Price Reversibility Before You Scale AI

The Bank for International Settlements describes an AI supply chain with five linked layers: hardware, cloud infrastructure, training data, foundation models and applications. High fixed costs, economies of scale and network effects shape several layers, while large technology firms increasingly operate across more than one. This is a market-structure analysis, not a forecast that every buyer will be locked in. It shows why a dependency recorded only at the application contract misses important upstream exposure. ¹

The Financial Stability Board's 2025 monitoring report identifies third-party dependencies and concentration among the vulnerabilities authorities need to track. Its financial-sector case study focuses on criticality, concentration and substitutability and also highlights data gaps. The finding is sectoral and supervisory; it does not measure the failure probability of an enterprise AI service. ²

A 2025 US Federal Trade Commission staff study of three large cloud-provider and AI-developer partnerships found contractual and technical terms that may increase switching costs, restrict multi-cloud use or link access to compute, models and sensitive business information. The study concerns specific partnerships and competition questions, not the terms offered to every enterprise customer. ³

These layers can reinforce one another. The same supplier group may influence the model, infrastructure, integration path and evaluation tooling. A buyer can therefore have several contracts yet still rely on one operating chain. Conversely, one contract can sit on components that are genuinely replaceable. The board needs a dependency map built around the business service and its upstream critical path, not a count of vendors.

The European Union's Data Act has applied since 12 September 2025. Its cloud-switching provisions require providers of data-processing services to remove specified contractual, commercial, technical and organisational obstacles, set contractual switching terms and support the portability of exportable data and digital assets. The rules include scope, technical and trade-secret limits; they do not guarantee that a buyer can reproduce every proprietary service or business outcome elsewhere. ⁴

This distinction is material for AI. A portable file may not include the provider's model behaviour, orchestration, safety settings, evaluation history, hidden dependencies or operational expertise. Even the same model and prompt can vary when the numerical execution path changes. A September 2026 preprint demonstrated cross-architecture output differences and a proposed deterministic kernel approach across three NVIDIA GPU families. That bounded technical result is not enterprise switching evidence, but it is a warning against treating identical inputs as proof of identical service. ¹³

Cloud research has described portability and interoperability problems for more than a decade. A peer-reviewed ACM review of more than 120 papers found continuing challenges in security, monitoring, identity, standards and vendor lock-in across connected clouds. Because the review predates current generative AI services, it supports the persistence of the systems problem rather than any claim about today's model market. ¹⁴

The practical unit of recovery is the accepted business service. The substitute does not need to reproduce every response. It needs to complete the bounded job at an agreed quality, cost and risk, with the required records and authority. That is why the enterprise needs its own acceptance pack: representative cases, thresholds, prohibited failures, human review, affected populations and evidence of which commitments can continue.

First, recoverable service. Define a minimum safe workflow that can operate through a supplier or model change. It may be slower, partly manual or narrower than normal. What matters is that the organisation knows which customer, safety, financial or legal commitments continue and which are paused.

Second, recoverable state. Keep authoritative source data, current instructions, identity and permission mappings, decision records and required history in forms the enterprise can retrieve and govern. Do not assume that an exported embedding store, conversation log or application database recreates the meaning, provenance or access controls of the original service.

Third, recoverable acceptance evidence. Retain the task set, baseline, thresholds, failure cases and sign-off record used to approve the service. A replacement can then be judged against the business promise rather than against the incumbent's interface. NIST's voluntary AI Risk Management Framework calls for inventory, third-party contingency, recovery, change management and safe decommissioning; it is guidance, not proof that an organisation has implemented them. ⁸

Fourth, recoverable operating knowledge. Preserve workflow logic, integration documentation, prompts where they are material, evaluation decisions, incident lessons, workarounds and the expertise needed to run the fallback. Singapore's voluntary 2026 framework for agentic AI emphasises bounded authority, meaningful human accountability, lifecycle controls and user capability. It applies to organisations using in-house or third-party agents but does not certify any deployment. ⁹

Fifth, recoverable authority. Name who can invoke the fallback, suspend commitments, accept reduced service, communicate the change and approve the new operating boundary. A recovery plan that requires fresh committee design during an incident is not an exercised decision right.

These assets do not all have to be owned or operated internally. Shared infrastructure, specialist evaluators, managed continuity services and sector arrangements can be stronger than a thin in-house substitute. The enterprise must be able to inspect the evidence and exercise the consequential decision.

A conventional business case prices licence, implementation and expected benefit. It often leaves the cost of changing direction outside the decision. The Institute proposes a reversibility account beside the investment case. This is not an accounting treatment or a prediction of supplier failure. It makes the accepted dependency and the cost of preserving an option visible.

Record five exposures for the consequential workflow: time to a safe minimum service; expected performance or quality loss during the fallback; data, rights and records that could become unavailable; internal knowledge and staff capacity required to recover; and the financial, customer or public commitments that would be delayed or breached. State the evidence behind each estimate and the owner who accepts it.

Then record the option cost: contract and transfer rights, maintained exports, independent tests, staff training, substitute capacity and rehearsal time. Some options will not be worth buying. A low-consequence internal drafting tool may justify a simple export and a restart plan. A service that commits money, affects employment, controls critical infrastructure or determines customer access may warrant a live reduced-service path and frequent exercises.

Recent US public-sector evidence shows why the decision should stay proportional. A 2026 Government Accountability Office report documents multi-vendor procurement challenges. In its response, the Department of Defense argued that provider-specific capabilities can be strategically correct when dependencies are understood and mission-critical workloads have a documented exit strategy; it also warned that multi-vendor SaaS for one function can be infeasible and expensive. The report and response concern US federal cloud procurement, but the counterargument is valuable: reversibility should manage concentration risk, not erase differentiation. ⁶

A separate 2025 GAO review found restrictive licensing and lock-in effects among selected agencies, including added fees, repurchase requirements, limits on architecture choice and one case in which a contractor required payment to regain data ownership. The sample is not representative of enterprise prevalence and the agencies could not always quantify the cost. It demonstrates the need to surface the actual rights and cost path before commitment. ⁷

A second contract is not evidence of continuity. Providers can share infrastructure, identity systems, data pipelines, model families or scarce staff. A substitute can also connect successfully while failing the business acceptance test. The exercise should therefore begin with the workflow and its commitments, not with a technology failover demonstration.

Select one consequential but containable service. Freeze a representative task set and the current acceptance thresholds. Trigger a realistic change: remove the preferred model, revoke one integration, withhold a proprietary context layer, apply a new data-location constraint or assume the supplier is unavailable. Restore the minimum safe service using the documented fallback and current people.

Measure elapsed time, accepted-task completion, quality loss, manual workload, unavailable records, security and rights exceptions, customer impact and the decisions that lacked an owner. Preserve every repair. The result is not a pass/fail badge; it is a priced gap between the approved loss boundary and the present operating capability.

United Kingdom financial regulators began direct oversight of designated critical technology providers in July 2026. Their regime complements rather than replaces regulated firms' responsibility for due diligence, risk management and contingency planning. The scope is a particular financial-sector regime, but the management principle is wider: supplier supervision does not operate the buyer's fallback. ⁵

US Treasury's financial-sector cloud review likewise found that cloud can improve security and resilience while firms still face concentration, incident and migration challenges. It is older sector evidence and should not be used to claim a universal failure rate. It supports an exercise that tests both the benefits of the managed service and the practical limits of substituting it. ¹⁵

The decision to set an acceptable reversal loss is portable. The recovery design is not. Law determines which records, rights and notices must survive. Language determines whether a substitute can meet the same population. Infrastructure and supplier availability determine whether parallel capacity is realistic. Contract enforceability and institutional capability determine which rights can be exercised under pressure.

The World Bank's 2025 global report describes large differences in connectivity, compute, locally relevant context and skills. High-income markets dominate data-centre capacity, while lower-income settings can face much thinner access and capability. The report concerns national foundations, not enterprise continuity. Its implication is that a company cannot copy a high-resource fallback design into every market and call the risk equivalent. ¹⁰

ECLAC's 2025 index compares enabling factors, research and adoption, and governance across 19 Latin American and Caribbean countries and finds pronounced variation. The African Union's 2024 continental strategy sets shared ambitions while recognising the need for national implementation, infrastructure, data and skills. Neither source supplies firm-level exit outcomes. Together they support a simple constraint: regional labels do not reveal the substitute, expertise or rights available to one service in one country. ^{11 12}

Where alternatives are scarce, the responsible answer may be a smaller service boundary, a manual minimum service, regional shared capability, a longer recovery objective or a consciously accepted dependency. A formal multi-provider design is not the global default. The board should receive the exception and its consequence, not a false promise of identical resilience everywhere.

This paper does not argue for technology neutrality. Provider-specific models, tools and managed services can create better performance, lower operating burden, faster delivery and stronger security than a portable design. The BIS supply-chain paper and US Treasury review both describe benefits alongside concentration and resilience concerns. ^{1 15}

Multi-provider designs can add identity, data-consistency, monitoring, skills and incident complexity. The 2026 GAO record includes a direct warning that multi-vendor SaaS for one business function may be infeasible and expensive. Duplicating two providers can also create a false sense of independence if both depend on the same upstream infrastructure or scarce expertise. ⁶

The proportional answer is to preserve the option that matches the consequence. For low-consequence or rapidly changing work, documented reconstruction may be enough. For a high-consequence service, the enterprise may need current data exports, a maintained acceptance pack, retained skills and an exercised reduced-service path. For a uniquely valuable capability, the board may approve a larger reversal loss—provided the exposure, benefit and decision owner are explicit.

Evidence remains limited. There is no representative global dataset measuring enterprise AI exit cost, successful model substitution, quality loss or recovery time. Official frameworks define expectations; they do not prove implementation. Technical research can isolate reproducibility and portability problems without measuring business continuity. The recommendations are therefore a decision framework to test, not an empirical claim that one architecture will minimise cost.

Name one accountable executive for each consequential AI-enabled service—normally the COO or the business executive who owns the outcome. The board or an authorised committee sets the acceptable reversal loss and approves material exceptions. The CIO or CTO owns the dependency map and recoverable technology assets. The chief risk officer or independent assurance leader challenges the acceptance evidence. Procurement and legal establish transfer, access, assistance and termination rights. The workforce owner ensures the minimum service still has competent people.

In days 1–30, select the three most consequential AI-enabled services. Map the five dependency layers and five recoverable assets. Set the acceptable interruption, minimum safe service, quality floor, required records and rights, critical knowledge and commitments that can be paused. Record every assumption that lacks evidence.

In days 31–60, exercise one service. Remove a material dependency and restore the minimum safe workflow using current people, current exports and current decision rights. Have an independent observer record time, accepted outcomes, manual work, missing assets, security or rights exceptions and customer consequences.

In days 61–90, put the reversibility account beside the benefit case. Decide whether to buy more option capacity, redesign the service, narrow its authority, renegotiate rights or accept the dependency. Release deployment capital only when the initial loss boundary and assets are documented. Release scale capital only when the exercised result sits within the boundary or the board has approved the exception.

Review the boundary when a model, provider, contract, integration, law, language population or critical upstream service changes. The board pack should show the benefit, the option cost, the exercised reversal loss, the accountable owner and every accepted exception.