Auditing business process automation: where does the task really end?
Review automation from trigger to received result, including data, exceptions, retries, permissions, owners and manual handover.
Article language: English

In this article
The dashboard shows success, but the recipient is still waiting for the result. The automation has checked off its task; the business has not. An automation audit should trace the journey from input to the right recipient, including failures and retries. Check that the right file was processed, no action was performed twice and an exception reached a person. Green indicators are useful, but they cannot certify themselves.
An automation audit starts by reconstructing the flow, then examines its boundaries. It applies to simple connections between systems and processes containing an AI stage. The aim is a map of operations, gaps and responsibilities. There is no need to rebuild the tools immediately. First establish what the automation does today, who receives its result and what happens outside the successful scenario.
Establish the start and end of the process
Record the triggering event: a form, a record change, a new file, a schedule or a manual decision. Check how the automation knows a task is new and ready to handle. If the source is filled in gradually, starting too early can produce an incomplete result. Understand that condition before analysing later steps.
Describe the end from the recipient's perspective. A created file, a message waiting to be sent and a received response are different states. Assign an owner and a completion signal to each. The AI use register (in Polish) can connect the process to a decision to maintain it, rather than treating the automation's existence as a reason to keep it running.
Reconstruct a real run
Choose an approved example and follow it from input to result. At each step record the system, action, data, executing account and outcome. Compare the configuration with a conversation with the person operating the process. Important work may happen manually between formal stages, such as correcting a file before sending it or checking whether the recipient received it at all. If the automation processes many cases at once, separately check mapping AI batch results to records (in Polish), including when responses return in a different order.
Keep the run identifier and observation time. If identifiers change between systems, establish how to connect them. Otherwise, an error at the end may be hard to trace to the source material. The map should include storage and sharing locations, as in the AI data-flow register (in Polish), not just arrows between product names.
Separate fixed rules from AI judgement
Check which steps have a fixed condition and which depend on a model's interpretation. Copying a field into another system needs a different review from assessing the meaning of a request. For the AI stage identify the instruction, sources, permitted result and behaviour under uncertainty. A result that cannot be assigned unambiguously to a category needs an explicit handling path.
Establish whether an AI answer automatically triggers a later action. If so, check the action boundary and necessary approval. The responsible agent workflow (in Polish) describes that decision point. Later changes to an instruction or model should be assessed using regression tests (in Polish), rather than relying on one historical run.
Examine retries and duplicate execution
Establish what happens after a connection breaks. An automation may retry even though the other system completed the previous attempt. If a task creates a record, sends a message or places a paid order, you need a way to recognise repetition. Check whether the process identifies the same task and can distinguish a resume from a new request.
Run the test on controlled material with an agreed way to reverse its effects. Record what happens after a retry, not merely the presence of a configuration option. Power Automate limits documentation shows that retry policies and limits depend on the service and plan. Review the configuration actually in use.
Check incomplete data and unusual input
Prepare an empty attachment, a missing required field, a changed file name and material in an unexpected format. Establish whether the process shows a clear error, sets the task aside or sends it to a person. Cases where the automation appears to continue and produces a result from incomplete input deserve particular attention.
Assign an owner to every exception path. If failed tasks go to an inbox, check who reviews it and how they identify a backlog. A notification without a recipient does not complete handling. Review frequency and escalation rules should match the consequences of delay. Do not invent an arbitrary deadline just because it is easy to put in a procedure.
Assess permissions and account dependencies
Check whether the automation uses an employee's account, its own application identity or another connection. Establish the owner, scope of capabilities and how access is renewed. Ask what happens when a password, role or service plan changes, or a person leaves. These events often reveal a dependency absent from the original process description.
Review reading, writing, sending and deletion capabilities. Each should correspond to a stage of the task. An AI integration permissions review (in Polish) helps plan a test of an allowed action and a refusal outside scope. The business owner should understand the result even if an administrator handles configuration. Technical permission is part of the process's operating conditions.

Verify stopping and work already in progress
Prepare a procedure for stopping new runs and a separate one for handling tasks already started. They can have different effects. Microsoft states in the referenced documentation that turning off a Power Automate cloud flow stops new runs while in-progress runs can continue. Check the rules of any other product separately.
In a test establish what the operator sees, where the queue remains and how to prevent an unintended restart. Record the authorised person and a substitute. If stopping may lose material or lead to repeated execution, address that in the instructions. If improper operation is suspected, use the data-incident procedure (in Polish), which organises coordination and protection of information.
Plan manual handover
Describe the minimum material a person needs to finish the task. They should know what has been done, what is waiting and which effects have already occurred. An error message alone may not be enough if some actions completed in another system. Handover should reach a named person or a queue with an acceptance rule.
Test handover with someone who did not build the automation. Check whether they can identify the state, perform the next stage and close the task without duplicating an effect. If they need a private message from the creator, record the documentation gap. A prepared substitute reduces dependency on one person, but must be tested in practice. Naming someone alone is not enough. Set out how to finish cases while a tool is unavailable in an AI outage continuity plan (in Polish).
Practice example: handling form enquiries
A form creates a request, AI proposes a category and the system directs the task to the appropriate person. The review covers a valid entry, missing contact information, an unusual question and resubmission of the same request. The reviewer also checks whether the category affects later actions and whether an uncertain answer reaches a person.
The team then interrupts a controlled run before the result is saved and examines resumption. It establishes how to recognise an already created record and who handles an exception. The example does not assume that AI improves quality or reduces cost. Only observation shows where an improvement makes sense and what maintenance it requires. A valuable outcome may be deciding to retain a simple step as a fixed rule without model interpretation.
Show the cost of exceptions and maintenance
Collect service fees, handling effort, checking and correcting tasks. Separate successful runs from those requiring handover. If automation shortens one team's work but creates an exception queue for another, the calculation should cover both. Otherwise a convenient demonstration can hide a shift of work.
The total cost of an AI process (in Polish) describes the calculation. Also check maintenance of documentation, permission reviews and reassessment after changes. Each needs an owner. If no one has time to do them, that limits the automation's scope and should be presented in the decision about continued use.
Record findings as work that can be accepted
For each gap record the observation, effect, evidence, owner and acceptance criterion. Instead of “improve monitoring”, specify which state must be visible, to whom and how it will be checked. Priority follows impact and urgency, not how impressive the technical change looks. Separate quick containment from later rebuilding where useful.
Acceptance of a fix should repeat the case that exposed the gap and check important dependencies. There is no need to run another full review after every small change if the evidence covers the scope. Keep the result and configuration version. The next owner can then understand what the fix addressed and which operating conditions remain current.
What should the owner see without opening configuration?
Prepare a view that distinguishes completed tasks, those awaiting a person, failed and stopped tasks. Check late status handling in automation (in Polish), so a later message is not mistaken for a newer case state. Assign each state a definition and a response owner. A count of runs can help but does not alone say how many results were received. If a technical dashboard and business report use different terms, document their mapping so the team does not compare different quantities under one name.
Check the view with a known example from each important path. The owner should be able to find a task and identify the next step without the automation author's help. Also establish how to recognise an absence of events: no new requests may be normal or may result from a broken trigger. The right check depends on the process. The review should identify a source that distinguishes those situations and a person to examine discrepancies.
How do you maintain the process map after an audit?
Identify events requiring an update: changes to the source, model, permissions, owner and acceptance method. Connect the description with company knowledge architecture (in Polish), so the current instructions are available to the operator. Include retiring the process and organising its data if the solution is no longer needed.
For a first review, choose an automation with a noticeable effect or one nobody can fully explain. The Audit offer helps define the scope.
A separate acceptance case is a retry without a duplicate effect (in Polish). Check what happens after a lost response or a second click. A connection error message does not prove that the first operation was not performed.











