Where AI dependency comes from
Provider Lock-In
The first source of fragility is leaning on one model or platform you do not control. When it reprices, restricts, or disappears, your workflows stop.
- Single point of failure. Core delivery runs on one model with no fallback
- Locked-in assets. Prompts, SOPs, and data live inside a tool you cannot export from
- No swap path. Switching providers means a rebuild, not a config change
- Geo and export risk. Access can be restricted by regulation, as Fable showed
The fastest way to see your own exposure is to run the AI Dependency Audit. Four minutes, scored across both sources of fragility, with the first fixes for each.
72 hrs
How long Claude Fable 5 was publicly available before a US export-control directive took it offline globally
Own the method, not the model
The model is rented infrastructure. The method is yours. Your frameworks, prompts, SOPs, brand voice, client context, decision rules, and QA standards do not live inside any model. They are portable, if you keep them that way.
A consultant who has encoded their methodology into structured prompts and process documents can swap the underlying model in an afternoon. One who built everything inside a single tool is starting over. This is the same discipline as owning your AI stack and writing model-agnostic client proposals.
Keep the model swappable: cloud, local, or hybrid
Cloud frontier model
Best capability per dollar for complex reasoning. The trade-off: data leaves your control and access can be repriced or restricted overnight.
Local open-weight model
Runs on your own hardware via Ollama. Within ~0.3% of frontier on many evals. No centralised API for anyone to switch off. Needs setup and ops.
Hybrid (recommended)
Cloud frontier for the hard 10-20% of work, local or cheaper models for the rest. Capability where it counts, resilience and a sovereign fallback everywhere else.
Keep a human in the loop
Resilience is not only about providers. Automating delivery with no review step erodes the quality that justifies a premium price. The rule: if a wrong or low-quality AI output would damage trust, trigger a refund, or mislead a client or student, a human reviews it before it goes out.
For cohort and course businesses, this is the difference between a premium program and a content mill with a human brand on it. The full playbook is in anti-fragile AI for cohort businesses.
Measure it, then fix the weakest part first
You cannot fix what you have not measured. Run the blackout test: if your primary AI tool went offline tonight, what halts in 24 hours? That list is your priority order. Track how often your team revises AI output rather than shipping it, and whether they could still do the core work without it.
The AI Dependency Audit → scores both dimensions and hands you the first fixes. It pairs with building durable AI search visibility so neither your delivery nor your distribution sits on a single switch.