Model dependency: why model-agnostic workflows are becoming a boardroom issue
Marc GasserSerial Founder · GTM & MarketingConnects AI with revenue operations and builds autonomous GTM systems for predictable growth.TL;DR
- The problem isn't Fable 5, it's dependency. For the first time a US agency pulled an already public frontier model by directive – worldwide, with no right of appeal. Build your AI pipeline on a single provider and you have a single point of failure you don't control.
- Dependency has layers: the model, the cloud beneath it, the API ecosystem of your workflows and political favour. The workflow layer is precisely the one product teams can steer themselves – through an abstraction that turns a provider switch from a re-engineering project back into a config change.
- The answer is neither “away from US models” nor open-source romanticism. The answer is the ability to switch: model choice per project, routing by criticality, hosting under your own control. That is exactly why we built Teklens model-agnostic – model choice as a project setting, not a matter of luck.
Key findings
- VentureBeat captured the lesson in a sentence that can hardly be put more precisely: “centralized, cloud-based frontier models exist at the absolute mercy of government oversight and vendor compliance.” Centralised, cloud-based models sit in the hands of oversight and vendor compliance – not yours.
- The expensive lock-in isn't the model name but the workflow around it: prompt architectures, tooling integrations, agentic workflows, written deep into one specific API. A switch is then a re-engineering project – unless you put an abstraction in between from the start.
- Bitkom president Ralf Wintergerst called for digital sovereignty to be “among the top political priorities”. But companies don't have to wait for Brussels: the diversification levers – inventory, abstraction, routing, hosting choice – are already in their own hands today.
What happened on 12 June 2026 – and why it isn't a tech-insider topic
On 9 June 2026 Anthropic launched Claude Fable 5, by its own account the most capable publicly available model to date, instantly on all major platforms. Three days later, on the evening of 12 June, a US government export-control directive arrived. And Fable 5 was gone. For everyone. Worldwide. Not because Anthropic wanted it, but because an agency decided so.
The mechanics behind it are sober and unsettling precisely because of that. The directive nominally targeted foreign nationals – but a selective block by nationality cannot be implemented in real time in a global cloud system. The only path to compliance was to switch everything off. Including for a company in Linz running its support over Claude APIs that stood there on Monday morning without its most important tool. No right of appeal, no warning period, no European authority that could step in.
You don't have to assess the conflict's background in detail to grasp the finding. What matters is this: it is the first time in the industry's history that an already deployed frontier model was withdrawn from the market by government order. Germany's digital association Bitkom reacted with alarm; president Ralf Wintergerst spoke of dependence “on the goodwill of the US government”.
This is no criticism of Anthropic, which publicly opposed the directive and still had to comply. There was no other choice – and that is exactly the point.
The four layers of model dependency
When teams think about their AI dependency, most think of the model: which version do I use, can I switch from Claude to Gemini? That is the top and most harmless layer. Beneath it lie three more that hurt more in an emergency.
Layer 1: the cloud. Claude runs on AWS, Gemini on Google Cloud, GPT on Azure. All three hyperscalers are US companies under US law. Even a European model would run on American infrastructure at most firms – the dependency reaches one floor deeper than the model name suggests.
Layer 2: the API ecosystem. Here sits the real lock-in. The last two years' workflows are written deep into one specific API: prompt architectures, tool definitions, agentic loops. A model switch is then not a config change but a re-engineering project – unless an abstraction sits between business logic and provider.
Layer 3: political favour. The layer least talked about. Export-control directives are at the executive's discretion – no parliamentary vote, no international coordination, no warning needed. If a trade conflict escalates, the availability of frontier models can become a bargaining chip overnight. I'm not saying it will happen. I'm saying that since 12 June 2026 it is demonstrably possible.
The vendor-risk check: how dependent are you on a single model?
You hang on one provider without steering it. Start with the inventory – it costs days and exposes where a letter from Washington would hurt.
Tick what holds true for you today – the result shows whether a letter from Washington could cripple you on Monday.
What model-agnostic workflows concretely mean
Model-agnostic doesn't mean constantly switching provider. It means being able to switch without rebuilding half the stack. The lever is an abstraction layer: your business logic talks to an internal interface, not directly to the API of Anthropic, OpenAI or Google. Behind that interface the provider is a setting – swappable, testable, selectable per task. The re-engineering project becomes a config change again.
That is exactly why we built Teklens model-agnostic from the start. The Context Engine – semantic search plus a knowledge graph over code, tickets and decisions – is not tied to one model. Model choice is a project setting, not a matter of luck: Claude, GPT or Gemini depending on the task, including EU or Swiss hosting for sensitive projects. The context, the genuinely valuable part, stays yours – whatever model hangs off the bottom. That rules out vendor lock-in, not as a promise but as architecture.
The second half is routing by criticality. Not every task needs the most expensive frontier model. A ticket summary, a classification, a standup text run fine on a smaller or locally hostable model. Frontier performance – multi-hour autonomous workflows, deep code analysis – you reserve for the cases that truly need it. That cuts cost, reduces per-task dependency and, in an emergency, makes part of your workflows instantly sovereign.
Sovereignty without ideology: neither US boycott nor open-source romanticism
This is where the debate likes to tip into the fundamental: away from the Americans, towards European models, ideally all open source. That is well meant and falls short. US models are today simply the most capable options for many tasks, and acknowledging that honestly is part of a grown-up strategy.
European providers like Mistral deliver impressive open-weight models for standard tasks – but for true frontier performance there is a gap today that shouldn't be talked away.
Sovereignty therefore doesn't mean avoiding the best tool. It means never depending on a single one. The EU AI Act governs which risks AI applications pose – transparency, high-risk duties. A sovereignty strategy answers a different question: who controls the infrastructure these applications run on? Both are needed, but they are not the same. The AI Act doesn't solve the dependency problem.
The good news: companies can steer the decisive part themselves, without waiting for Brussels. Put in an abstraction, route by criticality and know the hosting for sensitive data, and you translate a geopolitical risk into an architectural decision. That isn't activism but risk management – the same discipline with which you also accept no single payment provider, no single cloud location and no single supplier for a business-critical part.
Recommendations
- Do the inventory this week. List every workflow with its AI provider, its criticality and its data class. You can't steer a risk you can't see – and the list costs days, not months.
- Put in an abstraction layer. No direct provider API call in the business logic. Talk to an internal interface behind which the model is a setting. That turns the switch into configuration rather than a project.
- Route by criticality. Reserve frontier models for tasks that truly need them. The rest runs on smaller or locally hostable models – cheaper and, in an emergency, instantly sovereign.
- Make model risk a board topic. A named risk with an owner, a metric (how fast could we switch?) and an annual test. Supplier concentration in AI belongs on the same agenda as payments, cloud and logistics.
Scope & caveats
- The account of the Fable 5 events rests on public reporting (including Anthropic's statement, Bitkom, VentureBeat, TechCrunch, NBC News) as of 15 June 2026. It is a developing situation with no official restoration timeline – individual details may still change or be reframed.
- Model-agnosticism isn't free. Abstraction layers, eval harnesses across several models and portable prompts cost engineering time. Calibrate the investment to the workflow's criticality – internal draft text warrants less effort than clinical decision support.
- The capability gap is real: for true frontier tasks, European and open-weight alternatives don't reach the level of the best US models today. Model-agnostic means readiness to switch, not pretending the options are equivalent.
- This article is product-strategy guidance, not a geopolitical forecast and not legal advice. The concrete legal assessment of export controls and data residency belongs in the hands of specialised experts.
The takeaway
The lesson from three days of Fable 5 isn't “away from US models” but “never dependent on a single one”. Make model-agnostic workflows your strategy – model choice per project, abstraction before the provider, hosting under your own control – and you translate a geopolitical risk into an architectural decision you make yourself.
Matching use cases from the library
From the article straight into practice: these use cases put the concepts to work with Teklens.
No new piece without you.
New articles, new interactive tools, new evidence – in your inbox first. And when you reply, we reply: you write directly with the authors, not with a no-reply.
No spam, no sharing, unsubscribe any time.