Your accounting program sends an invoice and waits. Then it displays “connection interrupted”. You are tempted to click Send again. But the message describes what your program experienced, not necessarily what the provider did. The invoice may never have arrived. It may be waiting for processing. Or the provider may already have accepted it, while the answer failed to reach your program.
This is why a valid request can still lead to a business problem. Recognizing the sender, allowing the operation and checking the format are important. They do not, on their own, decide what to do when an answer is missing or a permitted action repeats. Security guidance separately considers harm from excessive use of legitimate business activities. 6
Sending and receiving an answer are separate events
A request is a message asking another program for information or work. A response is the answer sent back. The agreed arrangement for these exchanges is called an API. Your accounting program uses the provider’s customer API to submit invoice details or ask what happened to a document. The receiving program runs on the provider’s computers, often called servers. 13
Follow one invoice for the bicycle shop. First, your program prepares the submission request. Next, it sends that request. The provider then receives it and carries out whatever acceptance steps its service requires. Finally, it sends a response, which must travel back before your program can display the result. An interruption at different points can leave your program with the same visible uncertainty.
A waiting limit can also expire. This is a time limit after which a program stops waiting for an answer rather than waiting indefinitely. Reaching that limit tells you that the answer did not arrive in time. It does not by itself tell you whether the receiving service finished the work. Supplier instructions need to distinguish a known refusal, a known acceptance and an unknown result. Documentation of requests, responses and errors is part of API management guidance. 7
The next step should therefore not be left to a guess. Depending on the actual service, the program may be instructed to ask for the existing submission’s status, repeat a request in a defined way, or refer the uncertainty to a person. We are describing choices the supplier must explain, not prescribing one universal mechanism.
Another attempt need not mean another intended invoice
Imagine ordering goods by phone and losing the call just before confirmation. When you call again, you want to clarify the original order, not order the same goods twice. Software needs an equivalent way to distinguish a repeated attempt from a genuinely new business instruction.
One possible design gives the intended submission a reference that is retained across repeated attempts. Suppose our fictional connection labels an operation “submission-A”. If it has to make another attempt, it uses the same reference with the same intended content. The receiving service can, if designed to do so, recognize that this is the existing operation rather than create a new one. The reference is not an access password; in this example it identifies which intended operation the attempts belong to.
This explanation is a proposed design to discuss with the provider, not a claim about every API. If such a reference is supported, the instructions must say how long it remains meaningful, how the result is retrieved, and what happens if somebody changes the contents while retaining it. Otherwise the words “duplicates are prevented” may leave the practical problem unresolved.
A second real invoice or a correction is not merely another attempt at the first submission. An accountant may intentionally issue a different document. Rejecting everything that looks similar can therefore obstruct legitimate work. Your accountant and the person maintaining the connection need the same explanation of what counts as one intended submission.
Peppol does not erase this customer-side question. An Access Point is a service that exchanges messages in Peppol; the AS4 profile describes communication between Access Points. Your program’s customer API connection is another stage. A supplier operating both should explain the connection between their statuses instead of implying that the transport rules automatically define every customer retry. 10
An allowed action can be used at the wrong time or too often
Suppose the invoice service also offers payment reminders. A valid connection might be permitted to send one. If faulty automation asks repeatedly, the customer could receive reminders much more often than intended. There need not be an intruder. The program may be faithfully repeating a badly designed instruction.
The useful restriction follows the business consequence. Repeating a status question is not the same as sending another reminder or creating another charge. Similarly, editing a draft may be allowed while changing an already approved document requires a different process. These are examples of business-specific rules to agree with the service, not universal legal claims about invoice correction. OWASP treats protection of sensitive business flows as a separate concern because permitted operations can still cause harm when excessively used. 6
Now consider an export: a file assembled from stored records for use elsewhere. One request for a year’s records can require much more work than one question about an invoice’s status. The receiving service needs processing time, temporary working space and storage to build the export. Merely counting messages cannot explain all the workload. API guidance considers input sizes, processing resources, operations within a request and spending on connected services. 4
A limit on requests can help stop a broken connection from doing unlimited work. But a limit set too low can prevent a legitimate month-end task. Ask which activity is limited, what happens when the limit is reached, and how an unusually large legitimate export is arranged. Do not generate traffic to discover the limit yourself. Appropriate restrictions should be chosen for business needs. 4
A cost warning is not a spending ceiling. An email saying that expenses have risen can prompt a person to act; it does not itself stop the next charge. Guidance distinguishes spending limits from billing alerts where limits are not available. Clarify which protection the service supplies and who handles a malfunctioning connection or resulting charges. 4
A delay may come from another service
Your provider may use another provider for an additional task. If it does, it sends a request and waits for that service’s response too. An unavailable or unexpected answer can therefore affect the invoice workflow. The invoice service should check incoming information and handle waiting limits rather than trusting another API merely because it belongs to a supplier. 9
The business question is what remains unfinished. Does the invoice wait because the task is essential? Does only an optional extra wait? Is manual help required? These are alternatives to explain for the actual service. “Another service was slow” is not a complete answer unless your staff also know what happened to the invoice.
Ask the provider to demonstrate an uncertain result in an expressly authorized, separate test environment. It can use an invented invoice, temporary accounts and a simulated missing response. Low test limits can demonstrate handling without creating load; non-chargeable arrangements avoid sending paid reminders. Agree the expected result and stopping conditions first, then check final records and waiting work. No interruption of live work or real invoice transmission is necessary.
Before using a new connection, resolve this one sentence: “We sent it, but we did not receive the answer.” Your staff should have a documented way to establish what happened and what remains to be done, rather than an unexplained instruction to keep clicking Send.
Sources
6 — OWASP API6:2023 Unrestricted Access to Sensitive Business Flows
13 — MDN: Overview of HTTP — client/server requests, responses and sessions
7 — OWASP API9:2023 Improper Inventory Management
10 — OpenPeppol AS4 Profile 2.0.3
4 — OWASP API4:2023 Unrestricted Resource Consumption
9 — OWASP API10:2023 Unsafe Consumption of APIs