The day the AI tool stops: an SME's fallback plan

A question nobody asks when buying

When an SME adopts an artificial intelligence (AI) tool, it looks at what the tool does, what it costs and where the data goes. It rarely asks what it will do on the day the tool stops. The task it handed over, answering customers, preparing quotes or sorting invoices, does not stop with it: someone will have to take it over, with or without what the tool was keeping.

This guide goes through five ways an AI tool stops, each with a real case, then what an SME prepares for every system it uses. The sources were read on 24 October 2026; they are gathered at the end of the page.

First way: the service is shut down

A vendor can withdraw a service, with or without a replacement. On 24 March 2026, OpenAI told developers that its Videos API and its Sora 2 video generation models would be removed from the API on 24 September 2026. On the page that lists these shutdowns, the column for the recommended replacement is empty. On 26 August 2025, the same vendor had announced that its Assistants API would close on 26 August 2026, a year later, this time with a named replacement that still had to be programmed.

An agency producing videos through that API, or an SME whose internal assistant ran on the Assistants API, therefore knew its end date and had work to do before it. An outage of a few hours is the short form of the same problem: it is prepared the same way, with a tighter deadline.

Second way: the price changes

The tool stays available but costs more. On 4 December 2025, Microsoft announced a global price increase for its Microsoft 365 business suites, effective 1 July 2026 and accompanied by added features, including enhancements to its Copilot Chat assistant. Business Standard goes from $12.50 to $14.00 per user per month, Business Basic from $6.00 to $7.00: these are US list prices, which Microsoft says vary by country and currency. Existing customers keep the old price until their renewal.

Nobody guesses price increases in advance. What an SME can know is the price at which it would switch tools, and how much time the switch would take.

Third way: the terms change

The price stays the same, but what the vendor does with your data changes. On 28 August 2025, Anthropic changed the terms of its Claude assistant for consumer accounts (Free, Pro and Max). Existing users had until 8 October 2025 to choose whether their conversations could be used to train the models; without that choice, they could no longer use the service. Retention went up to five years for those who agreed, against thirty days for the others. Offers under commercial terms (Team, Enterprise and the API) were not affected.

An employee working for the company on a personal subscription therefore falls under the consumer terms, not the business ones. The fallback plan starts with a simple question: which contract covers each tool, and who receives the vendor's announcements when it changes something?

Fourth way: the vendor leaves Switzerland

An SME sometimes chooses a tool because its data stays in Switzerland. That reason can disappear without the tool changing. On 23 July 2025, when launching its AI assistant Lumo, the Geneva-based company Proton announced that it was moving most of its physical infrastructure out of Switzerland, citing the legal uncertainty around Swiss surveillance proposals, and that Lumo would be the first product to move. The proposal in question, the revision of the Ordinance on the Surveillance of Post and Telecommunications (SPTO), has since gone back to the drawing board: on 12 February 2026, the Federal Council announced an external regulatory impact assessment, followed by a new consultation.

The register should therefore say where each tool hosts the data today, and what the company will do if that place changes: accept it, export its data and leave, or switch tools.

Fifth way: the law prohibits the use

The tool can keep working while the way it is used becomes unlawful. Since 2 February 2025, the AI Act, the European regulation on AI (Regulation (EU) 2024/1689), prohibits among other things the use of AI systems to infer emotions of a natural person in the workplace, except for medical or safety reasons (Art. 5(1)(f)). It reaches a Swiss SME whose tool produces output used in the European Union (Art. 2). The list of prohibited practices grows on 2 December 2026, and the consultation on a preliminary draft Swiss AI law is planned for spring 2027: our AI Act module (/aiact) keeps these deadlines up to date.

An SME that had the tone of its employees on the phone analysed to rate their mood would not have to wait for an outage: overnight, the task would have to be done differently, or not at all. In Switzerland, Art. 26 of Ordinance 3 to the Employment Act (OLW 3) already prohibits systems intended to monitor the behaviour of employees at their workplace.

What an SME prepares for each system

For each AI tool a task depends on, five lines are enough. The task it carries, and what happens if it is not done for a week. Who takes it over without the tool, with what (document template, spreadsheet, telephone) and how quickly. What the company gets back: its data, but also its settings, saved instructions and histories, with their format. The replacement tool or method. And finally, the date of the last trial.

Exports are checked for real. You look for the export function in the tool's account, settings or administration, make a real export and note its format (CSV, JSON, PDF or none), then check whether another tool can open it. A tool without export is recorded as such: it is the first fallback to prepare.

The last trial is the line people skip most readily. A plan never tried describes what you believe you can do. Half a day a year is enough: you cut access to a critical tool, or act as if it were cut, and the person named does the task without it. You note how long it took and what was missing, correct the plan, and review it whenever prices, terms or the law change.

What the law says, and what it does not

No text requires an SME to have a fallback plan for each AI tool. It is first of all a good practice, business continuity applied to the tool, and only it covers a change of price or terms, a vendor leaving or a use newly prohibited by law.

For the outage of a tool that processes personal data, Swiss law covers part of it. Art. 8 of the Federal Act on Data Protection (FADP) requires a level of security of personal data appropriate to the risk. Its ordinance (DPO) requires that the data “are available when they are required” (Art. 2 let. b) and that “the availability of personal data and access to them can be rapidly restored in the event of a physical or technical incident” (Art. 3 para. 2 let. d). The outage of a service is such an incident; a change of price, terms or vendor is not. These texts were read on Fedlex on 24 October 2026, in the version in force.

AI-Karma's register and indicator

AI-Karma's free AI system register (/registre) carries, under each system, an “If the tool stops” block: task carried, takeover without the tool and lead time, recoverable data and settings with their format, replacement, date of the last trial. It prints with the rest of the register and exports to JSON and CSV. Everything stays in your browser: nothing is sent to AI-Karma.

In the KarmaScore, the indicator “Fallback plan for AI tools” in the technology dimension measures the same thing on five levels, from “no idea what would happen if an AI tool stopped” to a fallback tried at least once a year for each critical tool. The methodology (/methodologie) describes it with its legal references; the free self-assessment (/evaluation) measures it and proposes a five-step action plan.

AI-Karma applies the same grid to itself (/exemplarite). The written summary of its reports goes through Anthropic's API, under commercial terms; without it, the report without AI takes over automatically. It scores itself 2 out of 4 on this indicator, as no replacement has yet been named and tried.

What this plan does not promise

A fallback plan guarantees neither that no task will be interrupted nor that no job will change. It says who does what when the tool is missing, and what the company keeps in its own hands. This guide is not legal advice: for a use close to the AI Act's prohibitions, have your analysis checked by a specialist.

Sources, read on 24 October 2026

Service shut down: OpenAI, “Deprecations” page of the API documentation (developers.openai.com/api/docs/deprecations), entries “Sora 2 video generation models and Videos API” (announced on 24 March 2026, removed on 24 September 2026) and “Assistants API” (announced on 26 August 2025, removed on 26 August 2026).

Price: Microsoft, “Microsoft 365 Pricing and Packaging Updates”, page of 16 February 2026 (microsoft.com/en-us/licensing/news/2026-m365-packaging-pricing-updates).

Terms: Anthropic, “Updates to Consumer Terms and Privacy Policy”, 28 August 2025 (anthropic.com/news/updates-to-our-consumer-terms).

Vendor leaving: Proton, “Introducing Lumo, the AI where every conversation is confidential”, 23 July 2025 (proton.me/blog/lumo-ai); Netzwoche, “Bund überarbeitet revidierte Überwachungsverordnungen”, update of 12 February 2026 (netzwoche.ch).

Law: Regulation (EU) 2024/1689 (AI Act), Art. 2, Art. 5(1)(f) and Art. 113; FADP (SR 235.1), Art. 8; DPO (SR 235.11), Art. 2 let. b and Art. 3 para. 2 let. d; OLW 3 (SR 822.113), Art. 26.

Frequently asked questions