Invoices checked and matched, with your accountants deciding the exceptions

Fewer hours go into routine checking and more attention goes to the invoices that cost money: wrong prices, double invoices and charges for goods that never arrived. Clean invoices reach approval with the checks already done, and every deviation comes with its reason and the documents behind it.

Accounting is one of the fastest-growing AI areas in German companies. In Bitkom's 2026 survey of companies with 20 or more employees, 25% of the companies that use AI use it in controlling and accounting, up from 17% in Bitkom's 2025 survey. In KPMG's study ‘Digitalisierung im Rechnungswesen 2025/2026’, as reported by mybusinessfuture.com in March 2026, 53% of German companies already use AI in accounting or are implementing it.

We do not promise a percentage up front. We measure effort per invoice, the share of invoices that pass without manual work and the deviations found, on your own invoices, before and after.

• Less manual checking per invoice, so a growing invoice volume does not have to mean new hires in accounts payable.
• Money kept: double invoices, prices above the contract and charges for undelivered goods are caught before payment.
• Cash discounts used more often, because checked invoices reach the approver earlier.
• A cleaner month-end close, with a recorded reason for every open deviation.
• First drafts of variance comments for controlling, built only from figures in your systems.

Reading invoices is becoming the smaller part of the job. According to a summary by IT service provider Hagel IT from April 2026, all German companies must have been able to receive e-invoices in B2B since 1 January 2025, and the obligation applies fully from 1 January 2028. XRechnung and ZUGFeRD count as e-invoices, a plain PDF does not. That summary is a secondary source and is not consistent on the transition rules, so check the dates against the Federal Ministry of Finance FAQ or with your tax adviser.

Either way, an e-invoice still has to be checked. It arrives as structured data, but it can carry a wrong price, a wrong quantity or a second copy of last month's invoice just as easily as paper. Checking is the work that remains, and an agent can take it on. By agent we mean a program that carries out steps in your systems and presents the result for approval.

There is a public example. The Frankfurt air freight forwarder Cargo Movers reconciles invoices against its transport management system, and the software vendor Nemox reports more than 28000 invoice lines reconciled, of which 97% needed no correction of the extracted values (logistik-heute.de, September 2026). This is a vendor-reported figure, and Cloudsail did not build that system.

• Three-way match: each invoice line against the purchase order and the goods receipt in SAP or your ERP, within tolerances you set.
• Price and contract check: unit prices, surcharges, discounts and payment terms against framework agreements, price lists and contracts.
• Duplicate detection across invoice numbers, amounts, dates and bank details, including the invoice that arrives by email and again by post.
• Account assignment suggestions: G/L account, cost centre and tax code proposed from earlier bookings for the same supplier, with those bookings shown.
• Invoices with many lines, such as freight or service invoices, checked against the transport management system or the signed service report.
• Foreign and non-standard invoices that do not arrive as e-invoices: read, translated where needed and checked like the rest.
• Questions on your finance handbook, such as which cost centre takes trade fair costs, answered from your accounting policies and chart of accounts with the source shown.

Controllers spend much of the close explaining numbers rather than analysing them. An assistant can answer questions such as why freight costs rose against plan, draft the monthly commentary and prepare variance notes for the management pack, by pulling the figures from SAP or your BI system together with the bookings behind them.

One rule decides whether this is safe: the numbers come from your systems, never from the model. The model writes text around figures it was given, each figure links to the query that produced it, and a controller signs off before anything reaches management. A wrong number in a board pack costs more trust than the assistant saves time.

This field is at an earlier stage than invoice checking. In Bitkom's 2026 survey, 7% of the companies that use AI use it in management, and an undated KPMG page on AI in finance says half of the finance respondents plan AI in financial reporting. SAP also offers its own analysis of review booklets, Financial Business Insights, as described by SAP partner PIKON in 2025. If that covers your reporting, use it. We build where the questions need data from several systems or your own definitions.

The agent prepares, and your accountants decide. They confirm its suggestions and only need to step in where an invoice departs from the usual pattern. You set the approval rules, and they are written down before go-live.

• Every payment release stays with a person, under your existing approval limits and the four-eyes principle.
• Deviations above the tolerance you set, by amount or percentage, go to the buyer or the cost centre owner and are never booked automatically.
• New suppliers and changed bank details are always checked by a person before any payment.
• Account assignment suggestions are confirmed or corrected by the accountant, and corrections are used for the next suggestion.
• General ledger postings, closing entries and tax questions stay with your accounting team and your tax adviser.
• You decide which invoice types may go to approval without a manual check once they pass every test, and you can withdraw that at any time.

The agent sits between your invoice inbox and your accounting system. Invoices come in as e-invoices in XRechnung or ZUGFeRD format, as PDFs in an Outlook mailbox, or from the scanning step of a document management system such as d.velop or ELO. E-invoices are read as data, without text recognition. Other invoices are read by a language model and validated against your master data.

Orders, goods receipts, supplier master data, contracts and past bookings are read live from SAP or your ERP through approved interfaces, with a technical user that may read but not post. Arithmetic, tolerances and duplicate checks are written as ordinary code, so they give the same result every time. The language model does what code cannot: compare free-text invoice lines with order items, read unusual layouts and write the reason for each exception in plain words.

The result goes into the step where your team already works: a parked document in SAP that a person releases, or, where the books are kept in DATEV, the route your tax adviser already accepts. Many companies have their bookkeeping done by a tax firm on DATEV, so we agree the handover with them. Which DATEV route fits depends on your systems and contracts. We check that in the first step, before promising any automation.

Invoices carry supplier bank details, prices and personal data, so the system runs in your own cloud account or on your own servers. On your own servers it uses an open-weight model, a model whose weights you can run yourself. All model calls go through one LLM gateway, a controlled access point to all models with keys, budgets and logs, so you can see the cost per checked invoice.

Before building anything, look at what you already have. Invoice and document management products such as d.velop, ELO, DocuWare or Candis, and the invoice functions in SAP, already capture invoices and route approvals. If yours does this well, we do not rebuild it. We add what it does not cover: checks against your contracts and your freight or service records, a reason for every exception, and answers from your own finance policies. A custom build pays off when the checks need data from several systems, or when invoice data must stay on your own servers.

Before go-live, the agent passes an acceptance test on a set of your past invoices whose outcome you know, including the ones where your team found errors. We measure each check on its own: deviations found, false alarms, duplicates caught and account assignments right at the first suggestion. The pass mark is agreed with you in writing before the test.

Then the agent runs in shadow mode: it checks live invoices alongside your team without any effect, and we compare its findings with your team's decisions. Only invoice types that meet the agreed mark are switched on. Buyers are right to ask for this: in Bitkom's 2026 survey, companies that use AI name poor result quality (26%) and a lack of traceability of results (25%) among their barriers.

After go-live, evals show whether the system still works as it did in the test. Evals are repeatable tests of an AI system on fixed cases. We rerun them after every change to rules, prompts or models, and a monthly report shows exceptions, corrections and cost per invoice. You get the runbook and can run the system yourself or have us run it.

Buyers ask whether AI accounting tools are GoBD compliant. The technical side: the agent does not replace your accounting system, which stays the system of record. Every suggestion, the documents it used, each check result and who approved what are stored unchanged and can be traced from the posting back to the invoice, and your procedure documentation describes the new step. Whether the whole process meets the GoBD, the German rules for proper digital bookkeeping, you decide with your tax adviser or auditor. This is not tax or legal advice.

Invoices from sole traders and employee expense claims contain personal data, which is one more reason to keep processing on your own servers or in your own cloud account. Your data protection officer decides on the details. Approval logs can also show how fast each accountant works. Under § 87 (1) no. 6 BetrVG, co-determination applies when a system is objectively capable of monitoring performance or behaviour, as Bitkom's 2026 guide on AI and co-determination explains. We therefore design the logs for the audit trail, with pseudonymised user IDs and agreed retention, and leave per-person reports switched off unless a works agreement allows them.

Invoice checking that prepares work for people is a different matter from credit decisions. Creditworthiness assessment and credit scoring of natural persons are high-risk uses under the AI Act (Annex III), with obligations for deployers from 2 December 2027 after the Digital Omnibus. For a deployer that means, among other things, human oversight by trained staff, monitoring of the system and keeping its logs. We do not automate such decisions: in receivables or customer credit checks, the agent may collect documents, but a person decides.

Start with one invoice stream where the volume is high and the checks are clear: freight invoices, suppliers with framework agreements, or service invoices with many lines. In the free 30-minute first call we look at that stream, the systems it touches and how you check it today. If it is worth doing, the first step is a baseline measured on your own invoices, with a fixed price for each step, agreed in writing after the call.

• A rough number of invoices per month in that stream, and where they arrive.
• Your systems: SAP or another ERP, DATEV in house or at your tax adviser, and your document management system.
• Examples of errors your team found in the past, and what they would have cost.
• Who approves which amounts today.