AI vendor lock-in is invisible
while it forms.
It becomes visible when the vendor changes terms.
tointelligence · omer taki
· why AI lock-in is different

Why AI vendor lock-in is more dangerous than classic lock-in.

speed
Rapid formation
Where an ERP lock-in builds over 3–5 years, AI lock-in can form in 12–18 months. API integrations are fast, teams adapt quickly, and processes rebuild around the model.
invisibility
Implicit dependency
AI lock-in doesn't always come through an explicit contract. It forms in work habits, optimised prompts, fine-tunings completed, and data transmitted to the vendor.
asymmetry
Power imbalance
Major AI vendors have millions of customers. Your organisation is not their negotiation priority. When terms change, they change for everyone : including you.
· three forms of lock-in

Lock-in doesn't always appear where you look for it.

· model lock-in

Your teams optimised workflows around a specific model (GPT-4, Claude, Gemini). Prompts, evaluations, QA processes are all calibrated on this model. Switching models doesn't mean switching APIs. It means recalibrating months of optimisation work.

· data lock-in

Data transmitted to vendors for fine-tuning, embedding generation or RAG systems creates an asymmetric dependency. That data is no longer recoverable in its processed form. It is with the vendor.

· contractual lock-in

Major AI vendors' usage terms evolve. Initially attractive conditions can be modified unilaterally at renewal. An organisation that depends on a single vendor is not in a position to renegotiate.

AI lock-in cannot be negotiated away. It can only be prevented.
· the reversibility test

Are you already in a lock-in situation?

· test now

If your primary AI vendor doubled its pricing tomorrow, how long would it take to migrate to an alternative? If the answer exceeds 6 months : or if you don't know : you are in a significant lock-in situation.

If your AI vendor went offline for 48 hours, which critical processes would stop? That list is your real dependency perimeter.

· prevention principles

Three principles to avoid lock-in without abandoning AI.

Principle 1 : Concentrate dependencies at infrastructure level, not model level. Infrastructure is more standardised and more substitutable than model-level lock-in.

Principle 2 : Document exit conditions before you enter. Every significant AI integration decision must include an exit condition analysis. Not after. Before.

Principle 3 : Maintain an internal evaluation capability. Organisations that can benchmark models internally retain real negotiating leverage. Those that fully delegate this evaluation progressively lose their independence of judgment.