AI Dependency on Hyperscalers Becomes a Governance Issue for Boards and Executive Management
What this is about
In conversations with board members, I frequently find that few people ask whether and how a strategic technology provider could be replaced without losing control over data, knowledge and operations. Marc Beierschoder and Antonio Russo from Deloitte Switzerland address precisely this question in a partner contribution in Netzwoche from 9 November 2026. Their headline makes the point: “If you cannot switch, you have no technology strategy”.
AI agents shift the question. The authors point out that embedded agents can create their own corporate memory: context, decisions, prompts, assessments, integrations and learning histories. The language model can be replaced relatively quickly. This memory may not be. If it sits on a provider’s platform, that is where the real dependency lies.
What concerns me more is the argument that you simply cannot switch any longer. Whoever argues this way has accepted a risk without ever deciding to do so. It is then merely managed on paper. The fact that a critical provider cannot be replaced overnight is obvious. This is precisely why it is the board’s job to know how long a switch would take and what its consequences would be. Only with this knowledge can it arrange a replacement operation or consciously bear the dependency.
Why this matters for boards and executive management
Art. 716a OR does not mention technology. However, it assigns to the board non-transferable and inalienable duties: overall management of the company, determination of the organisation and overall supervision of executive management. Art. 717 OR requires these duties to be fulfilled with all due care. My assessment: if the company’s operational capability depends on a single provider, this touches on overall management. The question therefore belongs on the board’s agenda.
Whoever cannot replace a critical component relinquishes bargaining power. The provider can adjust prices and licensing models or change interfaces. The customer follows along until it reaches its pain threshold. The authors also note that this requires no outage. The Broadcom takeover of VMware demonstrates this: according to a report from the observation office of the European cloud association CISPE, licensing costs for individual members rose by 800 to 1,500 per cent. Flexible models were reportedly replaced by three-year subscriptions. Equivalent alternatives were largely absent according to the report.
There is also the risk of a provider failing through an attack, a technical fault or government intervention. The latter has already happened. After US sanctions against the International Criminal Court, AP reported in May 2025 that Microsoft had suspended the email account of the chief prosecutor. Microsoft emphasises that it never suspended services for the Court itself. The case nevertheless shows how directly foreign law can have an impact on operations. Whoever knows such scenarios and fails to prepare is deliberately leaving a risk unmanaged.
What regulation and supervisory authorities already require
For banks, insurance companies and selected financial institutions, exit capability has been mandatory for years. The FINMA Circular 2018/3 requires that orderly return of an outsourced function or its transfer to another service provider is ensured. In the EU, DORA in Art. 28 para. 8 requires financial undertakings to have exit strategies for ICT services that support critical or important functions. Exit plans must be tested sufficiently and reviewed periodically. The EU Data Act has regulated cloud provider switching since 12 September 2025; from 12 January 2027, switching fees are prohibited.
For an industrial or trading company, FINMA Circular 2018/3 and DORA do not apply. But they demonstrate the standard that supervisory authorities and legislators apply to exit capability. I expect this standard to serve in future as a reference for careful corporate governance even outside the financial sector. For groups with companies in the EU, the Data Act is moreover a concrete lever in the next contract negotiation.
What is different about AI agents
Antonio Russo formulates in the interview a principle I share: “The AI may access the corporate memory, but it should not own it.”
Open standards help with interfaces. The Model Context Protocol governs how AI applications access tools and data. The Agent2Agent Protocol governs communication between agents. Both now rest with the Linux Foundation and thus with no single vendor. For accumulated agent knowledge, the situation is less clear; Russo notes that portability is not yet fully resolved. This makes it all the more important to keep knowledge bases, decisions, assessments and prompts in your own structures. Equally, the contract should specify in what format the provider will release them on exit.
Consciously deciding on dependency
Not every dependency is a mistake. Access to best-of-breed technology or lower costs can justify it; the authors hold this view too. What matters is that the board enters into it consciously, with known exit costs and a documented resolution. An exit strategy that has never been tested remains an assumption. I agree with the authors on this point.
Roles must be clearly separated. Executive management develops exit strategy and migration plans and tests them. The board determines which dependencies the company is willing to bear. It demands evidence of how dependent the company actually is on individual AI and cloud providers and what an exit would cost. The authors propose “Time to Substitute” as a metric: the time until a replacement operates stably in real operations.
My suggestion: the board asks for this metric for the most critical providers once a year, together with the results of at least one switching test.
Questions for executive management
- Which providers and technologies would materially affect our operations in the event of failure or switching, and how long does it take in each case for a replacement to run stably?
- Do we have migration plans for these providers, do we know their costs and consequences, and have we tested a switch for at least one critical system?
- Do context, decisions, assessments and learning histories of our AI agents lie in structures we ourselves control, or are they tied to a provider’s platform?
- Do our contracts govern release of data in usable formats, a transition period with continuing service, and the provider’s cooperation in the switch?
- Which dependencies do we consciously bear, and where has the board decided on this?
Sources
Beierschoder/Russo, Netzwoche, 9 November 2026 (Deloitte partner post) FINMA Circular 2018/3 Outsourcing, version 4 November 2020 DORA Art. 28, Springlex (text OJ L 333) McCann FitzGerald, EU Data Act: Switching Cloud Provider, 17 September 2025 Deloitte Legal, Cloud Switching under EU Data Act IT Reseller, 27 May 2025 ECCO, Second Report on Broadcom, May 2025 Techzine, 4 June 2025 Anthropic, Transfer of MCP to the Linux Foundation, 9 December 2025 A2A Protocol