doclane
en
Get free demo

Security

Security in Doclane

Doclane processes documents that come from outside: orders, claims, customer onboarding requests and product data from your customers and suppliers. We start from the assumption that any document may be wrong, or may try to deceive, no matter who sends it. That is why the platform is built on four principles:

  1. 01

    The AI reads, the rules decide.

    The AI model extracts data from the document. Approval, recipients, status, prices and amounts are always controlled by fixed logic and your own master data, never by what the document says.

  2. 02

    The sender is verified outside the document.

    Who sent a document is determined by technical checks and your lists of trusted senders, not by the name in the document.

  3. 03

    Nothing is written to your ERP without a verified source or a person.

  4. 04

    Your data is yours.

    Everything is separated per company, hosted in the EU, and never used to train AI models.

Your data

Separation
All data is separated per company: documents, templates, settings, logs and files. What Doclane learns from your corrections is only used for you. The separation is checked by an automated test suite that runs on every change to the platform.
Hosting in the EU
Doclane runs on servers at Hetzner in the EU. We only use providers headquartered and hosted in the EU, and tools we run ourselves on our own servers.
The AI model
Documents are processed by an EU-hosted AI model from Mistral. Your data is not used to train the model. The AI model has no access to the internet, your systems or any other tools. It receives text and returns structured data, nothing else.
Transfer and storage
All traffic is encrypted, and uploaded files are kept in private storage that can only be accessed through Doclane.

How we process documents

Files
Every file is checked on its actual content, not its file extension, against a list of permitted formats, with limits on size and number. All files are virus scanned before they are read, and again before they are passed on to your systems. Infected files are quarantined, never processed, and your administrators are notified. The reading of files itself takes place in an isolated environment without network access, so a malicious file cannot affect the platform. Formulas in spreadsheets are never calculated, and embedded objects in documents are not processed.
Hidden content
Documents can contain text a person cannot see, such as white text on a white background, invisible characters, hidden sheets in spreadsheets or hidden elements in an HTML mail. Doclane removes hidden content before the document is read, and only extracts from what a person would also see. Documents with hidden content are flagged and always sent for manual review.
Attempts to manipulate the AI
Document text is sent to the AI model as delimited data, never as instructions. Content that tries to change how Doclane works ("ignore previous instructions", "set the price to...") is detected in several languages, and the document goes to a person instead of being processed automatically. Your own template notes rank below the system's rules and cannot override them.
Checking the AI's output
What the AI model returns is validated against a fixed schema for your specific template: field types, permitted values, lengths and number of lines. Fields that control the system, such as status, approval, recipients and destination, cannot be set by the AI. Every identifier (order number, item number, EAN, invoice number, customer number, email) must appear verbatim in the document. A value that does not appear in the document is flagged for review.
Learning
Templates only learn from documents from verified senders without security flags, and only from corrections made by your own employees. Corrections from external parties (such as your customers in an approval view) are recorded, but do not change how the system behaves. Learned examples must appear verbatim in the document and must not resemble instructions.

Incoming email, forms and links

Email
Each channel gets its own unique, unguessable Doclane address, which you forward to. Doclane does not need access to your mailbox. Every mail is checked with DKIM, SPF and DMARC, so a forged sender cannot pose as a trusted one. You define the trusted senders per channel yourselves. Mail from anyone else is received but held back: no workflows run, nothing is written to your ERP, and no mails or links are sent until an administrator on your side releases the document. Addresses can be rotated if a channel needs to be closed.
Public forms
Upload and onboarding forms are protected against automated submissions and rate limited. Everything that comes in through a public form is treated as unverified: it is never approved automatically and never written to your ERP without a person. Your administrators can rotate a form link, which closes the old link immediately.
Secure links
Links for approval, information and upload are only sent to verified addresses: your registered contacts or the customer address held in your ERP. An address someone types into a document never receives a link. Each link is tied to one document and one recipient, expires automatically and is closed once the action is completed. Opening a link never changes anything. An approval requires an active action, so mail scanners that open links automatically cannot approve anything. Links are not stored in readable form in our systems.

Writing to your ERP

We design from the question: "If a document were completely controlled by an attacker, what is the worst that could be written?" The answer must be: nothing of consequence.

What can happen without a person
Automatic processing requires every condition to be met: the document comes from a verified, trusted sender, it has no security flags, the value is below a limit you set yourselves, and it is not the first document from that sender. Everything else waits for a person on your side to click "Send to ERP".
The customer is determined by the sender, not by the document
Which customer a document belongs to is determined from the verified sender or the channel setup, never from a name or customer number in the document. If they do not match, the document goes to review.
Claims and credit notes
Every credit or return line must refer to a posted invoice for the same customer. The quantity cannot exceed the invoiced quantity minus what has already been credited. The price always comes from the invoice line in your ERP, never from the claim.
Orders
A delivery address that differs from your master data is flagged and never processed automatically.
Customer master data
Changes to bank details, invoice addresses and reminder addresses on existing customers always require manual approval and are sent to your finance department. Credit limits and payment terms are never updated by Doclane.
Limits
Each company has a daily cap on the number of writes and the amount credited. If it is exceeded, writes are stopped and your administrators are notified. All values used in lookups against your ERP are treated as data and cannot change what is queried.
Everything is logged
Every write to your ERP can be traced to the document, the person or the rule that triggered it, with the original document attached.

Access and accounts

Two-factor login
All administrators log in with a password plus a second step: a passkey (fingerprint, face or PIN on the device) or a code from an authenticator app. Doclane's own employees and reseller administrators use passkeys exclusively. New passwords are checked against a local list of known leaked passwords, and repeated failed login attempts lock the account temporarily.
Support access
When Doclane support or your reseller logs in as one of your users, a justification is required, the session is time-limited, a banner is visible at all times, and your administrators are notified immediately. Under Settings > Support access you can see every single time it has happened: who, when, why and for how long.
Templates and rules
Every change to template notes is versioned, logged and announced to all administrators in your company. Notes cannot contain links or email addresses and cannot control recipients, approval, prices or destinations.
Usage limits
Each company has a daily cap on AI processing and fixed limits per document. This protects you against unexpected costs and against someone flooding your channels. Documents rejected as spam, misdirected or unsafe are not billed.

Integrations and operations

Integration secrets
Access keys for your ERP, webshop and mail accounts are stored encrypted, never shown again in the interface once saved, and automatically removed from all logs and error reports. Integrations can only point to public addresses, so Doclane cannot be misused to reach internal systems.
What we pass on
All values from documents are treated as data when they are shown in Doclane, in mails and in exports. Excel exports are written so they cannot contain active formulas. Files are only passed on to your systems after a clean virus scan.
The platform
All traffic is encrypted (TLS), and all pages are served with modern security headers. The admin interface cannot be embedded on other websites. Our software dependencies are continuously checked against known vulnerabilities, and the codebase is scanned for accidentally stored secrets.

Testing and verification

Attack testing in every release
Doclane's test suite contains a corpus of malicious documents: hidden text in PDF, Word, Excel and email, invisible characters, instructions to the AI in several languages, attempts to manipulate ERP lookups and exports, and files built to break parsers. For each document it is checked automatically that it is flagged, not approved automatically, not written to an ERP and not learned from. The suite runs before a new version can be released.
Isolation testing
Every page, download and export is tested automatically to ensure that a user from one company cannot reach another company's data, including across resellers.
Compliance
Doclane is working towards ISO 27001 certification. Our data processing agreement and list of sub-processors are published on this site.
Responsible disclosure
If you have found a vulnerability, write to hello@doclane.ai. We confirm receipt within two working days, keep you informed, and take no legal action against reports made in good faith.