures, minimised the same way. It works from your directory, so a bare first name, a surname that is also an everyday word, or someone the directory does not list can still come through, and what the AI's tools look up, such as web search results, goes as found. The software-delivery pipelines are the exception: they send the documents and code they are working on as they are, because stripping would corrupt them.

  The AI steps: triage, classification, drafting, the assistant's answers. Off entirely in human-mode. With your own key, under your own agreement with them.

• Microsoft (Entra ID + Graph)Your own Microsoft tenant's region

  Your sign-in identity and group membership. Where you connect them: shared-mailbox contents, Teams messages, SharePoint list items, and the files in the SharePoint folders your platform admin names for the assistant, which are read and never written.

  Sign-in, department sync, email intake, Teams triage, the accounts list, the assistant's answers from your own documents.

• DigitalOceanUnited States (New York)

  Everything at rest and in transit: the databases, application traffic, attachments, HR documents and backups.

  Hosting, managed databases, object storage.

• SentryUnited States

  Error events, stack traces and a sample of performance traces, with personal data suppressed by design before they leave.

  Error monitoring.

• ExaUnited States

  Search queries, which may carry a user's own words from a question, and the software-delivery pipelines' research queries.

  Web search behind service-desk answers, market intelligence and the software-delivery pipelines' research. Only where web search is turned on.

• StripeUnited States

  Nothing outbound beyond credentials. The platform reads charges, prices and customer records from your Stripe account into the books.

  Revenue ingest and AI-cost accounting. Only where the accounting spoke is connected to Stripe.

• Better StackNot pinned to a region

  Application logs.

  Log storage and search.

• GitHubUnited States

  Source code, build logs and the output of platform-run automation.

  Development, builds and publishing.

• DopplerNot pinned to a region

  Secrets only. No personal data.

  Secret management for every environment.

• JumpCloudUnited States

  Names, email addresses, usernames, departments and device identifiers — and account actions: suspend, reactivate, unlock, reset a password or MFA, lock or restart a device.

  Identity and device management. Only where your deployment connects JumpCloud.

• Backblaze B2The bucket's own region

  Encrypted offsite copies of every database and of the file store, encrypted before they leave with a key the bucket never holds.

  Backups that survive the loss of the hosting account. Only where your deployment configures offsite copies.

• Azure SentinelYour own Azure tenant's region

  Security events from the SDLC engine, shipped to a Log Analytics table in your tenant.

  Security event forwarding to your own SIEM. Only where a deployment configures it.

What we do not claim yet

• No SOC2 or similar certification yet.
• No status page yet. Uptime monitoring exists; the public view of it does not.
• A DPA and an SLA come with the Enterprise Fork tier. They are agreements we sign with you, not templates we publish.
• That every processor named above is covered by a data-processing agreement we can hand you on request. Our register records which instrument covers each and which are still to be confirmed; we will say “confirmed” only when we can produce the document.
• No third-party penetration test report yet. The platform reviews itself hard and often — that is not the same thing, and we will not pretend it is.