Security and data.
Where your data goes, who can change what, and how your auditor checks the record without trusting us.
| What | Where it goes | Notes |
|---|---|---|
| Your books | They stay in your ledger. Koncernify reads Fortnox directly, or an SIE file from any ledger. | Koncernify never deletes, voids or reverses an entry in your ledger. |
| Files for the group check | Kept with the month in the same EU database EU | The SIE file and the reporting you upload. The group check shows payroll vouchers by number and amount only. The file is kept as uploaded, and only an owner can download it. |
| What Koncernify keeps | Cloudflare, in a database under its EU jurisdiction EU | Checks, drafts, decisions, the record and the ledger connection, which is encrypted. |
| Questions to the box | Off, until an owner turns it on. Then our Claude on Amazon Bedrock inside the EU EU, or Claude in your own AWS or Azure | What you type, the chart of accounts and the month’s figures. Not used for training and not shared with Anthropic. |
| Your own assistant | Copilot, Claude or ChatGPT, on your own account with its provider | It reads what your role may read. It sends drafts for review. Koncernify sees only what it sends. |
| Sign-in | Google or Microsoft, with your work account | We do not store passwords. Two-step sign-in is set in your Google or Microsoft account. Koncernify cannot see whether it was used. |
| Payments | Stripe, in Ireland | We never see card details. |
Every provider is listed on the subprocessors page. The terms are in the data processing agreement.
AI drafts. People decide.
No journal an AI drafts reaches the ledger until a person with the right authority approves it.
An owner can set up standing approvals for routine journals. Those post first and are signed off at month end, and the month cannot lock without the sign-off. The rules that refuse a journal are fixed checks. No model judges another model’s work.
Where AI is used.
The checks are code. The model only works on what a person asks it for.
- No modelEvery check, every drafted answer and every rule that stops a journal. The accruals too. The same files give the same answer every time.
- Your choiceAn owner sets where the AI runs for the workspace: off, our model, or your own endpoint. A new workspace starts with it off. The change is kept in the record.
- Our modelClaude Sonnet 5.5 from Anthropic, through Amazon Bedrock in AWS regions inside the EU. It drafts journals, writes checklist steps and answers questions in the box.
- Your ownClaude on Bedrock in your AWS account, on Microsoft Foundry in your Azure tenant, or your own Anthropic key. The books go where your own agreement says, and nowhere else.
- What it getsYour question and a summary of the month: the checks’ findings, entry lines with their text and amounts, and any voucher you name. Never the full ledger file.
- TrainingYour data trains no model, ours or anyone else’s. Bedrock does not share it with Anthropic.
- Marked as AIA journal the model drafted is labelled as AI, next to the person who approved it. It always waits for a person.
If our EU route is down, the model is switched off. It never falls back to a provider outside the EU. With the AI off, every check and drafted answer works the same.
Who can do what.
- ReviewersApprove up to the amount, accounts and ledgers an owner sets. Never their own drafts.
- Assistants and agentsDraft within their own ledgers, accounts and amounts. They cannot approve or post.
- AuditorsRead only. A link opens one report and expires on the date you set.
- OwnersLoosen a rule, reopen a locked month or remove a person. Each change is signed.
A record nobody can quietly change.
Each decision is signed with the reviewer’s name, the time and the exact lines. An outside authority timestamps the signatures every hour.
Once a decision is timestamped, not even we can change it without it showing. Decisions made through the agent hook are also kept on your own machine. There each entry carries the hash of the one before, and the service countersigns the latest one. Journals from the box, the connector and the form are not chained. The ledger is read back every night, so a journal changed in Fortnox after approval shows in the record. Changes to the rules, roles and limits are signed too. How an auditor checks a record.
Keys and limits.
- Ledger connectionsEncrypted with a key for your workspace alone, destroyed when the workspace is deleted.
- Auditor linksWe store only a hash of the token, so a copy of our database opens nothing.
- Large approvalsAn owner can require a fresh sign-in above an amount they choose.
- RequestsCapped per key, address and workspace. Repeated wrong keys shut an address out.
- LeavingAn owner can download everything at any time. We close a workspace 90 days after the subscription ends. The record of journals approved for your ledger is then kept read-only until the bookkeeping retention period is over, and handed over on request; everything else is deleted.
For your IT and security review.
The questions a security review sends a new vendor, answered before you ask.
We do not yet have ISO 27001, SOC 2 or an ISAE 3402 report. The answers say what we do instead. The record can be checked without trusting us.
Report a problem.
If you find a security problem, or a way to change a record without it showing, write to security@koncernify.com.
We reply within two working days. Give us reasonable time to fix it before you publish, and do not access other customers’ data. We will not take legal action against research done in good faith.