Musketier Software
01 AI Consulting 02 Software Development 03 About 04 Blog
DE EN
Arrange a call
← All posts

AI Consulting

Context Lock-In in AI: Why Switching Providers Costs More Than You Think

July 5, 2026 9 min read Leon Lotz

“Swap a model? That’s just a different API endpoint — an afternoon’s work.” That’s how most people think when they start with an AI provider. A survey of 542 executives shows how thoroughly that misleads: 90 % believed they could switch within four weeks — yet 58 % of actual migrations failed or cost significantly more than planned.

The reason has a name: context lock-in. It isn’t the model that ties you down, but everything you build around it. And that raises an uncomfortable question — because the very company behind the industry’s most open standard is accused of building exactly these traps with its current offerings. Let’s look honestly.

At a glance

  • What: It isn’t the model call that binds you (that is swappable) but the accumulated context: finely tuned prompts, memory, agent scaffolding, RAG indexes coupled to an embedding. It grows the longer you stay.
  • The evidence: 90 % consider switching trivial — 58 % actually fail.
  • And Anthropic? Leading on open standards (MCP) and data policy, just as self-interested at its own product layer as any provider. No provider is “lock-in-free”.
  • Your protection: Four affordable architecture and contract steps cover most of the risk.

What context lock-in really is

Everyone knows classic vendor lock-in: proprietary formats, costly exits. With AI, the trap has moved — and that’s exactly what most people miss. The model call is cheap and standardised; an abstraction layer translates between providers in minutes. The switching cost sits in the context you accumulate around the model over months:

  • Prompts and guardrails, meticulously tuned to the behaviour of one model.
  • Memory, projects, assistants — configured in the provider’s format.
  • RAG indexes, coupled to an embedding model: a better embedding forces you to re-embed your entire knowledge corpus (more in RAG vs. fine-tuning).
  • Fine-tunes — trained on provider A, worthless on provider B.
  • Agent scaffolding and institutional knowledge no one has documented.

Cloud lock-in is mostly static — data sits, formats are fixed. Context lock-in accumulates: the longer you use a model, the more ingrained tuning you lose when you switch, and the new provider won’t reproduce that behaviour one-to-one. That context is the real capital of modern AI systems, as we’ve described elsewhere — and it’s that capital that binds you. The most dangerous form is the silent one: building ever deeper out of convenience and never testing whether an exit is still possible.

And what about Anthropic?

This is exactly where the criticism now circulating begins: that Anthropic’s current offerings — Projects, Memory, Skills, the Agent SDK, Claude Code — are themselves such traps. The honest answer has two sides.

Anthropic does more for openness than its competitors. The Model Context Protocol (MCP), released as an open standard in 2024, is now adopted by OpenAI, Google and Microsoft and was handed to a foundation under the Linux Foundation in late 2025 — it makes tool and data integrations provider-independent. And, crucial for GDPR-bound companies: API and enterprise data is not used for training, zero data retention is contractually available. Agent Skills are an open, file-based format that also runs outside Claude.

At the product layer, though, Anthropic very much builds binding. The Agent SDK is proprietarily licensed and tightly coupled to Claude models. Developers criticise Claude Code for using its own directory conventions instead of an open one — “exactly the lock-in MCP was meant to eliminate”, as a widely noted GitHub issue puts it. In early 2026 it was also reported that Anthropic cut third-party tools off from access to its top model without notice and pushed paying users into its own app; at the same time it moved to dynamic pricing. And the 2025 shift on consumer accounts toward opt-out training use shows: data policy is a business decision, not a law of nature.

A small, weightless form of light floats in the upper third of the frame; from it, countless hair-fine bright threads grow downward and weave ever more densely into a firm mesh that grips deep into a dark substrate. At a single point, one thread is vermilion and has fused inseparably with the ground.

It isn’t the model that holds you — it’s the context beneath it: the glowing element on top could be swapped out, but the fine threads it has woven into the substrate over months tear when you pull it free. And one connection has fused for good.

The verdict: openness at the protocol and data layer, binding at the product layer — and this is no Anthropic peculiarity. OpenAI is shutting down its Assistants API on 26 August 2026 and forcing customers to rebuild; Google binds deeply through its Workspace suite. No provider is lock-in-free. Anthropic leads on the standard and data flanks and is just as self-interested at the product flank as everyone else — a commercial company whose openness is strategy, not altruism. For you the conclusion is not “avoid”, but: don’t rely on good behaviour — hedge.

What this means for your company

Pricing and roadmap power is real. If you’re bound deeply, a price or terms change leaves you no credible alternative — and therefore no negotiating power. Reported cases from 2026 range from sharply raised API prices to withdrawn model access. The problem is less any guaranteed increase than the unpredictability you’re exposed to without an exit.

Your workflows break without you touching them. Providers often retire model versions within 12 to 18 months and update hosted models sometimes without notice. The same prompt then yields different results — a process that ran yesterday degrades today. This behavioural drift is one of the most common causes of production incidents in AI systems.

For companies under EU law, exit capability is mandatory. If you have to switch for data-protection or data-residency reasons, context lock-in must not block it. The ability to switch is part of your compliance, not just your IT strategy.

The good news: you don’t need a million-euro architecture. Four steps cover most of the risk — affordable even for mid-sized firms:

  1. An AI gateway between application and model (LiteLLM, OpenRouter) turns a model switch into a configuration change rather than a rebuild. Gartner expects that by 2028 around 70 % of organisations running multiple models will use such a gateway — in 2024 it was under 5 %.
  2. Prompts and data in your hands: prompts as versioned, tested artefacts in your own repository, your knowledge base in your own vector database with a swappable — ideally open — embedding model.
  3. Contracts that secure the exit: a data processing agreement with zero data retention and an exclusion of training use, no multi-year exclusivity, clear data-export clauses.
  4. Keep a second model running for testing: benchmarking two or three models regularly on your own tasks keeps switching practised rather than theoretical — turning a deprecation into planned maintenance, not an emergency.

Conclusion

Context lock-in is not a question of the right provider but of the right architecture. The trap doesn’t spring because you chose Claude, GPT or Gemini — it springs because you build deep without ever testing your own way out. Even the provider with the most open standard remains a commercial actor with its own interests. Your leverage isn’t distrust, but independence from goodwill: portable assets, clear contracts, a practised alternative.

This intersection of technology and law is exactly where I work — as a developer and business lawyer in one person. If you want to know how deeply your AI use is already bound — and how to protect yourself with reasonable effort — let’s talk.

FAQ

What is context lock-in in one sentence?

Your tie to an AI provider does not come from the model itself — that is easily swapped behind a gateway — but from the accumulated context around it: finely tuned prompts, memory, agent logic, and RAG indexes coupled to a specific embedding model. This context grows the longer you stay and cannot be transferred one-to-one to another provider.

Are Claude Projects, Memory and Skills lock-in traps?

Partly. Knowledge files and prompts are exportable text and therefore relatively portable; for memory, two-way import tools exist as of 2026. The real bind is the proprietary Agent SDK and model-specific agent scaffolding. Anthropic leads on the open-standard and data flanks, but on its own product flank it is just as self-interested as OpenAI or Google. Don’t rely on provider goodwill — secure your exit contractually and architecturally.

Is MCP genuinely open — or just marketing?

It is genuinely open: released in 2024, since adopted by OpenAI, Google and Microsoft, and handed to a foundation under the Linux Foundation in late 2025. That is more than rhetoric. Two caveats remain: a standard steered by a consortium of the largest providers is not the same as one controlled by its users — and MCP deliberately does not enforce security, which operators must provide themselves.

How do I concretely protect my company from context lock-in?

Four steps cover most of the risk: an AI gateway as an abstraction layer between application and model, prompts and data in your own hands (your own vector database, a swappable embedding), a data processing agreement with zero data retention and export clauses, and a second model you regularly keep running for testing. For most mid-sized companies this is achievable with reasonable effort.

Sources — as of 2026-07-05

Note: The reported blocking of third-party tools (early 2026) and individual pricing figures come from provider-critical or secondary reporting and are marked here as “reported”, not as confirmed fact.

This article is general information, not legal advice for an individual case. It covers a fast-moving market; as of 5 July 2026, please verify the current state before making decisions.

Leon Lotz

Leon Lotz

Leon Lotz is a business lawyer and founder of MusketierSoftware. He combines legal depth with real software craft.

AI-assisted, editorially reviewed and under editorial responsibility. AI transparency

Related posts

A single tall, solid blade-like form of bone-colored light rises from the lower edge into deep ink-black — solitary, upright, sharply precise. A single vermilion point glows near its tip. The left half of the frame stays quiet negative space. AI Consulting
July 27, 2026 · 11 min

Claude Opus 5: What Anthropic's New Model Can Do, What It Costs — and Where the Record Numbers Demand a Second Look

Read →
Three slender, blade-like forms of bone-colored light stand side by side in ascending size on a deep ink background — small, medium, large. A single vermilion point glows precisely at the tip of the largest form. AI Consulting
July 10, 2026 · 11 min

GPT-5.6: Sol, Terra, Luna — What OpenAI's New Release Can Do, What It Costs, and Where the Benchmarks Mislead

Read →