AI vendor lock-in: the question to ask before you pick a platform

At some point in the last year, you will have heard the advice: pick an AI platform and standardise on it. It is reasonable-sounding guidance. Consistency helps with training, pricing, security, and governance. A single vendor relationship is easier to manage than four.

The advice is not wrong. It is just incomplete. AI vendor lock-in is a real risk - and the gap between what "pick a platform" covers and what it leaves out became a lot more visible between May and June 2026.

What happened

In May 2026, Anthropic acquired Stainless - a startup that had built SDK generation tooling used across the AI industry. Stainless was not building tools for Anthropic alone; it was the infrastructure that competitors including OpenAI and Google used to make their APIs accessible to developers. Within days of the acquisition being reported, Anthropic had announced it was winding down Stainless's service to third parties. Developers who had relied on that infrastructure to build integrations with other AI providers suddenly needed to find alternatives.

The deal was reportedly worth over $300 million, according to coverage at the time - though Anthropic and Stainless did not confirm terms. The point is not the number. The point is the speed. One acquisition, and a piece of shared developer infrastructure was gone in weeks.

Around the same time, OpenAI acquired Tomoro, a business built around AI deployment and implementation. And reports emerged - unconfirmed at the time of writing - that Anthropic was separating billing for its Claude chat product from its Agent SDK, effectively creating distinct cost pools for different types of usage.

Three separate events, each individually explainable, but taken together describing something that has been building for a while: the major AI providers are consolidating tooling, acquiring adjacent infrastructure, and adjusting the commercial terms under which businesses use their platforms.

Why this is a workflow decision, not a tool decision

When people talk about AI vendor lock-in, the framing is usually about data portability or subscription contracts. Those matter. But the more practical risk is different.

It is not that a provider will hold your data hostage. It is that the platform will change something - a pricing model, a billing pool, an API version, a third-party integration - and the workflows you have built around it will need rebuilding too. The more deeply a workflow is coupled to one provider's specific implementation, the more expensive that change becomes.

This is not a new risk. It is the same risk that applied to building too deeply on any single software vendor. What has changed is the pace. AI providers are moving faster than traditional software companies, the market is still consolidating, and the tooling layer underneath the main products is shifting in ways that are hard to predict from the outside.

Standardising on one platform is a sensible default for most small businesses. The risk is not standardisation itself - it is building workflows that cannot function if the platform changes something you did not anticipate.

Three questions to ask before committing to a platform

Where does the model appear in the workflow, and what would it take to swap it?

If your workflow is "we use ChatGPT to draft client emails and a human reviews them", swapping the model takes an afternoon. If your workflow is "we have a custom GPT trained on our knowledge base, integrated with our CRM via a proprietary API, and three team members are trained on its specific quirks", swapping the model is a project. Both are valid - but they carry very different exposure if the provider changes something material.

The more a workflow depends on provider-specific features (custom GPTs, Claude Projects, Gemini Gems, proprietary fine-tuning), the more work a migration requires. That is worth knowing before you build.

Which parts of this workflow would still function if the AI component disappeared tomorrow?

This sounds extreme. It is not meant to suggest AI platforms are unreliable - the major providers are robust. The question is really about design. A well-designed AI workflow has clear handoff points where a human can step in, or where a different tool could be substituted. A poorly-designed one has the AI at the centre of something that collapses if the model changes behaviour, the API goes down, or the pricing makes a specific use case uneconomic.

The businesses that get the most durable value from AI tend to design around outcomes, not around specific tools. The tool is a means to an end. The end is a task completed, time saved, output produced. If you can describe the outcome precisely without naming the AI platform, your workflow is in reasonable shape.

What are you agreeing to in the terms, particularly around data and model training?

This is the traditional lock-in question, and it is still relevant. Some AI providers reserve the right to use inputs for model training by default; others do not. Some enterprise tiers offer stronger data isolation than consumer plans. If your workflows involve client data, commercially sensitive information, or anything subject to sector regulation, the default terms may not be appropriate - and changing them later, or migrating to a provider with different defaults, takes time you may not have budgeted. The post on the SME AI stack audit covers a structured way to review what you are actually committed to and where subscriptions are earning their keep.

The portable-workflow principle

The practical answer to vendor risk in AI is not to avoid AI platforms. It is to build workflows where the model is a component, not the architecture.

Think of it this way: a good workflow has a clear input, a defined process, and a measurable output. The AI sits inside that process - it handles the part of the work that benefits from speed, pattern recognition, or language generation. But the workflow itself - the input format, the review step, the output destination, the human decision point - is yours, not the platform's.

When a provider changes something, a workflow built this way requires a change to one component. A workflow built around the provider's specific features may require a rebuild from scratch.

This principle is not complicated to implement. It mostly means resisting the temptation to use every platform-specific feature available just because it is there. A prompt that works in a plain text interface will work anywhere a large language model is available. A process that depends on a bespoke integration with one provider's memory or file storage system will not.

Own the process. Rent the tool. The providers will keep acquiring, consolidating, and repricing. The businesses that will weather that without disruption are the ones that designed for it from the start.

What this means practically

For most small businesses, none of this argues against using a major AI platform. ChatGPT, Claude, and Gemini are all capable, reasonably priced, and improving. The Stainless situation affected developers building infrastructure, not business owners using AI for day-to-day tasks.

What it does argue for is spending time understanding your workflows before you standardise on any tool. Not which tool is best in the abstract, but which tasks in your business are genuinely good candidates for AI assistance, what the right implementation looks like for each one, and how much of your workflow would be at risk if the platform changed its terms or pricing.

That diagnostic step is what most "pick a platform" conversations skip. It is also where most of the value is.

The HoursBack Assessment works through exactly this: a 60-minute conversation that maps your workflows, identifies where AI genuinely earns its place, and produces a five-day plan written in plain English - with specific tools recommended only where there is a clear case for them. If you want a quicker read first, the free two-minute quiz gives you a rough picture of where your business stands.

Sources: Anthropic's acquisition of Stainless and the subsequent wind-down of third-party SDK access were covered in developer community discussions in May 2026, including Hacker News. A reported deal value of $300 million-plus appeared in commentary at the time; neither party confirmed terms. OpenAI's acquisition of Tomoro was reported by the AI Enablement Insider newsletter in June 2026; no deal terms were disclosed. Reports of changes to Claude's billing structure are attributed to industry commentary and have not been confirmed by Anthropic at the time of writing. Any pricing or billing details should be verified directly with each provider before making decisions.