November 17, 20265 min read

    What Happens When Your AI Stops? How to Build an AI Business Continuity Plan

    By MASSIVUE Team

    AI Business ContinuityOperational ResilienceAI Risk ManagementAI GovernanceMAS Guidelines

    For CIOs, COOs, chief risk officers and the transformation leaders who now run critical work on AI.

    Contents
    1. The short answer
    2. Why this became a board question in 2026
    3. Three ways AI stops
    4. The number to calculate first
    5. The AI Continuity Test: four questions per service
    6. The manual fallback problem
    7. A worked example
    8. What to do this quarter
    9. Where MASSIVUE fits
    10. Frequently asked questions
    11. Sources

    The short answer

    An AI business continuity plan answers one question: if an AI system stops working, or you have to switch it off, how does the business keep serving customers?

    Most organisations already have a business continuity plan. What they usually lack is AI inside it. The AI estate grew one use case at a time, through internal builds and through features that arrived inside software the business already owned. Nobody added those dependencies to the continuity map, and the fallback written into most plans, "revert to manual processing", assumes a team and a skill set that may no longer exist.

    Start with one number: how many of your critical business services depend on the same AI provider? Then run four questions against each of those services. What stops? Who can stop it? What runs instead? When did we last prove it works?

    For financial institutions in Singapore this is no longer optional good practice. The Monetary Authority of Singapore's Guidelines on AI Risk Management, issued on 7 October 2026, expect tested contingency plans for high-risk AI. For everyone else, the outages of 2026 have made the same point less politely.

    Why this became a board question in 2026

    Three things changed this year.

    AI moved into critical work. Accenture's CIO Pulse, published on 3 November 2026 and based on 300 technology leaders at organisations with more than US$1 billion in revenue, found 64% of CIOs scaling enterprise AI deployment and 37% already using agents for end-to-end orchestration. Once AI is answering customers, checking documents or triaging alerts, its failure is a service failure, not an IT inconvenience.

    The outages became real and public. Thoughtworks described a major global disruption to Anthropic's Claude on 2 June 2026, with enterprise systems that relied on its API meeting a wall of errors. A Cloud Security Alliance research note recorded significant ChatGPT incidents on 20 April and 23 July 2026. The point is not any one provider. Any external service has outages, and every enterprise that builds on one inherits them.

    Regulators wrote it down. The MAS Guidelines on AI Risk Management list "over-reliance on a few dominant Generative AI providers for mission critical areas" and "potential service disruptions due to Generative AI outages" among the risks AI introduces. For a use case assessed as high risk, they expect contingency plans with fallback options, such as alternative systems or manual processes, reviewed and tested regularly. Where AI has kill switches, they expect activation protocols that are also tested. In the United States, the Office of the Comptroller of the Currency (OCC) issued an examiner framework for generative and agentic AI on 4 November 2026. It is guidance for examiners rather than new rules for banks, but it treats operational resilience as covering "both loss of service and incorrect operation", and names kill switches and circuit breakers as containment tools.

    Confidence has not caught up. Accenture found 53% of CIOs confident in their ecosystem's cyber resilience, but only 17% with real-time visibility and automated resilience across their technology partners. Wolters Kluwer's US Banking AI Risk and Governance Index, based on 230 US banking professionals and published in May 2026, found 72% naming either AI kill-switch protocols or regulatory reporting of AI failures as the area where their bank was least prepared.

    Three ways AI stops

    Continuity plans are usually written for one kind of event: the system is down. AI can stop in three different ways, and each needs a different response.

    How it stops What it looks like Why it is different
    It goes down The provider has an outage, an application programming interface (API) returns errors, a quota runs out The classic case. A second provider may solve it, if one is ready
    You switch it off It starts producing wrong, harmful or out-of-policy outputs, or an agent acts outside its limits The system still works, which is the problem. Rerouting to a similar model may repeat the fault
    It changes or disappears A model version is retired, a vendor changes the model inside its product, a contract ends, a vendor withdraws support Slow and quiet. Often discovered only when behaviour shifts

    The MAS Guidelines cover all three. For third-party AI they ask whether contingency plans exist for failures, for unexpected behaviour and for vendors discontinuing support. The OCC framework makes the same split between loss of service and incorrect operation. A plan that only covers outages covers one row of this table.

    The number to calculate first

    Before writing any plan, count how many of your critical business services depend on a single AI provider.

    This sounds simple. It rarely is, for one reason: much of an organisation's AI was never "deployed" by anyone. It arrived inside a customer relationship management (CRM) upgrade, a contact-centre platform, a document management system or a fraud tool, and runs on a foundation model the buyer may never have been told about.

    AI CONTINUITYSix critical services.One point of failure.Before you write an AI continuity plan, count how many critical services depend on one AI provider.Customer service chatKYC document checksFraud alert triagePayment exceptionsComplaints routingCredit memo draftingOne AI model providerIf it stops, all six stop.Reached through software you boughtIllustrative example. MAS Guidelines on AI Risk Management (7 October 2026) list reliance on a single provider as a possible risk appetite measure.MASSIVUE
    Illustrative. Two of the six dependencies arrive through vendor software, which is why they are usually missed.

    The MAS Guidelines point in exactly this direction. Among the examples of quantitative measures a board could set in its AI risk appetite, they list reliance on a single provider, alongside thresholds for AI incidents. That turns a vague worry about concentration into a number a board can track.

    To count it properly:

    1. Start from your critical business services, not from your AI inventory. The question is which services would stop, not which models you run.
    2. Trace each service to the model underneath it, including AI reached through software you bought. Ask each vendor directly which model provider their feature uses and what their product does when that provider fails.
    3. Group by provider. The result is often a short list, with one or two providers carrying most of the critical exposure.

    If the answer is "four of our six critical services run through one provider, and two of them we learned about this week", you have found the most useful fact in your continuity programme.

    The AI Continuity Test: four questions per service

    Run these four questions for every critical business service that depends on AI. Each one is simple to ask and uncomfortable to answer.

    1. What stops?

    Map what actually breaks when the AI is unavailable, end to end. Not the model, the service: which customers are affected, which obligations are missed, which downstream teams stall.

    For Singapore financial institutions, this is the work the MAS Business Continuity Management (BCM) Guidelines of June 2022 already expect for critical business services: identify end-to-end dependencies and set a Service Recovery Time Objective for each one. Those guidelines also expect an independent BCM audit every three years, so many institutions face their next one by June 2027. The practical step is to add AI dependencies to that existing map before then, not to build a separate AI continuity framework beside it. Two parallel frameworks tend to drift apart, and one of them goes stale.

    2. Who can stop it?

    A kill switch is a decision, not a button. Three things must be written down before an incident, not debated during one:

    • Who has the authority to switch off or restrict a specific AI system, by name or role, including out of hours.
    • What triggers it. Error rates, complaint spikes, a confirmed harmful output, a provider status alert. Pre-agreed thresholds stop the decision becoming a negotiation.
    • Who tells whom. Customers, business owners, and where required the regulator.

    Where the AI sits inside a vendor's product, the switch may not be yours at all. Find out now whether you can turn the AI feature off without turning the whole product off. Sometimes the answer is no, and that changes which fallback you need.

    One more point: a kill switch without a fallback simply converts an AI incident into a service outage. Questions 2 and 3 have to be answered together.

    3. What runs instead?

    There are four fallback levels. Most services need more than one, because no single level covers every way AI stops.

    Level What it means Best for The catch
    1. Reroute Switch to a second model or provider that is already approved for this use Outages Must be tested and approved before you need it. Does not help if the fault is shared, or if you switched off because of behaviour
    2. Narrow Run a smaller, safer version: fixed answers, rules, an agent with fewer permissions, or a human approving every action Misbehaviour, partial failure Throughput drops. Customers notice the difference
    3. Hand back People do the work the AI was doing Short disruptions in lower-volume services Depends on people, skills and capacity that may no longer exist (see below)
    4. Pause and tell Stop the service deliberately, queue the work, tell customers when it will resume When no safe fallback exists Needs prepared customer messages and a recovery time you can actually meet

    The OCC framework describes the same idea in its own words: testing whether work can continue through "a narrower fallback agent, or another approved process". Level 4 is not a failure of planning. Choosing it deliberately, with customers informed, is far better than discovering it by accident.

    4. When did we last prove it?

    This is the question most plans fail, and it is where the manual fallback breaks.

    The manual fallback problem

    Almost every AI continuity plan ends the same way: if all else fails, people will do the work by hand. That sentence carries three assumptions, and AI quietly breaks all of them.

    The capacity is still there. Teams get resized around the AI. A service that once needed 110 people may now run with 45, because the AI handles most of the volume. The manual fallback assumes the 110.

    The skills are still there. People who did the work by hand move on, and people who joined after the AI arrived have never done it at all. A skill that is not used fades, and nobody notices until the day it is needed.

    The process is still there. The manual procedure in the plan often describes forms, systems and handoffs that were redesigned when the AI went in. People are asked to fall back to a workflow that no longer exists.

    A manual fallback is a promise made on behalf of people who may no longer be in the room. The only way to know whether it holds is to try it: real people, real volume, a timer running. The MAS Guidelines ask for contingency plans to be reviewed and tested regularly, and testing a manual fallback means testing the people, not the document.

    This is also why AI continuity is a workforce question as much as a technology question. Keeping a fallback alive means deciding, deliberately, which skills the organisation must keep current even though the AI does the work most days, and who practises them.

    Your provider's uptime commitment is not your continuity plan

    A service level agreement with an AI provider typically compensates you with service credits when availability falls short. Credits do not answer a customer, process a payment exception or clear a fraud alert. Read the agreement for what it is: a pricing term, not a recovery strategy. Your obligations to customers do not transfer to your vendor.

    A worked example

    The following is illustrative, not a benchmark. The numbers are constructed to show the method.

    A mid-sized Singapore bank runs six critical services on AI. Four were built internally on one model provider. Two arrived inside vendor platforms, for payment exceptions and complaints routing, and when the bank asked, both vendors confirmed they run on the same provider. The count is six out of six.

    Take the contact centre. It handles 12,000 contacts across a 12-hour day, about 1,000 an hour. The AI resolves 60% of them, 600 an hour. A team of 45 people handles the other 400 an hour and is fully loaded doing so.

    Now the provider has a four-hour outage.

    • 2,400 contacts arrive that the AI would have handled, and the 45 people have no spare capacity.
    • Hand back: 20 people are borrowed from other teams. They have not handled contacts in 18 months and work at about half pace, around 4.5 contacts an hour each. Over four hours that clears roughly 360.
    • Narrow: a fixed-answer menu with no generative model behind it, prepared in advance, deflects 30% of the AI's usual volume, 180 an hour. Over four hours that clears 720.
    • What is left: about 1,320 contacts. These go to pause and tell: a callback queue and a message that names a realistic time.

    Three lessons come out of this simple arithmetic. No single fallback level covered the gap. The manual fallback covered only 15% of the shortfall, because the people who could do the work no longer exist at the scale the plan assumed. And the bank would only learn any of this from a rehearsal, or from the outage itself.

    The fix is rarely one large investment. Here it might be a second, pre-approved provider for the contact centre (level 1), a maintained fixed-answer menu (level 2), a small rotating group who practise contact handling each quarter (level 3), and pre-written customer messages (level 4).

    What to do this quarter

    Five steps, in order.

    1. Count the single-provider dependencies. Start from critical business services. Include AI inside vendor software, and ask every vendor which model provider they use and what their product does when it fails.
    2. Put AI into the continuity map you already have. Add each AI dependency to the end-to-end dependency map and recovery objective for the service it supports. Do not build a second framework.
    3. Write down who can switch each high-risk AI off, on what trigger, and who tells customers and, where required, the regulator. Confirm you can actually disable vendor AI features.
    4. Choose fallback levels for each service and test the manual one for real. Real people, real volume, timed. Record the gap.
    5. Rehearse on a schedule and after material change. MAS asks for regular testing. Our view is that twice a year, and after any change of model, provider or process, is a sensible starting point. Keep the results as evidence.

    For MAS-regulated firms, the timing matters. The Guidelines take effect on 7 October 2027, with the lifecycle control expectations in Section 5, where the contingency planning sits, to be met by 7 October 2028. The single-provider count, however, belongs to the board's risk appetite work in Section 3, which applies from October 2027. That makes step 1 the first job, not the last.

    Where MASSIVUE fits

    MASSIVUE is an AI and enterprise transformation consultancy headquartered in Singapore. Our role in this problem is specific.

    Consult. We map AI dependencies against critical business services and set the decision rights for who can restrict or stop an AI system. That work sits inside the AI operating model, because a continuity plan is only as good as the authority behind it. For AI that arrives inside vendor products, our guide to assuring third-party AI under the MAS Guidelines covers contracting for exit and contingency.

    Upskill. A manual fallback depends on people who can still do the work. MASSIVUE Academy designs role-based capability programmes, which is how an organisation keeps fallback skills current instead of discovering they have faded.

    Sustain. Protum, MASSIVUE's AI governance platform, covers the controls that make a switch-off safe in practice, including runtime containment and re-certification when a model or provider changes.

    What we do not claim is a proprietary statistic on AI outages. The figures in this article belong to MAS, the OCC, Accenture, Wolters Kluwer, Thoughtworks and the Cloud Security Alliance, and each is listed below so you can check it.

    If you want to run the single-provider count with your own leadership team, Sandeep Joshi, MASSIVUE's Founder and Managing Director, takes these conversations directly. You can book a 30-minute conversation with him.

    Frequently asked questions

    What is an AI business continuity plan?

    It is the part of a business continuity plan that covers AI. It identifies which critical services depend on AI, who can switch an AI system off and on what trigger, what runs instead when it stops, and how recently that fallback was tested. It should sit inside your existing continuity framework, not beside it.

    Do the MAS AI Risk Management Guidelines require an AI continuity plan?

    For AI use cases assessed as high risk, the Guidelines expect contingency plans with fallback options, such as alternative systems or manual processes, reviewed and tested regularly, plus tested activation protocols where kill switches exist. Where poor performance or unavailability of an AI tool is unlikely to have a material adverse impact, basic governance may be enough. The Guidelines take effect on 7 October 2027, with the Section 5 lifecycle expectations to be met by 7 October 2028.

    Is having a second AI provider enough?

    No. A second provider helps with outages, if it is already tested and approved for the specific use. It does not help when you have switched AI off because of harmful or incorrect behaviour that a similar model could repeat, and it does not help when the AI sits inside a vendor product you cannot reroute. Most critical services need more than one fallback level.

    What is an AI kill switch?

    MAS describes kill switches as mechanisms used to deactivate AI quickly when it exceeds risk tolerances. In practice the switch matters less than the decision around it: who can use it, on what trigger, and what runs once the AI is off. A kill switch without a fallback turns an AI incident into a service outage.

    How often should we test AI fallbacks?

    MAS asks for regular testing without fixing a frequency. Our view is that twice a year, plus after any change of model, provider or process, is a reasonable starting point. Test with real people at realistic volume, and time it.

    Does this only matter for banks?

    No. Banks face the most explicit regulatory expectations right now, from MAS in Singapore and in the OCC's examiner framework in the United States. But any organisation that runs customer service, operations or compliance work on AI faces the same three failure modes, with or without a regulator asking.

    Sources

    This article reflects MASSIVUE's interpretation of published regulation and research. It is not legal or compliance advice. The worked example is illustrative and is not a benchmark. Regulatory instruments change: confirm the current position with the issuing authority before relying on it.

    Share this article

    Help others discover this insight