How much does a subscription to an external AI model really cost?
The most immediate cost of contracting access to an external AI model is the subscription or token-based consumption. It is a visible cost, easy to budget for and, in the early stages, relatively low. The problem begins when that model stops being an occasional tool and becomes an operational dependency.
The providers of the main models on the market have progressively moved towards billing structures based on actual usage. Microsoft integrated Copilot into its enterprise suite with staggered price rises from 2025, and the APIs of the most advanced models from OpenAI or Google follow a variable consumption logic that can scale significantly in production environments. For a company that started by paying a fixed monthly amount for an AI tool and has integrated it into its daily operations, the jump to usage-based billing can represent a budgetary change that had not been anticipated.
Added to this is a historical pattern worth bearing in mind, although it comes from enterprise software in general and not specifically from AI: a Capiche study of one hundred business applications (reported by The Next Web) found that 67 of them raised their price by an average of 98% between 2009 and 2019, and that software prices rose by 62% over that same decade, three times above average inflation. There is no structural reason to think AI will behave any differently.
What hidden costs come with depending on an external AI provider?
Beyond the bill, dependence on an external provider generates a layer of costs that few organisations calculate before committing.
The cost of outbound data
Moving massive context windows between cloud regions involves high egress fees, and the vector storage needed for retrieval-augmented generation (RAG) pipelines carries minimum monthly fees that accumulate even when the application is idle.
The uncoordinated proliferation of tools
When different teams buy AI tools independently, for writing, document analysis or direct integration with language models, the result is duplicated functions that raise the total cost of ownership. Each additional integration also creates a new exposure surface: according to Cyberhaven Labs' 2026 AI Adoption & Risk Report, 39.7% of interactions with AI tools already contain sensitive corporate data, compared with 34.8% in the 2025 report and 27.4% in the 2024 report. The trend is clear and goes in only one direction.
The cost of migration
When data and processes are deeply integrated with the provider's platform, migrating requires redoing much of the infrastructure and code. The licensing, consultancy and training costs needed for that change act as an exit barrier. The longer an organisation has been operating with a specific provider, the higher that potential cost, and the lower its real negotiating power; in documented cases of forced migration after a provider closed down, the cost of rebuilding dozens of AI workflows has exceeded 300,000 dollars, not counting engineering time or the temporary degradation of service.
Why is vendor lock-in in AI different from other software?
Vendor dependence in technology is not a new concept, but in artificial intelligence it takes on specific dimensions that deserve management attention. It occurs when the cost and effort of switching to another provider are so high that the customer is practically forced to stay with the original one, even if the quality of service declines, the product changes drastically or the provider imposes price increases.
In the context of AI, this dynamic is amplified because the model is not just a tool: it is the logic on which workflows, automations and business decisions are built. When the model changes, its outputs change. When the provider updates its usage policies, your company needs to adapt. When there is a service outage, operations stop. A Boston Consulting Group study on dependence on digital platforms in general (not specific to AI, but with a pattern directly transferable to model providers), carried out in November 2025 among 500 technology procurement leaders, found that 62% of them express active concern about lock-in, while 70% are already analysing alternatives on the market and 64% operate with more than one platform simultaneously. The signal is clear: dependence is perceived as a strategic risk, not just a technical problem. Designing the architecture to operate with more than one model from the outset is, precisely, one of the most direct ways to reduce that exposure.
How does the EU AI Act affect companies that use external models?
There is a factor that many management teams have not yet incorporated into their cost analysis: compliance with the EU AI Act. The EU AI Act (Regulation (EU) 2024/1689) is the first comprehensive legal framework on AI and establishes rules for model providers, including transparency and systemic risk assessment. Its rules on general-purpose AI models have been in force since August 2025.
The timetable has changed in recent weeks and it is worth keeping it up to date: on 29 June 2026 the Council of the European Union gave final approval to the simplification package known as the Digital Omnibus, following the backing of the European Parliament on 16 June. All that remains is its publication in the Official Journal of the EU, expected before 2 August 2026. With that text, the obligations for Annex III high-risk AI systems (recruitment, credit scoring, biometrics, education, among others) move from 2 August 2026 to 2 December 2027, and those of Annex I (AI embedded in regulated products) to August 2028. What does not change is the transparency obligation in Article 50 (notifying people when they interact with an AI system or when synthetic content is generated), which still comes into force on 2 August 2026, nor the prohibitions or the penalty regime, in force since 2025.
The key for companies using external models remains the same: even if the provider meets the obligations that apply to it, the deployer (the company using the model) is not exempt. It must verify the suitability of the use case, inform users where appropriate and ensure human oversight with proportionate controls. That responsibility cannot be transferred.
What this means in practice is that, if your company operates with an external model for processes that affect people, and that model does not allow full auditing or traceability, the regulatory risk falls on your organisation. Fines for non-compliance with the EU AI Act can reach 15 million euros or 3% of worldwide turnover for general breaches, and up to 35 million or 7% for prohibited practices. In Spain, supervision falls to the Spanish Agency for the Supervision of Artificial Intelligence (AESIA), which has had the power to investigate and impose penalties since August 2025.
What is lost when all the intelligence lives outside the company?
There is an opportunity cost that does not appear in any spreadsheet but that we at LurNova see frequently: the knowledge about the model that a company never develops when it operates on someone else's infrastructure. Knowing which model works best for each use case, how to structure processes to reduce consumption, what data the system really needs and what is superfluous: all of that is organisational knowledge that does not accumulate when the key intelligence lives outside the company.
According to McKinsey & Company's The State of AI 2025 report, 88% of companies already use AI in at least one function, but only a third have truly scaled it at enterprise level. The remaining two thirds are still at the pilot stage, in many cases because they do not control the system sufficiently to be able to scale it safely. That is exactly the trap of external dependence: the tool is available, but the autonomy to deploy it with judgement is not.
Companies that have opted for private models or their own AI infrastructure tend to develop a learning curve that becomes a lasting competitive advantage. They not only know their tools better, they also retain the ability to change them, audit them and adapt them when the business requires it. That has a value that does not appear in the contract with any external provider.
Conclusion
Depending on an external AI model has a real cost that most organisations underestimate because they calculate it by looking only at the visible part: the subscription or the cost per call. The full cost includes the variable escalation of prices as usage grows, the migration costs that act as an exit barrier, regulatory exposure under the EU AI Act and, above all, the progressive loss of autonomy over a capability that matters more and more to the business.
At LurNova we work with companies that want to implement AI in a way that makes sense not only today, but three years from now. That means analysing the total cost of ownership before choosing a provider, understanding what level of traceability and auditing regulation requires for each use case, and designing an architecture that preserves the client's operational independence. If you want to review how your AI dependence is currently structured and what options you have to reduce it, we can start there.
CEO of LurNova, with extensive experience in technology and digital transformation. Involved in developing technology projects and applying artificial intelligence in organisations, he combines technical knowledge with business vision.
