SIAgentNews.com
← All news

The risk of AI agents performing duplicate actions after cloud outages

Organizations must maintain independent override authority to prevent autonomous agents from repeating high-impact tasks like payments once connectivity returns.

· First reported by Eurasia Review

24 hoursDuration of the Microsoft 365 outage

The short version

  • The November 12–13 Microsoft 365 outage left organizations to manually verify which transactions were completed.
  • AI agents with independent identities may automatically retry tasks the moment a cloud connection is restored.
  • A hospital agent might send a second order for supplies if it doesn't receive a confirmation during a network drop.
  • Experts suggest separating an agent's ability to prepare work from its authority to execute high-impact actions.

Why it matters

As companies move toward autonomous AI agents for procurement and payments, they need a way to reconcile records after technical failures. Without human-in-the-loop verification, a simple system reboot could trigger multiple unauthorized financial transactions.

Lessons from a global outage

A major Microsoft 365 disruption on November 12 and 13 lasted about 24 hours, affecting services like Exchange Online. While the provider eventually restored the cloud, the outage demonstrated that a healthy dashboard does not mean all data transactions were successfully completed. Organizations were left to figure out what work remained unfinished and what was already done.

The city council in Hamilton, New Zealand, serves as a primary example of this governance challenge. While the tech provider began fixing the system, the local government kept its online portals shut over the weekend to perform manual audits. They only brought the digital services back online one by one after confirming the status of all payments and applications.

This delay highlights a critical gap in modern infrastructure. A technology provider can report that a system is back online, but only the specific organization using that system can establish which permissions should return. As companies begin to entrust AI agents with payments and records, this gap becomes a significant governance problem.

The risk of the 'blind retry'

The arrival of AI agents with independent identities, such as Microsoft's Copilot Autopilot, makes the recovery question more urgent. These agents are designed to follow up on recurring work and use tools to complete tasks. However, if a connection drops mid-action, the agent may not know if the external world received the request.

Imagine a scenario where a medical facility uses an automated purchasing tool to request supplies right as the cloud connection fails. If the vendor receives the request but the confirmation fails to reach the software, the agent's internal record will show the task as incomplete. Once the network is restored, the software might automatically try to place the order again, resulting in a second delivery of the same items.

This scenario illustrates that an unsuccessful request is not always an unsuccessful transaction. While engineering solutions like idempotent tool-action design can help, they still require specific transaction identifiers and current authorization checks. A cached approval from before the outage should not be treated as an eternal permission slip for the agent to act.

Who is affected by these errors

The consequences of these technical failures vary depending on the industry and the specific services used. In Brazil, the National Council of Justice had to suspend procedural deadlines because lawyers could not receive the email codes required for digital filing. This shows how a lack of reliable connectivity can halt legal proceedings across an entire country.

In England, the NHS Supply Chain saw a slowdown in incoming orders and the forms used for home-delivery patient referrals. Although the physical warehouses and transport fleets remained active, the lack of a functional portal forced staff to use phone calls to contact customers. These examples show that different countries face different business outcomes even when sharing the same cloud dependency.

For businesses, the danger lies in the first few minutes after access returns. If an agent is allowed to resume work immediately without a reconciliation step, it could create duplicate purchases or irreversible actions. Organizations need a way to preserve staff notes and check whether the original spending approval remains valid before letting the software proceed.

Designing for safe recovery

To mitigate these risks, experts argue that the authority to pause an agent must remain reachable even when the cloud is down. If the same unavailable identity service controls both the agent and its override switch, the backup system is effectively useless. A sensible design separates the work an agent prepares from the high-impact actions it is allowed to execute.

Authorization for high-impact actions should be enforced outside of the agent's own logic. This ensures that designated staff can restrict the agent's power while it continues to read records or prepare drafts. Such principles were discussed in an October exercise by Singapore’s Cyber Security Agency and Google, which focused on external safeguards and independent verification.

Organizations should not assume that every office task requires a human to sign off on every single step. Instead, they should focus on the first consequential transaction after an outage. The goal is to ensure the customer can account for the work and safely release authority rather than simply trusting a green light on a dashboard.

Testing the recovery process

A practical way to test these systems does not require creating a new regulatory committee. Companies can ask their vendors to simulate a scenario where a supplier accepts an order but fails to send a confirmation back to the system. This allows the organization to see how the software handles the missing information during a period of no connectivity.

During this test, an employee should verify the order using an approved alternative channel, such as a phone call. Once the connection is restored, the organization can then observe how the agent behaves. The test should determine if the next shift can avoid a duplicate and complete exactly one authorized purchase.

Ultimately, institutions adopting AI agents should retain the right to decide when it is safe to reopen services. Like the city council in Hamilton, organizations must be able to verify the state of the world before letting autonomous systems resume high-impact work. This human oversight is the only way to ensure that a system recovery does not lead to a series of automated errors.

Next steps for organizations

Companies should begin by identifying which high-impact actions their AI agents are currently authorized to perform. They should then create a manual reconciliation process for these specific actions in the event of a service disruption. This process should include a way to verify transaction IDs with external suppliers or partners.

It is also important to ensure that the 'kill switch' for an agent is not dependent on the same cloud service that the agent uses for its primary tasks. If the cloud goes down, the ability to stop the agent must remain functional. This prevents a scenario where an agent continues to attempt actions in a loop because its own control interface is also offline.

Finally, organizations should conduct regular drills to practice these recovery steps. By simulating a connection drop, teams can learn how to coordinate with suppliers and verify records. This preparation ensures that when the cloud eventually returns, the transition back to automated work is controlled, deliberate, and safe.

Try it today

Identify one high-impact action your team currently automates (like a payment or order) and write down the manual steps you would take to verify that action if the internet went out for four hours.

What to watch

  • How vendors will integrate 'idempotent' design to prevent duplicate transactions.
  • Whether major cloud providers will offer better 'state recovery' tools for AI agents.

Try it

Identify one high-impact action your team currently automates (like a payment or order) and write down the manual steps you would take to verify that action if the internet went out for four hours.

Sources and links

Read the original at Eurasia Review

First reported by Eurasia Review. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news