AI Vendor Lock-In: Risks for Firms and SMEs in Italy

What really happens when the API goes down

The first question a professional firm should ask is not 'which AI model is the most accurate?' but 'what happens to my processes if this service is unavailable tomorrow morning?'. The APIs of major public providers — OpenAI, Anthropic, Google — offer high availability, but not absolute availability. Partial outages, traffic throttling and unscheduled maintenance do occur. And when your email-routing workflow, your draft-generation process or your document-analysis pipeline is built on a single external API, every interruption becomes an operational emergency.

But the risk is not purely technical. Providers can revise their pricing and terms of service with little notice: price increases, changes to usage limits, deprecation of specific models. If you have built automations on a particular model, the cost of migrating can be significant — in time, in development effort, in staff retraining. Dependency on a single external provider is technical debt that accumulates silently and is repaid all at once, invariably at the worst possible moment.

GDPR and professional secrecy: the concrete risks of public LLMs

When a professional firm sends a document containing personal or confidential data to a public LLM (Large Language Model) — an AI model accessible via API or web interface — it opens up a range of regulatory problems that many practitioners underestimate. The GDPR requires that personal data be processed only by parties that have been formally appointed as data processors, with precise contractual guarantees. Most public AI providers offer these guarantees only in their paid enterprise tiers, not on standard plans.

For accountants, lawyers and employment consultants, the problem is twofold: in addition to the GDPR, professional secrecy comes into play. Client data is not merely personal data — it is information covered by professional ethical obligations that practitioners cannot transfer to third parties without explicit consent and adequate confidentiality guarantees. A public LLM that processes an employee's payslips for a client company, or the documents in a legal dispute, could potentially be used to train future models — unless the contract explicitly excludes this.

The European AI Act adds a further layer: AI systems used in high-risk contexts require technical documentation, conformity assessments and, in certain cases, registration. Anyone using an AI tool in an unstructured way — without a risk analysis — is operating in a regulatory grey area that could prove costly. The article on Italian Law 132/2025 published elsewhere on this blog examines in depth the obligation to inform clients about the use of AI within your firm.

Is your firm exposed? Check these points

• ✕
  DPA signed with the AI providerDo you have a valid Data Processing Agreement with every AI provider that handles your clients' data?
• ✕
  Data hosted in the EUDoes processing take place on servers located within the European Union?
• ✕
  No use for trainingDoes the contract explicitly exclude your data from being used to train models?
• ✕
  Client disclosure up to dateAre your clients aware that you use AI tools to process their data?
• ✕
  Documented emergency planIs there a written procedure for managing critical processes if the AI provider is unavailable?

Geopolitical risk: what if Washington pulls the plug?

The question of European digital sovereignty is not an abstract concern for policy-makers — it is a concrete operational variable for anyone running a professional firm or an SME. The leading AI providers — OpenAI, Anthropic, Google — are US companies subject to American jurisdiction. This means that, in the event of geopolitical tensions, trade sanctions or regulatory decisions, access to their services from Europe could be restricted or cut off entirely.

These are not remote hypotheticals. In recent years we have seen how US–EU dynamics can rapidly affect access to digital technologies: the debate over data transfers, the discussions around the American CLOUD Act, and the tensions over TikTok and other digital services all demonstrate that geopolitics reaches into the server room. As AI4Business notes, the answer to this risk is private AI: artificial intelligence systems that run on infrastructure controlled by the organisation itself or by trusted European providers, rather than on third-party APIs from outside the EU.

For a professional firm or a mid-sized SME, this does not necessarily mean installing physical servers on-premises. It means choosing providers with EU data centres, open-weight models hosted on a European cloud, and architectures that do not depend on a single external access point. The choice is not between 'everything public' and 'everything private' — it is between maximum exposure and managed risk.

How to reduce dependency: practical steps for firms and SMEs

Reducing dependency on a single AI provider does not require doing everything at once. It requires a progressive strategy that starts with an analysis of current risks and builds resilience step by step. The starting point is not choosing a new tool — it is mapping the processes that already rely on AI and assessing their criticality.

For every AI-dependent process, three questions must be answered: how long can it be down before causing serious impact? What would happen if the data flowing through it were exposed? Is there a manual or technological alternative already available? The answers define the priority of interventions. High-criticality processes — those handling confidential data or sitting on the critical path of daily operations — must be protected first.

How to assess whether an AI provider is reliable in the long term

Evaluating an AI provider for a professional firm cannot be limited to accuracy benchmarks or a product demo. The criteria that truly matter for long-term sustainability and regulatory compliance are different. The first is contractual solidity: a reliable provider clearly sets out its registered office, the applicable jurisdiction, the conditions under which service terms and pricing may change, and the data-portability guarantees upon termination.

The second criterion is transparency over data management: where data is processed, who has access to it, how long it is retained, and whether it is used to train future models. The third is financial stability and service continuity: AI start-ups with only a few months of history and no contractual commitment to continuity are a genuine risk for anyone building critical processes on top of them. As the debate on security as governance makes clear, resilience is not a technical feature but an organisational choice that is built over time.

One final element that is often overlooked: portability. If you had to switch providers tomorrow, what would it cost you in terms of data migration, automation rewrites and retraining? Providers that make it difficult to leave exercise a form of lock-in that is as risky as any technical vulnerability. Give preference to providers that use open standards and document their APIs in a way that allows substitution without overhauling your entire architecture. The article AI Agents: Risks and Costs examines the topic of AI system governance in professional firms in greater depth.

Criteria for evaluating an AI provider for the long term

• ✓
  Clear registered office and jurisdictionDoes the contract specify the applicable jurisdiction and the provider's registered office?
• ✓
  Data Processing Agreement availableIs the DPA accessible without requiring an enterprise plan?
• ✓
  Data in the EU and no-training guaranteedIs data processed in the EU and not used to train models?
• ✓
  Contractual notice for changesDoes the contract provide adequate notice (at least 30–90 days) for price or service changes?
• ✓
  Data portability guaranteedCan you export all your data in a standard format upon termination?
• ✓
  Open API standardsDo the APIs follow standards compatible with other providers, reducing migration costs?

AI vendor dependency: the technical debt nobody wants to pay

There is a paradox at the heart of this issue: the SMEs and professional firms that adopted AI fastest are often the most exposed. They integrated powerful tools into their operational workflows, but without laying the foundations — solid contracts, documented alternatives, distributed architectures. The result is technical debt that grows every day the system works without incident, because every additional day of exclusive dependency raises the cost of exit.

The good news is that reducing this exposure does not require giving up AI altogether. It requires using it more maturely: with a strategy that separates critical processes from non-critical ones, that spreads dependency across multiple providers, that keeps confidential data on controlled infrastructure and that has a written, tested Plan B. This is precisely the kind of work that distinguishes those who use AI as a tool from those who are captive to it as a dependency.

For a professional firm, value does not lie in whichever AI model happens to be the most powerful at any given moment. It lies in having an AI infrastructure that works tomorrow, the day after and three years from now — regardless of what Washington decides, how much OpenAI changes its pricing, or which new European regulation comes into force. Strategy and security generate value. AI is the instrument.