Intelligent features in OXS.eco accounts payable

What do we mean by intelligent features at OXS.eco?

Intelligent features at OXS.eco are three clearly defined capabilities. They support your accounting, they do not replace it.

This separation keeps the process auditable. For your internal audit, for external auditors, for the tax authorities.

How does OXS.eco process an incoming invoice?

OXS.eco processes incoming invoices in seven steps: intake, classification, AI extraction, plausibility check, rule matching, account assignment with a suggestion, approval and posting. Every step is documented.

Seven steps from the e-mail inbox to the posting. Every step logs what happens. Where the system assists you, it learns from your earlier choice. Where it makes suggestions, you decide.

Every step creates an audit trail. When an invoice enters the workflow, OXS.eco documents which classification was chosen. Which values were extracted and how. Which checks applied. Which learning rule was used. Which of your decisions fed into it.

This traceability is central to record-keeping duties aligned with GoBD. And to every internal audit in which you have to explain why an invoice took the route it did.

Which fields does the AI recognize on an invoice?

AI document recognition in OXS.eco extracts the central standard fields of every invoice: supplier, tax number, invoice number, date, due date, amounts, IBAN, payment term, purchase order reference and line items.

OXS.eco uses a specialized cloud AI for document extraction, specifically the invoice parser from Google Document AI. That is a cloud service with recognition models trained on business documents. We use the variant for incoming invoices.

This is how a recognition run works:

1. The document arrives in OXS.eco and has been classified as an invoice
2. OXS.eco hands the document to the AI recognition
3. The AI returns structured results: the recognized fields, their position in the document, a confidence value per field, and the full raw text
4. The data comes back into OXS.eco
5. The plausibility check applies (see the next section)
6. In the account assignment process you see everything as a suggestion

Important: the AI recognition is a recognition component, not a generative language model.

Where recognition reaches its limits:

• Handwritten notes or stamps on documents are usually not captured reliably. This is where you correct.
• Poor scan quality (washed out, skewed, missing pixels) reduces recognition quality. Low confidence values are flagged in the workspace.
• Special formats (for example specialized invoices from particular industries, unusual credit notes, older invoice layouts from some suppliers) are sometimes not classified correctly. OXS.eco then puts the document in front of you and you correct it.
• Complex multi-page documents with embedded attachments, delivery notes in the appendix for example, sometimes need to be split manually.

In all of these cases the system learns from your correction, see the next section. Over time, every manual intervention produces clearer learning rules.

How does OXS.eco check the recognition against your master data?

The plausibility check runs on fixed rules inside OXS.eco, not in the AI. Every extracted value goes through a series of fixed checks against your master data and against mathematical rules.

The check rules run in OXS.eco, not in a language model. Which rule applies, and whether it warns or blocks outright, we define together with you and show in the demo on sample data. That is different from a pure end-to-end AI: when a posting goes through in OXS.eco, it stays traceable which steps it passed. That matters for financial audits, for a tax audit and for any four-eyes approval.

How does OXS.eco learn from your classification?

A learning rule in OXS.eco is tied to the supplier and the document type. When you classify an invoice, the system can store that choice as a template and apply it to the next document from the same supplier.

Every learning rule stores how you classified this supplier in this combination. Which mandatory fields held which values. Which business partner was assigned.

For the next document from the same supplier, this is what happens:

1. The document arrives
2. AI recognition runs
3. The plausibility check applies
4. OXS.eco looks for a learning rule for this supplier and document type
5. If one exists, the stored classification is suggested, or applied automatically, depending on the setting
6. You see the result in the account assignment step and give the approval

The learning is rule-based, not statistical. Meaning: the system does not learn from patterns across many documents in a large database. It stores your concrete decision per supplier. If you say that this is how you classify this supplier, the system remembers exactly that.

Three modes for every supplier

You choose per learning rule how active the system should be:

In all three modes (Skip / Prefill / Automate) you keep control. A learning rule can be overwritten, deactivated or deleted at any time.

How you manage learning rules

OXS.eco provides a central administration view. You see all learning rules per tenant. You can:

• List active rules, filtered by tenant, supplier and document type
• View the details: which fields are pre-filled, which values are stored
• View the history: when it was created and who changed it
• Deactivate or delete rules

This transparency is deliberate. A learning system you cannot inspect is not a help, it is a black box.

How is the AI recognition covered under data protection law?

The AI recognition in OXS.eco runs as a sub-processor under a data processing agreement. Document data is processed briefly in the cloud for the recognition, and is not used to train the models.

In concrete terms: there is a chain of data processing agreements, with you as the controller, MAAS as the processor and the AI provider as the sub-processor. Processing is configured for the EU region europe-west3 (Frankfurt); details on provider roles and data flows are set out in the agreement. The training exclusion is contractually assured, and all downstream steps (plausibility check, account assignment, approval, posting) run in our hosting environment in Germany.

For your data protection documentation and for audits we provide:

• Data processing agreement (DPA) between MAAS and you
• Sub-processor list with purpose, data category and location
• Technical and organizational measures (TOMs) documented in line with Art. 32 GDPR

We answer the most common questions about our use of AI in short form on our trust page. We provide details on provider roles and data flows on request.

What does this look like day to day?

Three scenarios from everyday mid-sized business, from the regular supplier to the special case. Each one shows where the system assists and where your team decides.

Common questions about intelligent features

Does OXS.eco post automatically?

No. OXS.eco classifies automatically, suggests posting records and passes documents into the approval workflow. The final posting into your accounting system happens only after authorized people have approved it, with an audit trail.

Who is liable if the AI misreads an invoice?

Responsibility for accounting decisions rests, as always, with the person responsible for accounting in your company. OXS.eco is a tool that makes suggestions and extracts data. The plausibility layer and the account assignment workflow are designed so that every document is put in front of an authorized person before it is posted.

How soon does the system know a new supplier?

A learning rule can be created after a single manual classification, and it applies from the second document onward. You decide at which point pre-filling or automatic classification starts.

What happens to our document data if we end the contract?

Workflow data in our hosting infrastructure is handled according to the deletion and return periods agreed in the data processing agreement. The AI recognition is designed to process each request on its own, so it does not build up a document store of its own. What the sub-processors we use commit to contractually, we disclose before the contract is signed.

Can we switch off the AI recognition?

OXS.eco also runs without AI recognition. Automatic field recognition is then not available. Classification runs entirely manually, or through learning rules with no cloud component. Some companies choose this for regulatory or security reasons. We discuss it in the setup call.

What sets OXS.eco apart from competitors that advertise AI?

We cannot speak for every single competitor. But in general: where AI is advertised as a blanket claim, it is worth asking about the components. Which models? Which cloud providers? Which training basis? Where is data stored? Where are decisions made? Clear answers to these questions are a robustness test. Marketing vocabulary instead of answers is a warning sign.

Does OXS.eco learn from other customers' data?

No. Learning rules are tied to your tenant. One customer's rules are never applied to another customer's documents. The AI recognition uses pre-trained models whose training basis is independent of OXS.eco customers. Contractually, the documents you hand over are not used for further model training.

Related areas in the OXS.eco universe

Where intelligent features show results first

Three industries in which learning rules and AI recognition typically show a noticeable effect early, each linked to its industry page.