Skip to content
Interface language: English
Audit / Signal

An automation exception queue that does not become a digital waiting room

Give stopped cases an owner, a deadline and a safe next step. Design an exception queue that the team can actually resolve.

Article language: English

A person holds an instrument near the dark blue hull of a boat supported in a coastal workshop.
In this article

“Sent for manual review” sounds reassuring. But who will review it, and by when? If the answer is a shared folder nobody checks, the automation has created a waiting room. An exception queue should help resolve stopped cases. Each entry needs a reason for stopping, an owner, a next step and a condition for closing it.

Design it alongside the automation, before the first backlog appears. A mechanism that sets aside failed messages does not itself serve the customer, correct data or retry safely. It is only the waiting area. The company still has to arrange cover and know what happens when a number is called.

What is an exception and what is an ordinary task?

Distinguish a technical outage, missing data, an ambiguous decision and a case outside the process scope. They require different actions. A retry may help after a brief service interruption, but it cannot supply a missing case number. Three more attempts will not establish whether a customer changed an order's scope either.

Imagine an enquiries workflow in which a model identifies a category and drafts an answer. An unavailable ticketing system goes to technical support. A message with two conflicting deadlines goes to the case owner. A request about a service the company does not provide needs a separate handling rule. Putting everything in a bag labelled “AI failed” makes the process harder to improve.

Record these distinctions when auditing business-process automation (in Polish). A well-defined exception shows the boundary of automation. Not every exception must eventually be removed by a better model: some should remain a deliberate human decision.

A technical queue and a team's work list

Infrastructure tools can move undelivered or unprocessed messages to a dead-letter queue (DLQ), a separate queue for such messages. For example, the Azure Service Bus documentation describes inspecting, correcting and resubmitting them. This particular service does not automatically clean its DLQ. That is product-specific behaviour, not a universal rule for every queue.

Such a message may not contain the information an employee needs. An operational list should identify the case, the meaning of the problem, the responsible person and available actions. The team need not inspect the whole technical request to ask a customer for a missing attachment.

Conversely, a short task saying “check error” does not tell an administrator which step failed. Connect the two layers using an event identifier. Design AI logs without unnecessary data (in Polish) so the team can trace the flow without copying confidential content into every subsequent list.

An exception card should make the next decision easier

Prepare a short set of information for each entry:

  • Which case stopped, and at what stage?
  • What has already happened, and which outcome remains unconfirmed?
  • Why does it need a person or a technical repair?
  • Who owns the next step, and by when?
  • Which actions are allowed, and what will confirm completion?

Do not equate the owner with whoever happens to be signed in. You also need cover and an escalation route. A transfer to another department should leave a record and clearly identify the new responsibility. Clicking “assign” does not guarantee that the other person has accepted the work.

Connect role rules with a responsible AI-agent workflow (in Polish). An agent may prepare a description of the problem, but should not invent a business decision simply to clear the queue. A clean screen does not always mean work is in order. A lack of grounds for an AI answer (in Polish) may itself be a reason to send a case to a person. Record that separately from a technical outage: it needs a different resolution.

Define when a case's age starts

Age may be measured from the event's creation, its first failure or its transfer to the queue. These are not interchangeable moments. A message that has circulated for two days does not become new when it changes folders.

AWS explains for Amazon SQS that, in a standard queue, expiry is based on the original enqueue time, while the DLQ's ApproximateAgeOfOldestMessage metric indicates time since transfer. This example shows why you must check the definition of the particular metric before treating it as a customer's waiting time.

In the work view, distinguish the business deadline from time spent in the exception queue. Priority may depend on both, as well as on the effect of delay. An uncommon order with an imminent deadline should not disappear behind a thousand easy, low-priority records. Agree the rules with the people who bear the consequences of the handling order.

Retry when conditions have changed

A “try again” button needs a reason: the provider is available again, a field has been completed or a rule corrected. If nothing has changed, retries may merely add events to the same problem. A loop of hope is still a loop.

Before retrying, check whether the earlier attempt completed part of the operation. A message may have been sent even though the system did not record the confirmation. Rules for duplicates and retries (in Polish) should protect the recipient from a repeated effect and let the operator establish the state without guessing.

After correcting data, show exactly what will be used. Approval of the old case version does not automatically cover new content. For a consequential action, use human approval of AI actions (in Polish), with the retry's scope and effect visible.

Check handling capacity

For an exercise only, suppose 12 exceptions arrive each day and the team closes nine. The backlog grows by three a day; after five such days there are 15 additional cases. A prettier dashboard will not change that arithmetic. Reduce arrivals, increase handling capacity or limit the automation's scope.

Also count the time spent on an entry and recurring causes. A dozen exceptions requiring one click pose a different problem from a few cases that each need agreement across three departments. An average can conceal that difference.

Aeneia conceptual photography

If uncertain cases increase after an AI model change (in Polish), the migration assessment should include the load on the queue. The model may classify content more cautiously, while the company needs additional handling capacity. Decide on the effect across the whole process.

Closure must describe what happened to the case

Distinguish completion after correction, manual handling, deliberate cancellation and combination with an existing ticket. Each outcome needs a short reason and confirmation of the appropriate effect. Removing a record from the view does not explain whether the customer received an answer.

Run a test while the automation's author is absent. Another authorised person should find the exception, understand its state and safely complete the work. If they cannot, improve the instructions or interface. This scenario also supports a continuity plan for an AI outage (in Polish).

Distinguish waiting for an external answer from inactivity

A case may legitimately wait for information from a customer or a provider's repair. That state needs a reason, a review date and an owner. Moving it to an “awaiting” tab that nobody reviews is insufficient. The move does not stop time for the person who needs the result.

Record when the question was sent and the next permitted action. After the deadline, that might be a reminder, contact through another channel or deliberate closure under an agreed rule. Not every case needs automatic escalation to a manager. The choice should be agreed and reconstructible rather than depend on the person on duty remembering.

When transferring work to another team, distinguish the case owner from the person performing the repair. A customer representative may still own communication while an administrator fixes an integration error. This keeps the external recipient informed while technical work proceeds. The card only needs to show the two tasks and their dependency clearly.

Also check cover during an absence. An authorised person should see the context, last agreement and deadline without asking the absent worker for private notes. If cover requires broader access to content, assess that scope. Convenient rota cover does not justify opening the entire queue to everyone in the company.

Group a shared outage without losing individual cases

One unavailable service can create many exceptions. An operator needs a shared view of the cause, but each case still has its own outcome. Grouping by outage helps assess arrivals and prepare a repair. It must not turn dozens of different operations into a single task whose closure automatically confirms every effect.

Use a synthetic batch of three records: the first never began the operation, the second partly completed it, and the third succeeded without its confirmation being recorded. Then simulate service recovery. Check whether the process selects the next action according to each record's state. A shared retry button needs the same checks as an individual operation, applied across the group.

Limit notifications about the same cause while retaining access to detail. A hundred identical alerts can effectively hide the hundred-and-first, unrelated case. The person on duty should see whether the group is growing, what caused it and whether a case requiring a different response has appeared. Silencing messages does not remove the work.

After repairing the cause, inspect the rest of the queue. Some records may be stale, handled manually or supplied with new data. Do not replay an old intention without checking the present state. A rule rejecting stale work should retain a reason and say whether further human handling is required. For late confirmations, check handling delayed automation statuses (in Polish) so an older message does not decide an already handled case again.

Accept the queue by testing one complete shift handover

Prepare several synthetic cases with different causes and priorities. The first person reviews them, takes some actions and hands the rest to the next person. The second works only from information recorded in the workflow. This test shows whether the case card is sufficient when the author of the previous decision is not beside them.

Include an overdue task, an entry awaiting a provider and a failure requiring confirmation of its effect. Ask the second person where they will start and why. If the order follows only the table layout, establish clearer priority rules. Manual handling needs design too; a person is not a universal plug-in for everything the process left undefined.

After the exercise, count cases resolved, transferred and still open. Explain the difference between closing the cause and completing the handling. An administrator fixing a fault does not prove the outstanding actions were performed. The team needs to know who confirms each case's final state and where that evidence is recorded.

Choose one organisational and one technical improvement from the test. These might be adding the next contact deadline and displaying partial execution more clearly. Repeat the affected scenarios and retain the result. A queue becomes useful through verifiable improvements, not merely because its project description includes “monitoring”.

Before acceptance, add a sample exception card, the age definition and a closure record to the materials for an AI audit. A reviewer should be able to follow the path from cause to confirmed effect. They do not need the entire error archive if a few controlled examples show the rule and its boundaries. For preparation, see the guide to preparing AI audit materials (in Polish).

Also check what a user waiting for a result sees. An “in progress” message after a stop, with no deadline, may encourage them to submit the same order again. An appropriate status should explain the need for clarification and how to make contact, to the extent suitable for the recipient. Infrastructure details rarely help them decide what to do next.

After launch, hold a short review of the most frequent causes. For each one decide whether better validation, an integration fix or a human decision is needed. Do not try to eliminate every exception at any cost. Some indicate a properly functioning process boundary and protect against hasty execution.

A closed exception can contribute to a future test, but should not automatically change rules. First check whether the resolution was correct and suitable to generalise. Otherwise the automation may start learning shortcuts adopted in haste during a shift. Also check who can close a case without taking action. That ability can be needed, but its use should require an appropriate reason. “Cancelled” must not become a convenient substitute for confirmation the operator did not find. An unexplained entry needs resolution even when it interferes with a report.

The Aeneia Audit offering is a starting point for assessing such backlogs and handling rules.

Aeneia / Aeneia editorial

Editorial article. Source links are included in the text.

About Aeneia

Brands we have worked with

  • L’Oréal
  • Garnier
  • Netflix
  • Carrefour
  • Warner Bros. Discovery
  • Biedronka
  • HBO Max
  • Wolt
  • Żabka
  • COSRX
  • Embassy of China in Poland
  • Mixa
Aeneia / Initium

From knowledge to a decision

Tell us what you want to put in order. We will agree on where to start.

Contact Aeneia

We partner with and use solutions from:

  • Google
  • OpenAI
  • Microsoft
  • Stripe
  • Anthropic
  • Linear