{ "@context": "https://schema.org", "@type": "Dataset", "name": "AI Vendor Lock-in Matrix — severity + portability tactics 2026", "description": "Vendor-by-vendor analysis of lock-in severity (model API, fine-tuning, vector DB, agent framework, eval platform, observability) plus concrete portability tactics. For each vendor: lock-in axes scored 0-5, switching cost band, exit timeline estimate, mitigations available pre-commitment. Originally compiled for SLAtech client portability audits — published as open data.", "version": "1.0.0", "datePublished": "2026-06-20", "dateModified": "2026-06-20", "license": "https://creativecommons.org/licenses/by/4.0/", "publisher": { "@type": "Organization", "name": "Slavin AI (SLAtech LTD)", "url": "https://www.slavin.ai" }, "isAccessibleForFree": true, "keywords": ["vendor lock-in", "portability", "exit strategy", "AI procurement", "vendor risk"], "methodology": "Lock-in axes scored on a 0-5 scale based on (1) proprietary surface area (custom formats, APIs without open equivalents), (2) data extractability (training-data and fine-tuned weights egress), (3) tooling migration cost (downstream code rewrites), (4) commercial lock-in (long-term commitments, custom contracts), (5) operational lock-in (runtime dependencies). Switching-cost bands derived from real client migrations 2024-2026. Exit timeline = elapsed time from decision-to-switch to full cutover with feature parity.", "scoringScale": { "0": "No lock-in — full open standard, drop-in alternative exists", "1": "Minor — superficial config changes", "2": "Mild — some code refactor, no business impact", "3": "Moderate — significant refactor, possible workflow changes", "4": "Heavy — system rebuild required, business disruption", "5": "Severe — effectively irreversible without major business cost" }, "vendors": [ { "id": "openai", "name": "OpenAI (API)", "category": "Frontier LLM API", "lockinAxes": { "modelApi": 2, "fineTuning": 4, "embeddings": 3, "tools": 2, "commercial": 2, "operational": 2 }, "overallSeverity": "moderate", "switchingCost": "1-3 person-months for typical RAG stack", "exitTimeline": "4-8 weeks", "primaryFriction": "Fine-tuned model artifacts don't transfer; tool/function-calling schema syntax differs from Anthropic; specific behaviors (formatting quirks, refusal patterns) may not replicate exactly", "preCommitmentMitigations": [ "Abstract model client behind own interface (OpenAI-compatible adapter pattern)", "Don't depend on Assistants API state — keep state in your own DB", "Avoid fine-tuning until volume + utility justifies portability cost", "Use LiteLLM, Helicone, or OpenRouter as a routing layer from day one" ] }, { "id": "anthropic", "name": "Anthropic (Claude API)", "category": "Frontier LLM API", "lockinAxes": { "modelApi": 2, "fineTuning": 3, "embeddings": 0, "tools": 3, "commercial": 2, "operational": 2 }, "overallSeverity": "low-moderate", "switchingCost": "1-2 person-months", "exitTimeline": "3-6 weeks", "primaryFriction": "XML-style tool definitions slightly different; system prompt structure (top-down hierarchy assumption) deeper than other models; constitutional AI training shapes outputs in subtle ways", "preCommitmentMitigations": [ "Adapter pattern over Anthropic SDK", "Use OpenAI-compatible Anthropic endpoints where available", "Don't depend on Computer Use API in production until v2", "Embeddings: Anthropic doesn't sell — already portable via Voyage / Cohere / OpenAI" ] }, { "id": "google-vertex", "name": "Google Cloud Vertex AI / Gemini API", "category": "Frontier LLM API + ML platform", "lockinAxes": { "modelApi": 3, "fineTuning": 4, "embeddings": 3, "tools": 3, "commercial": 3, "operational": 4 }, "overallSeverity": "heavy", "switchingCost": "3-9 person-months if heavy Vertex usage", "exitTimeline": "8-16 weeks", "primaryFriction": "Vertex AI Pipelines, AutoML, Feature Store, Model Registry — each adds lock-in beyond just the API call. Tight IAM/VPC-SC integration makes egress harder. Gemini fine-tunes are non-portable.", "preCommitmentMitigations": [ "Use Gemini via direct API (not Vertex SDK) where possible", "Keep pipelines in Kubeflow/Argo, not Vertex Pipelines", "Maintain non-Gemini eval branch in CI to catch silent drift", "Avoid Vertex Search if you can use Pinecone/Weaviate/PGVector" ] }, { "id": "azure-openai", "name": "Azure OpenAI Service", "category": "Frontier LLM API + enterprise integrations", "lockinAxes": { "modelApi": 2, "fineTuning": 3, "embeddings": 2, "tools": 2, "commercial": 4, "operational": 4 }, "overallSeverity": "heavy commercial", "switchingCost": "API compatibility is high but Azure ecosystem (AAD, AKS, App Service, Key Vault, Azure AI Search integration) costs 3-6 person-months to detach", "exitTimeline": "6-12 weeks", "primaryFriction": "API itself is OpenAI-compatible (low technical lock-in), but multi-year EA commitments + ecosystem integration (Azure AI Search, Cognitive Services, Bot Service) create commercial and operational gravity", "preCommitmentMitigations": [ "Negotiate shorter terms", "Keep Azure AI Search optional — use Pinecone/Weaviate/PGVector as primary", "Use openai-python with Azure base_url to maintain SDK portability", "Don't tightly couple to Azure Cognitive Services for AI subcomponents" ] }, { "id": "aws-bedrock", "name": "AWS Bedrock", "category": "Multi-model LLM gateway", "lockinAxes": { "modelApi": 1, "fineTuning": 3, "embeddings": 2, "tools": 2, "commercial": 2, "operational": 3 }, "overallSeverity": "low-moderate", "switchingCost": "1-2 person-months", "exitTimeline": "4-8 weeks", "primaryFriction": "Bedrock is intentionally multi-model so API lock-in is low. But Bedrock Knowledge Bases, Guardrails, and Agents add bespoke surface area. Bedrock IAM integration assumes AWS as control plane.", "preCommitmentMitigations": [ "Use raw model APIs through Bedrock — avoid Knowledge Bases unless time-bounded experiment", "Implement Guardrails portably (open-source NeMo Guardrails / LLM Guard) before Bedrock Guardrails", "Custom Models (fine-tuned) only when you need them in production" ] }, { "id": "pinecone", "name": "Pinecone (vector DB)", "category": "Managed vector DB", "lockinAxes": { "api": 2, "data": 3, "infra": 1 }, "overallSeverity": "moderate", "switchingCost": "2-6 weeks for re-embedding and re-indexing", "exitTimeline": "2-4 weeks", "primaryFriction": "Embeddings tied to model version — switching vector DB usually triggers re-embedding cost. Pinecone's metadata filter syntax is proprietary. Hybrid search (sparse + dense) implementation differs from PGVector/Weaviate.", "preCommitmentMitigations": [ "Keep raw text + chunks in your own DB; treat vector DB as a derived index", "Use a vector-DB abstraction layer (LangChain VectorStore, LlamaIndex)", "Choose embedding model independently from vector DB choice" ] }, { "id": "pgvector", "name": "PGVector (Postgres extension)", "category": "Open vector DB", "lockinAxes": { "api": 0, "data": 0, "infra": 0 }, "overallSeverity": "minimal", "switchingCost": "0-1 weeks (mostly data-export logistics)", "exitTimeline": "1-2 weeks", "primaryFriction": "Performance at scale (>10M vectors) requires careful tuning; HNSW vs IVF tradeoffs. But standard SQL + extension model = full portability", "preCommitmentMitigations": [ "Postgres-anywhere makes PGVector inherently portable", "Use cosine_distance / l2_distance / inner_product — all standard" ] }, { "id": "langchain", "name": "LangChain", "category": "AI application framework", "lockinAxes": { "api": 4, "data": 1, "infra": 2 }, "overallSeverity": "moderate-heavy", "switchingCost": "3-6 person-months for non-trivial chains", "exitTimeline": "6-12 weeks", "primaryFriction": "LangChain's abstractions leak through your code; agents, callbacks, memory, and chain composition often have no clean equivalents in LlamaIndex or Semantic Kernel. LangGraph state machines are particularly framework-locked.", "preCommitmentMitigations": [ "Use LangChain for prototyping, then refactor critical paths to raw SDK before scaling", "Avoid LangChain Hub (proprietary prompt registry) for production prompts", "Keep chains shallow — deep nested chains are exit-tax" ] }, { "id": "huggingface-inference-endpoints", "name": "HuggingFace Inference Endpoints", "category": "Managed model hosting", "lockinAxes": { "api": 1, "data": 0, "infra": 2 }, "overallSeverity": "minimal", "switchingCost": "1-2 weeks (deployment to alternative runtime)", "exitTimeline": "1-3 weeks", "primaryFriction": "Custom handler.py + Dockerfile patterns; ecosystem assumes HF model card format. But model weights and tokenizer always exportable.", "preCommitmentMitigations": [ "Use standard transformers / vLLM / TGI patterns — they run anywhere", "Don't depend on HF Inference's autoscaling guarantees for SLA-critical workloads" ] }, { "id": "weights-biases", "name": "Weights & Biases (W&B)", "category": "ML experiment tracking + production observability", "lockinAxes": { "api": 3, "data": 3, "infra": 2 }, "overallSeverity": "moderate", "switchingCost": "2-6 person-weeks for history migration; ongoing logging migration", "exitTimeline": "3-8 weeks", "primaryFriction": "Run history, artifacts, and project structure don't export cleanly to MLflow/Comet. Custom panels and reports are W&B-specific.", "preCommitmentMitigations": [ "Use MLflow Tracking format alongside W&B (dual-write)", "Don't depend on W&B Sweeps if Ray Tune / Optuna can do the job", "Avoid W&B Launch for orchestration — use Argo/Kubeflow/Prefect" ] } ], "categories": ["frontier-llm", "vector-db", "framework", "platform", "tracking"], "lastReview": "2026-06-20", "nextReviewDate": "2026-12-20", "feedback": "Inaccuracy or want a vendor added? Contact info@slavin.ai" }