A customer reports a change of address, a policyholder submits an invoice, a supplier sends an order. In many companies, each of these transactions ends up on an administrator’s desk, even though none of them requires a decision. Back-office processing describes the alternative: The transaction runs completely automatically, from receipt to posting, without anyone having to touch it.
The name comes from the fact that the processing remains invisible to employees—in other words, it takes place „behind the scenes.“ In banking and finance, the same principle is called straight-through processing. This article explains how such a process works from a technical standpoint, where it is used in insurance, banking, and industry, and why it regularly fails in practice.
The most Important in a Nutshell
Automated processing refers to business transactions that are processed automatically and without human intervention. It is measured by the automated processing rate, which is the proportion of such transactions out of all incoming transactions.
It is most widely used in areas where a large number of similar processes occur: in insurance companies' input management, in banks' payment processing, and in order entry in the manufacturing industry.
The limiting factor is rarely recognition; rather, it is the exception. Any process for which no clear rule exists falls outside the scope of automation.
What is dark processing?
„Dark processing“ refers to business transactions that take place fully automatically in the background without any human interaction. They occur "behind the scenes" without a clerk seeing, reviewing, or approving the transaction. Dr. Torsten Oletzky describes them in the Gabler Insurance Dictionary as „the IT system’s final processing of a business transaction.“ This form owes its name to the „lack of employee intervention“ and the fact that the processing takes place „out of the employee’s sight (in the ‚dark‘).“.
The key term here is “end-to-end.” A transaction is considered to have been processed “in the background” only after it has gone through the entire process: recognized, verified, posted, and responded to. A system that reads an invoice and then submits it for approval has sped up the process, but has not processed it “in the background.”.
Not to be confused with dark data
Because of the similarity in names, „dark processing“ is often equated with “dark data.” These terms refer to two different things. “Dark data” refers to data sets that a company collects but never analyzes, such as log files or archived correspondence. The goal here is to gain insights from existing data sets.
Back-office processing, on the other hand, pertains to day-to-day operations: a specific business transaction that is received today and is to be completed today. When people look for metrics related to automation, they almost always mean the latter.
Dark processing in 5 steps
A process handled in the background follows the same chain as a manually processed one, only without interruption.
Receipt and Entry. The document arrives as an email with an attachment, as mail in a scanning line batch, via a portal, or as a structured data record. For paper documents and image files, the process begins with optical character recognition; structured formats skip this step.
Classification. The system determines what the message is: a cancellation, a damage report, an invoice, or a change of address. Only the automatic document classification determines which process applies and which fields are actually needed.
Extraction. The transaction is used to extract the relevant details, such as the contract number, date, amount, and bank account information. The Data extraction It does not use fixed coordinates, but instead evaluates text, position, and layout together.
Check against rules and inventory. This is where it is determined whether the transaction will remain off-book. The system verifies the values against the contract portfolio, master data, and business rules: Does the contract exist? Is the deadline met? Is the amount within the permitted range? Do the totals of the line items match? Each extracted value also carries a confidence score—that is, the probability with which the model considers its own output to be correct.
Case-closing entry. If all checks are passed and the confidence level exceeds the specified threshold, the transaction is posted, the policy is updated, the payment is authorized, and the confirmation is sent. Otherwise, it is routed to manual processing, and the correction is fed back as a training signal.
Dark processing is not a one-time project, but rather a state that must be maintained. If a major sender changes its form, the rate drops until the model has caught up.
Fields of Application of Dark Processing
Batch processing is worthwhile when there are a large number of similar tasks and the processing itself requires little judgment.
Public health
In health insurance, “automated processing” primarily refers to claims processing: submitted medical bills, prescriptions, treatment plans, and cost estimates. A case processed automatically is reviewed, assigned to the appropriate policy, checked against the list of covered services and reimbursement rules, and paid out without a claims adjuster ever seeing it.
A survey by the German Insurance Association: In health insurance, „nearly one-third of the processes are now fully automated.“.
The line is drawn based on discretion. A standard billing process within the standard rates proceeds as usual. An invoice that includes a justification for exceptional treatment requires a professional assessment and is therefore not a candidate for automation.
Financial services
Back-office processing in banking begins with payment transactions, where transfers have been processed for decades without any human intervention under the term “straight-through processing.” In document-driven business, the proportion is smaller, but the areas for improvement are clearly defined: opening an account with identity verification, submitting pay stubs as part of the loan application process, address changes, and tax exemption requests.
The loan application illustrates this pattern particularly clearly. An applicant’s documents—such as pay stubs, bank statements, and identification documents—can be scanned and checked against creditworthiness criteria. The bank retains the final say on the loan decision, but everything leading up to that point can be handled automatically.
Retail
In retail, “dark processing” primarily involves the reconciliation of documents between stores, headquarters, and suppliers: goods receipt documents against purchase orders, supplier invoices against terms and conditions, and return slips against sales receipts. The key to efficiency here lies less in the individual process than in the volume, because these same three types of documents are generated daily across all locations.
Manufacturing
In the manufacturing industry, the focus is on order entry. Many manufacturers still receive orders as PDFs or faxes, which are then manually entered into the ERP system. Instead, a “dark-processed” order is scanned, checked against the product master and price list, and created directly.
Logistics and transport
Waybills, delivery slips, and customs documents largely follow standardized forms and are generated per shipment, not per customer. This combination of high volume and consistent format makes them the ideal candidates for automation. Verification against the shipment data in the transportation management system is fully automated, provided the scan quality at goods receipt is adequate.
insurance
"Dark processing" in the insurance industry is the best-documented case, as the GDV regularly tracks the rate here. According to its figures, property and casualty insurers reached a hidden claims processing rate of 33.5 percent in 2023, up from 23 percent four years earlier. That’s an increase of 10.5 percentage points over four years—a significant acceleration rather than gradual progress.
According to the GDV, this encompasses business processes „such as new applications or the issuance of insurance policies, changes to contract data, or the processing of claims or contract cancellations.“ These exact types of transactions are the typical candidates. Separately, the GDV reports how many contracts are concluded without any human intervention: 19.1 percent across all lines of business, and 24.1 percent in auto insurance—just under a quarter. A glass damage claim with a repair shop invoice within the standard framework can be processed entirely automatically. A liability claim involving a disputed question of fault cannot.
Energy and supply
Utilities process meter reading reports, supplier changes, and advance payment adjustments in large volumes through a largely standardized process. The market process is regulated and therefore easily manageable.
Education
Colleges and educational institutions process applications, enrollments, and certificates in seasonal waves. It is precisely these peak periods that make automation attractive, because they can be handled without additional staff.
Advantages of Dark Processing
The benefit does not stem from the fact that a machine types faster than a human, but rather from the fact that an entire processing step is eliminated. This results in five interrelated effects.
Lead time. A process handled efficiently can be completed in minutes rather than days. This has an immediate impact on customer satisfaction because the response arrives while the customer still has the issue fresh in mind.
Cost per transaction. Setup is a one-time cost, whereas processing is required for each individual document. That is precisely why the impact increases with volume and is barely noticeable for small quantities.
Scaling During Peak Loads. Automated routes can handle the New Year's holiday or severe weather conditions without the need for temporary workers. For seasonal businesses, this is often a stronger selling point than mere cost considerations.
Consistent quality. A machine applies the same rule to the thousandth operation as it does to the first. Deviations do not result from fatigue, but from poor templates, and can therefore be systematically corrected rather than being random.
Relief when staff is short. When there is a shortage of skilled workers, "Dunkelverarbeitung" shifts available working hours to the cases that actually require assessment. Administrative work does not disappear; it is dedicated to handling the difficult cases.
Dark Processing Challenges
Obstacles rarely lie where projects expect them to. They are spread across the incoming data, the setting of thresholds, the monitoring of the invisible, and the regulatory framework.
The exception determines the ratio. It is not recognition that limits automation, but rather special cases. Partial shipments, credit memos, different recipients, or transactions with free-text explanations are technically correct but rare. That is precisely why the model lacks examples of these cases, and that is precisely why they end up being processed manually.
The quality of the evidence is more important than any model. A photographed document with creases and shadows produces poorer results than a digitally generated PDF. Where input channels can be controlled, switching to structured formats has a greater impact than any model optimization.
Threshold values come at a cost on both sides. A confidence threshold that is too high sends correctly identified transactions to the audit, thereby reducing the accuracy rate. A threshold that is too low allows errors to slip through that are only detected during post-processing, where correcting them is more expensive than the initial audit.
Errors become invisible. If no one looks at something, no one notices it. That’s why dark processing requires spot checks and metrics for the rework rate; otherwise, a systematic extraction error will go unnoticed for months. Visible checks also solve the acceptance issue within the team: People who have been reviewing processes for years are suspicious of a machine that does the same thing out of sight.
Regulatory commitment. In insurance and financial supervision, automated decisions must remain justifiable. With regard to personal data, the following also applies: Article 22 of the GDPR, which makes automated decisions in individual cases subject to certain conditions. This does not rule out automation, but it does require documentation and defined options for human intervention. This is where it pays off to log every automated step: With manual processing, it often remains unclear why a case was decided the way it was.
AI software for dark processing
Back-office processing stands or falls on the question of how many transactions successfully pass through the verification stage. Konfuzio’s software covers the entire process: It reads incoming documents of any type, assigns them to a category, extracts the relevant data, and transfers them—along with confidence scores—to the downstream system.
The difference from traditional template systems lies in how they handle variations: A rule-based layout template breaks as soon as a sender changes their form, whereas a trained model can still identify the value even in a new location. In practice, three characteristics are crucial. The models can be trained on a company’s own document types rather than being limited to standard forms. The confidence threshold is adjustable, allowing each company to choose its own balance between automation and manual oversight. And the system can be operated both in the cloud and on the company’s own servers, which is often a requirement in regulated industries.
The full extent of the chain's impact only becomes apparent afterward: Only when the data flows into the subsequent process—for example, into a Digital Invoice Approval, the manual step really does disappear. Anyone who simply extracts the data and then passes the values along manually is just shifting the work around instead of eliminating it.
Conclusion
Automatic processing takes place in the background, without a staff member seeing the process. It is therefore not an analytical method, but rather a statement about the operation: How many business transactions are processed from start to finish without requiring anyone’s involvement.
The size of this proportion depends less on the model than on the boundary conditions: the quality of the input channels, the number of similar transactions, appropriately set thresholds, and whether the data actually reaches the subsequent process. Anyone who wants to increase this rate should therefore not start with the detection rate, but rather with the question of which exceptions are currently disrupting the workflow. Our article on Key Metric: Dark Processing Rate.
A robust test doesn't cost much: 50 to 100 real transactions from your own inbox, deliberately including the problematic cases. How a system handles real-world exceptions says more than any specification in the datasheet. Talk to our experts, if you want to measure and increase the water flow in your home.
EN 
DE 
ES 
FR