AI vendor lock-in: the real risk behind the model war

AI vendor lock-in rarely appears in the marketing material of the platforms now competing for enterprise customers. Google launched Gemini 3, which independent tests place above GPT-5.1 on most measures. OpenAI signed cloud agreements worth USD 38 billion with Amazon. Anthropic reached a valuation of USD 350 billion. For an Italian small business, the benchmark race matters far less than a practical question: what does this competition do to the tools the company already depends on?

Beyond benchmarks, strategy decides

The numbers tell one story. Gemini 3 reached 1,501 Elo points on LMArena, ahead of its competitors. GPT-5.1-Codex-Max is designed to work across millions of tokens and to complete tasks that run for over 24 hours. These are technically relevant results.

Then Oracle lost USD 315 billion in market capitalisation after announcing a USD 300 billion agreement with OpenAI. The market signalled that concentrating resources on a single AI platform is an existential risk, even for a technology giant. If the lesson holds at that scale, it holds even more for a small business with a limited budget and no internal AI team.

The real competition is between two business models rather than two models. Google pushes an integrated ecosystem: Gmail, Drive, Calendar, Search, with Gemini as the layer that ties the customer closer to those services. Microsoft and OpenAI do the same with Azure and Office 365. The convenience is real, and so is the dependency underneath it.

The gilded prison of vendor lock-in

The mechanism is easy to describe and hard to escape. A company adopts an AI system to automate customer service, choosing a well-known platform, often the one its cloud provider recommends. The system works. Six months later the team has customised hundreds of workflows, integrated the system with CRM and ERP, trained the model on its own data, trained the staff and accustomed its customers.

Then the conditions change: the vendor raises prices by 40%, or rewrites the contract, or the model falls behind competitors. At that point the company discovers that leaving costs more than staying. The data sits in proprietary formats, the integrations need rewriting, the staff need retraining. Migration often exceeds 200% of the initial investment.

The MIT NANDA Initiative found that 95% of generative AI pilots deliver no significant results, and the dominant cause is badly designed enterprise integration and dependency on vendors that push generic solutions. Micro and small businesses are the most exposed: they have no internal IT team to evaluate alternatives, no negotiating power to improve terms, and no capacity to absorb a failed migration. When Builder.ai collapsed, hundreds of companies were locked out of their own applications, with inaccessible code, trapped data and critical systems down.

The hidden costs that turn an investment into a loss

A vendor quote usually lists a monthly price per user or a cost per token. The real budget starts after that. Data preparation comes first: a model performs as well as the data behind it, and cleaning, categorising and structuring years of information spread across spreadsheets and legacy databases takes weeks or months of internal time that rarely appears in the initial estimate.

Integration with existing systems follows, because no company operates in isolation, and connecting AI to ERP, CRM, e-commerce and warehouse platforms requires custom development, APIs and middleware. Quoted budgets of EUR 10,000 routinely become EUR 30,000. Continuous maintenance adds a third layer: models age, data shifts and user behaviour evolves, so retraining, performance monitoring and adjustment never stop, and the company pays for all of it.

Compute consumption grows with usage, and a chatbot handling a thousand conversations a day can cost ten times a traditional solution. Staff training is the fifth line: people must learn to write prompts, interpret output and recognise failure, or the result is workslop, content that looks correct, reads superficially and takes longer to fix than to write from scratch. Among Italian SMEs, 42% of AI projects are abandoned within the first year, more than double the 2024 figure. The main cause is economic rather than technical: real costs exceed the initial forecast by 150% to 200%.

Hardware depreciation and the Oracle lesson

Oracle lost USD 315 billion in capitalisation within days of its OpenAI agreement because the market saw the risk: committing USD 300 billion to AI infrastructure means holding hardware that may be obsolete in 18 months. Nvidia H100 GPUs, the reference point in 2024, have been overtaken by Blackwell chips, and the next generation arrives within twelve months. Each cycle brings performance gains of 200% to 400% and makes the previous generation economically hard to justify.

For a small business the mechanism repeats on a smaller scale. A company builds on the cloud infrastructure of provider X. Two years later provider Y offers more powerful hardware at a lower cost, but switching means rewriting code, rebuilding integrations and reconfiguring systems. The migration cost exceeds the saving, so the company stays on older technology and pays more for lower performance. Lock-in is not a defect of the vendor model; it is part of the design.

Five questions before signing

The first question is whether the data remain the company's property. A clear contract covers legal ownership, storage location, whether the vendor uses the data to train its own models, and whether everything can be exported in a standard format. Anything short of an unqualified yes is a warning sign.

The second is the cost of leaving. A transparent vendor answers this directly and accepts a migration support clause in the contract. A vague answer means the exit path is expensive.

The third is what happens if the vendor fails or changes strategy. Builder.ai left customers without access to their systems, and dozens of AI startups close every year. Code escrow arrangements, standalone operation of critical systems and credible alternatives on the market all reduce the exposure.

The fourth is whether the underlying model can change without rewriting workflows and integrations. Architectures built on open standards, such as the Model Context Protocol supported by Anthropic, Cloudflare, OpenAI and Microsoft, separate the model from the business logic and make substitution realistic.

The fifth is the three-year total cost of ownership, covering licences, compute, data preparation, integration, training, maintenance, upgrades and support. A serious quote for an AI project usually carries at least 60% of cost beyond the base licence.

The false promise of simplicity

Vendors sell simplicity: one click, no technical skills required, live within 24 hours. The claim is marketing. Applied AI is complex because business processes are complex, and it demands an understanding of how the company works, what data exists, which use cases are real, how integrations are designed and how people adapt. A monthly subscription does not remove that work.

The scepticism many Italian entrepreneurs show toward digital projects is experience rather than ignorance, earned through expensive disappointments and consultants who promised transformation and delivered complexity. AI can create real value for a small business, and successful cases share a pattern: bounded and measurable goals, gradual integration, strong staff involvement and independence from any single vendor.

Building strategic independence

Three principles lower the risk. The first separates business logic from AI technology, so that workflows, integrations and process rules live in a layer independent of the model underneath, which allows a change of vendor or technology without a full rewrite. The second protects complete ownership of data and its format, keeping every record exportable in a readable, standard format. The third makes alternative evaluation continuous, because the AI market changes every quarter and strategic choices need room to adjust.

These principles carry a cost in initial effort and design. They also protect against forced obsolescence and economic dependency later.

The real battlefield

The competition between AI models is not decided on benchmarks. It is decided by how the relationship between vendors and customers is structured. OpenAI, Google, Microsoft and Anthropic are building ecosystems that maximise customer dependency, which is a business model rather than a conspiracy: the more services a company integrates, the more expensive leaving becomes, and the more proprietary formats it uses, the more it is tied.

Oracle lost USD 315 billion because the market understood that dynamic. Small businesses have no capitalisation to burn, and the same mechanism applies. The difference lies in awareness. An entrepreneur who chooses an AI system knowing the hidden costs, the lock-in, the hardware depreciation and the available alternatives can make sustainable decisions. One who relies on the vendor's promises ends up trapped inside them.

Lock-in also has a hardware side: the effect of AI demand on hardware prices shows how quickly infrastructure choices age.