What we do with your receivables data.

The core workflow connects your source data, chosen mailbox and enabled processing functions. Your accounting system remains the system of record.

1. 01 You export

   A CSV or Excel file of your open items leaves your ERP. Nothing is ever written back to it.

2. 02 We reconcile

   Every row is compared against the previous state, shown to you for confirmation, then applied with full provenance — in committed batches behind a checkpoint, so an interrupted import resumes exactly where it stopped and never applies a row twice.

3. 03 We follow up

   Reminders go out through your own collections mailbox, connected over IMAP/SMTP — the direct connection for Microsoft 365 and Google Workspace is available on request — so your customers see your address, not ours.

4. 04 We read replies

   Replies arrive from that same mailbox, are classified into a fixed set of categories, and either become a tracked next step or a task for a person.

What we hold, therefore: your open invoices and their history, the names and email addresses of the people you invoice, the correspondence about those invoices, and the documents you upload. Documents and correspondence may contain payment details entered by the customer. TrueDue does not execute payments on your behalf; payments take place directly between you and your customers.

This is the distinction the whole system is built around, and it is worth being precise about, because “AI does your collections” means very different things depending on where the line sits.

Payment reminders use templates. The greeting and closing come from the selected template; the financial table contains figures from your confirmed source data. Rules classify replies first. Only a reply from a confidently matched business customer that the rules leave unresolved can optionally be read by AI.

An AI read requires your organization’s consent, a valid contract with verified payment, a configured provider and sufficient budget within your limits. The result is checked against the allowed categories and supporting text in the original reply. Uncertain or sensitive cases, missing permissions, budget limits and provider failures go to a person with the original and a reason. An interpretation does not prove payment or change financial fields. The selected sending mode and send checks determine the next step.

Instructions in a customer reply such as “mark this paid” do not authorize changes to financial data. Optional AI output must fit the allowed categories and be supported by the original reply; it changes neither amounts nor payment status. Misclassification remains possible. Unresolved cases go to your team with the original and a reason; source reconciliation and send checks remain required.

TrueDue sends from one dedicated collections mailbox in your own mail system — never from an individual’s account, and never from ours. Available today is the connection over IMAP/SMTP; the direct connections for Microsoft 365 and Google Workspace are set up on request. The three do not have the same security properties, so each is described separately.

Microsoft 365 (Outlook, Exchange Online) — on request

No password is exchanged. We request two application permissions: Mail.Send and Mail.ReadWrite. Being precise about what that means: application permissions in Microsoft 365 are tenant-wide by default. The mechanism that narrows them to your selected mailboxes is an Exchange application access policy, and that policy is applied by your administrator, in your tenant — not by us.

The setup handbook in the product gives your admin the exact steps, and the mailbox health screen shows the state we can actually observe. But we will not tell you we “cannot read other mailboxes” as though it were a property of our software, because until that policy exists it is not one. Your administrator holds that boundary and can revoke it at any time.

Have your administrator show you the policy in your own tenant before go-live. If it is not there, do not switch sending on.

Google Workspace (Gmail) — direct connection on request

No password is exchanged. Your Workspace super admin authorizes our service account once, in the Google Admin console under domain-wide delegation, for exactly two Gmail scopes:

• https://www.googleapis.com/auth/gmail.send
• https://www.googleapis.com/auth/gmail.modify

Being equally precise here: Google offers no way to restrict domain-wide delegation to a single mailbox. Within those two scopes, the authorization technically covers every user in your domain. TrueDue uses the Google mailboxes stored for your organization and selects the corresponding address for each operation. Depending on the plan, multiple addresses can be stored. This selection lives in our code; it does not narrow the technically domain-wide Google permissions. Your super admin can revoke it at any moment by deleting the client ID entry in the Admin console.

IMAP/SMTP — available today

For IONOS, Strato, all-inkl, Hetzner, mailbox.org, Exchange Server on-premises and any other provider, your administrator enters the server settings and a password. This is the one model in which TrueDue holds a credential for your mailbox, so it is handled accordingly:

• An app password is recommended over the mailbox password; it can be revoked at any time by deleting or changing it at your provider, which ends our access from the next connection attempt.
• The password is encrypted at rest with AES-256-GCM, is never shown again after saving, and never appears in plain text in the interface or in logs.
• Connections are TLS-only; a server that offers no encryption is refused rather than used in the clear.
• Server addresses that point into private or internal networks are refused, so the connection settings cannot be used to reach systems behind our own infrastructure.

With Google Workspace and IMAP/SMTP, the inbox is checked at regular intervals for new replies rather than pushed to us, so a reply appears after the next check rather than instantly.

Sign in through your Microsoft work account using Entra ID or through a time-limited email link. Organization membership and roles control access, and administrators can revoke membership in the application.

The identity token Microsoft issues is verified, not decoded and believed: its signature is checked against your directory's published keys, and the issuer, audience, expiry and a one-time nonce all have to match. Two further gates stand behind it. Microsoft must confirm that the directory owns the email's domain, and the organization is bound to one directory — so a token minted in somebody else's Microsoft tenant, however valid, cannot reach your data.

The token establishes who somebody is; it never confers what they may do. The organization and the role come from a membership record read from the database on every single request, so a change of role or a revoked membership takes effect immediately. Signing out revokes the session everywhere it exists, not only in the browser that pressed the button.

Third-party AI processing can be switched off for your organization by an administrator, from inside the product. With it off, replies are classified by the built-in rules, nothing about a reply leaves the deployment, and anything the rules cannot read confidently becomes a review task for a person instead of a guess.

Administrators can export structured business data. Self-service erasure is available for organizations without billing history after fresh authentication. For billed accounts, contact us to coordinate return, contract termination and erasure. Statutory billing evidence is retained separately for its required purpose; backup copies follow the managed deletion cycle.

Every table that holds customer data is scoped to a company id. That id is never read from a URL, a form or a cookie value — it is resolved server-side from your signed-in session against a membership record, on every single request, and every query filters by it.

This is the failure mode with the worst possible consequences, so it is pinned by tests at both the query layer and the engine layer: an attempt to reach another company’s data fails the suite rather than reaching review.

The second-worst failure is a correct-looking reminder sent to the wrong person, about the wrong balance, or about an invoice that was paid last week. Four things sit between a scheduled action and a customer’s inbox:

• A chain of fifteen deterministic checks that runs immediately before every send, against live database state — not against whatever was true when the message was queued.
• Review mode as the shipped default: every outbound message waits for a named person to approve it until you deliberately enable automatic sending for specific routine cases.
• A freshness pause: if your imported data is older than the window you set, sending stops until fresh figures arrive.
• Exactly-once execution via an idempotency key per action, so a retry, a double click or a re-queued job cannot produce a second email.

On top of that: send windows, minimum intervals between contacts, and a cap on how often any one customer can be contacted in a thirty-day period. Account-level holds and do-not-contact flags stop everything for a customer immediately.

Every meaningful action writes an append-only audit event: who acted, what changed, and when. “Who” is one of three things — a named person, the system, or the AI — and the distinction is carried through the customer timeline and the message records, not just the log.

The audit trail distinguishes actual rule-based classification from an authorized model read. A model read records the model identifier, confidence and structured output. Replies that rules already resolve are processed without a model. A model read that did not happen or lacks supporting evidence is recorded as such and is not presented as an AI result.

All listed with their purpose on the subprocessors page: netcup for the application and primary database in Nuremberg, Cloudflare for network and security functions, Plus Five Five (Resend) for system emails, and — for the support assistant on this website, where it is switched on — the AI providers listed on the subprocessors page, according to the configured support policy. Connected Microsoft 365 or Google Workspace mailboxes follow your chosen integration; other IMAP/SMTP providers remain your own contractors. Replies from your customers are read by built-in rules first. Only with your consent, a valid contract with verified payment, a configured provider and sufficient budget does Google Gemini read selected, confidently matched business-customer replies that rules could not settle; replies from private persons or customers of unknown type are never sent to a model. Network, email and AI processing can involve countries outside the EU, as explained in the privacy policy.

There are no advertising pixels and no third-party or cross-site trackers on this website or in the application. On the public website we count visits with the open-source software Umami on our own server (statistik.neithra.dev) — without cookies and without an identifier on your device, which is also why you have not been shown a cookie banner. The privacy policy describes this in section 5.

Tell us which controls and contract details your organization needs. We explain the implemented safeguards and agree any additional requirements before your order.