Skip to content
Interface language: English
Audit / Signal

An AI outage at work: how to prepare a continuity plan that works

Prepare priorities, manual takeover, a case queue and a controlled return when an AI service becomes unavailable.

Article language: English

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

AI stops responding halfway through a task. One person waits, another clicks again, and a third starts manually. Suddenly the team has to establish which of the three versions is right. A continuity plan should prevent that: it should say what can wait, what a person will take over and how to recognise an operation already performed. Prepare it for a specific process. A second subscription does not become a contingency plan simply because it sits beside the first.

The procedure below is for preparing an exercise for a particular process. Its owner must confirm assumptions about timing, staffing and priorities.

Which outcome must survive the tool's unavailability?

Describe the minimum outcome in the recipient's language. A customer should receive confirmation that a case was accepted, support should know the response deadline, and an important request must not disappear. Generating an extensive summary can wait. This distinction lets the company maintain a simpler service without recreating every assistant feature.

Use the register of AI uses (in Polish) to identify processes most dependent on model availability. Do not select them by call volume. The most frequent task may be easy to postpone, while a rare step blocks a document handover or another department's work. Priority follows the effect of the interruption.

For each outcome, set the tolerable delay and the required currency of information available after recovery. Google Cloud describes these categories as RTO and RPO: recovery time objective and recovery point objective. Explain them in ordinary language in the company plan so the process owner understands the commitment.

An objective is not a guarantee. If the company promises manual takeover within an hour, an exercise should demonstrate that people, documents and access to the target system are available. Only observation lets the team assess feasibility. If resources are insufficient, change the promise or arrange additional support before launching the dependent process.

How do you recognise an outage and start the fallback mode?

Define symptoms that require a response: recurring errors, no reply within an agreed time, connection unavailability or no operation confirmation. Slower model text alone need not mean the whole process has failed. A single person's access problem is different too. A short diagnostic card limits unnecessary switching for all employees.

Name the role that declares fallback mode and informs the team. The message should give the problem's scope, the way to work and the time of the next decision. It need not promise a repair deadline outside the company's control. “Record new cases in this queue” can help employees more than an unconfirmed recovery forecast.

Treat a provider status notice as an additional signal. Your own integration may remain broken after a reported fix, for example because a connection expired or tasks accumulated. The state of your own process determines when to switch modes. Prepare the verification method with the materials for an AI audit (in Polish).

Also record when diagnosis must end. If safe resumption cannot be confirmed within a company-defined period, the operator should start the agreed fallback path. Otherwise a day can disappear into short attempts, each apparently the last necessary fix.

How do you take over without doing the same work twice?

First establish the last confirmed state. Distinguish starting generation, preparing a proposal, approval and execution in the target system. If no response arrived after a request was sent, the outcome may be unknown. Blind repetition can be dangerous when the task creates a record, sends a message or changes an order status.

For each stage, prepare instructions saying where to find the record and which identifier links it to the case. A business automation audit (in Polish) can help reconstruct those confirmations. A contingency plan does not need the whole technical manual: the operator needs a short answer about whether the next action is permitted.

Manual takeover should change the task owner. If automation can resume in the background, parallel execution becomes possible. Prepare a way to stop the relevant scope or mark cases being handled manually. Check that all subsequent stages, including delayed ones, respect that mark.

Do not expand an employee's permissions as a side effect. An outage is not a reason to send data to a personal account or approve operations without agreed checks. AI-agent boundaries (in Polish) need a manual equivalent: who prepares, who approves and who performs the action.

How do you prepare a manual queue?

A fallback queue needs a case identifier, intake time, priority, owner and next step. Choose a place available during the outage with suitable information protection. Do not assume the everyday documentation tool will work if it depends on the same broken connection or account.

Agree priority rules before an incident. They can account for deadlines, customer impact and whether postponement is safe. Priority should not depend on who complains most loudly. The operator also needs authority to defer less important work, otherwise the manual queue will quickly exceed the team's capacity.

Aeneia conceptual photography

Prepare a simple message for the recipient. It should state the real situation and next step without pretending automated service is normal. If response times lengthen, the company can deliberately change scope or order of work. Concealing the problem often produces more enquiries and additional load on the same people.

Check how the queue is updated. Who marks completion? Who sorts abandoned entries? How does the next shift learn which cases it has taken over? These everyday details matter more to continuity than an elaborate architecture diagram. A good procedure fits a tool the team can use under time pressure.

What might an enquiry-handling exercise look like?

Suppose, hypothetically, a company receives service enquiries that AI classifies and routes to specialists. For a time the classification service does not respond, while the intake system still works. The process owner starts manual assignment for urgent cases only and leaves the rest in a visible waiting queue.

The exercise uses fictional tickets: urgent, ordinary and one whose priority cannot be determined without a conversation. The operator has short instructions and contact details for the person who decides exceptions. Acceptance checks whether the operator can assign an owner and mark an uncertain case without guessing. It does not measure clicking speed alone.

One enquiry was classified just before the outage, but confirmation did not reach the panel. The operator checks the target system and finds the assignment, avoiding a duplicate task. This shows why the exercise must include a partial state. A test that only switches the service off completely would not reveal the same difficulty.

After the exercise, manual takeover works, but one person does not know the urgency rules. The team improves the instructions and repeats the test with the next shift. The result is local evidence for this procedure. It does not justify announcing that the whole company is resilient to every technical outage.

How do you return to automation after repair?

Start with a small control task using synthetic data. Check connection, reading, writing and confirmation wherever the process requires them. A provider's availability notice is not a substitute. If configuration changed, restoration may require checking behaviour again on known cases.

Then agree what happens to the queue. Cases completed manually should not execute again. Cases already started need an owner's decision. Only new or explicitly handed-back tasks return to automation. Record the switch time and the person responsible for comparing case counts before and after resumption.

Watch load during the return. Accumulated work can trigger a service limit or delay current work. Decide the order for draining the queue and how to stop resumption. If you cannot safely distinguish backlogged from new tasks, keep the simpler mode until their states are organised.

Close the incident with a short review: what worked, what was missing and which assumption was wrong. Include staff time in the cost of the whole AI process (in Polish). This helps decide whether to improve instructions, reduce dependence or prepare another technical option.

When should the fallback not use a second model?

A second provider is not automatically a ready fallback. It may have different processing terms, integration scope and interpretations of instructions. If the company has not checked these in advance, spontaneously pasting material during an outage creates a new, unaccepted process. For a simple task, manual work in an approved system can be more predictable.

A second-model option is worth considering when the process warrants maintaining configuration, access and tests for both solutions. Record the minimum scope the substitute must actually take over. It need not provide identical convenience or every feature. It must meet agreed requirements for data, output quality and confirmation of target operations. For automatic fallback, assess routing between AI models (in Polish) so the team can establish which executor took the task and on what terms.

Check shared dependencies too. Both models may use the same intermediary, account, disk or internet connection. For a solution described as local, separately test what works without internet (in Polish) before relying on it as a fallback. Changing a model's name then does not remove the common cause of interruption. The plan card should state which outage the fallback covers and which problems still require manual work or deferral.

Include the automation exception queue (in Polish) in the continuity plan. A case needs an owner, deadline and closure method after manual takeover as well. Thoughtful change communication (in Polish) also helps: a clear status, instructions for continued work and the time of the next update.

What should be checked before approving the plan?

The plan should pass an exercise with someone who did not build the automation. If they need the author's constant help, the procedure is not yet self-sufficient. Fix the places where information was missing and repeat the relevant part. You do not have to simulate the whole company to uncover a significant gap in one process.

  • The process owner approved the minimum outcome and tolerable delay.
  • Outage symptoms and the person activating fallback are defined.
  • The last confirmed state of each important operation can be established.
  • Manual takeover prevents parallel execution of the same case.
  • The queue works without the dependency whose failure it must handle.
  • The exercise includes a partial state and missing confirmation.
  • Return requires a functional check and reconciliation of the queue.

To exercise the unavailability of one important process, see the AI audit offering.

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