After 16 years as a HubSpot partner, we watched companies get locked into SaaS platforms from the inside. We built the integrations. We watched the switching costs compound. We saw the mechanisms of lock-in tighten until leaving felt impossible.
Now we’re watching the exact same pattern play out with AI. From both directions.
The obvious vector: SaaS vendors are bolting AI onto their existing platforms. HubSpot Breeze, Salesforce Einstein, and Microsoft Copilot all use the identical playbook that worked for a decade. You adopt their AI, train it, and within two years, switching becomes the one project nobody wants to approve.
The less obvious vector: the AI companies themselves are becoming SaaS vendors. Anthropic, OpenAI, and Google are selling enterprise subscriptions, building workspace features, shipping connectors to every platform in your stack. Same bundling playbook HubSpot perfected, except the product is the AI itself. And the lock-in compounds faster because the AI learns your business while you use it.
Companies are getting squeezed from both ends. SaaS vendors want to own the AI layer on top. AI providers want to own the platform underneath. Both want to be the thing you can’t leave.
Is AI creating the same lock-in that SaaS did?
Yes. AI vendor lock-in happens when your AI capabilities rely on a specific platform, creating switching costs that often exceed traditional SaaS lock-in. The additional dependencies, like proprietary training data and embedded workflow logic, make AI lock-in harder to unwind than CRM dependency ever was.
The SaaSpocalypse that erased $800 billion in SaaS market value wasn’t just about overpriced subscriptions. It was about the structural dependency that made those subscriptions feel mandatory. Bain & Company documented the pattern: SaaS price-to-sales ratios compressed from 9x to 6x as enterprises realized they were paying premium prices for commodity software wrapped in switching costs.
AI adoption is following the same arc at three times the speed. Global AI spending hit $2.52 trillion in 2025, up 44% year-over-year (Gartner). Mayfield’s 2026 survey found 72% of enterprises are in production or piloting agentic AI. But Deloitte found only 11% are actively using agents in production — meaning the vast majority are still in the “adopt and experiment” phase where vendor lock-in takes root.
This is the window where dependency is created. Not when the technology is mature, but when companies are experimenting, building workflows, and training models within vendor ecosystems. By the time you realize you’re locked in, the switching cost is already priced into your operations.
Why are companies adopting AI from the vendors they’re trying to leave?
Because it’s the path of least resistance. When your CRM vendor adds AI, the pitch is frictionless: it already has your data, it already knows your workflows, and the feature is “included” in your existing contract. The integration work is zero. The learning curve is minimal.
This is exactly how SaaS lock-in worked for the last 15 years. The initial adoption was always easy. HubSpot’s genius was never the product — it was the bundling. CRM, CMS, marketing automation, email, analytics — all in one platform. Each addition felt logical. Each addition deepened the dependency.
Now the same vendors are adding AI to the bundle. HubSpot launches Breeze and it works across your contacts, your content, your workflows. Salesforce announces Einstein GPT and it’s embedded in every cloud. Microsoft adds Copilot to the entire Office suite. The pitch is always the same: “It’s already integrated. Why build something separate?”
AI vendor lock-in adds a new dimension on top of the existing five: the Intelligence Lock. Your AI learns from your data, inside a vendor’s system. That learned intelligence is the asset you can never take with you.
The answer we learned across 2,000+ projects is that integration today becomes dependency tomorrow. Having AI pre-integrated with your vendor’s platform is convenient, but the AI only sees what the vendor shows it. When you want to change providers, the AI layer is the hardest thing to extract.
We watched this with HubSpot workflows for over a decade. Clients would build automation logic — hundreds of workflows, scoring models, behavioral triggers — and that logic existed only inside HubSpot. It was The Logic Lock in action. AI vendor lock-in adds a new dimension on top of the existing five: the Intelligence Lock. Your AI learns from your data, inside a vendor’s system, using a vendor’s framework. That learned intelligence is the asset you can never take with you.
What does AI vendor lock-in actually look like?
AI vendor lock-in manifests in four ways that compound over time: proprietary data enrichment, embedded model training, platform-specific AI frameworks, and priced-in AI features that inflate your contract while reducing your leverage to leave.
Proprietary data enrichment. When your vendor’s AI processes your customer data, the enriched output — intent scores, behavioral predictions, content recommendations — exists only within the vendor’s system. Salesforce Einstein’s predictive lead scoring doesn’t export. HubSpot Breeze’s content analysis doesn’t port to other platforms. The AI creates new data that you can’t take with you, and over time, that derived data becomes more operationally important than the raw data you originally brought.
Embedded model training. As your team provides feedback to vendor AI — “this suggestion was useful,” “this lead score was wrong” — the model adapts to your patterns. That adaptation lives inside the vendor’s infrastructure. When you switch platforms, you don’t just lose the model. You lose the months or years of implicit training your team invested. Deloitte projects over 40% of agentic AI projects will fail by 2027, and legacy system limitations are the primary cause — but switching AI providers also means restarting the training cycle from zero.
Platform-specific AI frameworks. HubSpot’s AI operates only on HubSpot data, through HubSpot’s interface. Salesforce’s AI operates only on Salesforce records. Microsoft’s AI operates only within the Microsoft ecosystem. Each vendor has built an AI framework that works nowhere else. The automation your team builds — the configurations, the decision logic, the workflows — none of it transfers. Every hour your team spends building inside these frameworks is an investment with zero portability and zero residual value outside the vendor’s walls.
Priced-in AI that inflates your contract. Vendors are bundling AI into higher-tier plans, using AI as a pricing lever. The features are “free” — but only at the enterprise tier you wouldn’t otherwise need. This is the same strategy that turned $200/month CRM subscriptions into $43,000/year platform commitments. Retool’s 2026 data shows 35% of enterprises have already replaced at least one SaaS tool. AI bundling is the vendor’s response: make leaving even harder by embedding the technology every executive is asking about.
Are the Five Locks repeating in AI?
Every lock from The Five Locks framework has a direct AI parallel. The mechanisms vendors used to prevent departure over the last decade are being rebuilt into AI adoption patterns.
The Code Lock becomes the Framework Lock. Proprietary template languages trapped your website code inside one vendor. Now, vendor-specific AI agent configurations — Salesforce’s Agentforce, HubSpot’s Breeze workflows, Microsoft’s Copilot Studio — trap your AI logic inside the platform. The implementation detail is different but the business consequence is identical: work product that has zero value outside the vendor’s ecosystem. Every dollar you invest in building on these frameworks is a sunk cost trapped in someone else’s infrastructure.
The Data Lock deepens. Traditional SaaS locked your existing data — contacts, deals, content. AI vendor lock-in adds a new layer: derived data. The predictions, scores, and enrichments that AI generates from your data are new data assets that the vendor controls. You brought the raw material; they own the refined product.
The Logic Lock compounds. Workflow automation in SaaS was hard enough to extract. AI decision logic is harder because it’s often implicit — embedded in the vendor’s system in ways that can’t be documented, exported, or reproduced elsewhere. The tribal knowledge problem that plagued SaaS automation becomes even more acute: the “intelligence” your team spent months refining is an asset you can never take with you.
The Audience Lock extends to AI audiences. When AI segments your audience based on behavioral patterns the vendor’s model identifies, those segments exist only inside the vendor’s system. The AI-generated audience definitions are as locked as manually-built ones — more so, because you may not even understand how the AI drew the boundaries.
The Dependency Lock becomes existential. When your team relies on vendor AI for daily operations — content generation, lead scoring, customer service — the organizational dependency goes deeper than software. You’re dependent on the vendor’s AI roadmap, their model quality, their rate limits, and their pricing decisions. If they deprecate a feature, raise API costs, or pivot their AI strategy, your operations absorb the impact with no alternative.
Are AI companies becoming the new SaaS vendors?
The packaging looks different, which is why most companies aren’t seeing it. OpenAI, Anthropic, and Google aren’t selling CRMs or marketing platforms. They’re selling intelligence as a subscription. But the business mechanics are converging fast.
Look at what ChatGPT Enterprise and Claude for Work actually are. Monthly per-seat pricing. Workspace features. Admin consoles. Usage-based tiers that reward consolidation and punish switching. These are enterprise SaaS products. The pricing model is structurally identical to what Salesforce and HubSpot have been running for 15 years.
Then there are the connectors. ChatGPT and Claude now ship integrations to Slack, Google Workspace, Salesforce, Jira, Notion, and dozens of other platforms. The pitch sounds like liberation: “Use AI across all your tools.” But every connector deepens the dependency. Every workflow your team builds through a single AI provider’s connector layer is another workflow that breaks when you switch.
MCP (Model Context Protocol) is a good example of where this gets tricky. On paper, MCP is a step toward interoperability — an open standard for connecting AI to external tools and data. In practice, most of the early implementations are provider-specific and half-built, designed to pull your workflow into one vendor’s orchestration layer rather than make it portable. The promise is “connect everything.” The reality, so far, is “connect everything to us.”
This matters because the model market doesn’t sit still. OpenAI led in 2023. Anthropic pulled ahead on reasoning in 2025. Open-weight models from Meta and Mistral closed the gap on commodity tasks. Google’s Gemini keeps shifting the price-to-capability ratio. A company locked into one provider’s enterprise subscription — with workflows wired through that provider’s connectors, their team trained on that provider’s interface, institutional knowledge stored in that provider’s conversation history — can’t follow the market. They’re paying whatever their vendor charges, regardless of what the market offers.
We watched this exact dynamic at HubSpot. The CRM was never irreplaceable. The workflows were. The automations, the reporting logic, and the team habits were all built inside HubSpot’s walls. When a better option appeared, switching meant rebuilding from scratch. AI enterprise subscriptions work the same way. The model is replaceable. The organizational dependency around the model is not.
Where does vendor AI have legitimate value?
Vendor AI is genuinely useful for commodity tasks that don’t create meaningful dependency: basic content drafts, simple workflow suggestions, and surface-level data summaries within a platform you’re already committed to. For companies under 50 employees with modest AI ambitions, pre-integrated AI may outweigh the lock-in risk.
The honest assessment of specific platforms:
HubSpot Breeze has legitimate uses for basic content generation, email subject line testing, and simple chatbot deployment within the HubSpot ecosystem. If you’re staying on HubSpot and your AI needs are limited to these commodity tasks, Breeze adds value without creating additional lock-in beyond what HubSpot already has. But it operates only on HubSpot data, cannot integrate with external AI frameworks, and cannot be extracted or ported.
Salesforce Einstein is genuinely strong for lead scoring and opportunity prediction within the Salesforce CRM — if your sales data lives in Salesforce and you don’t need AI that works across other systems. For pure CRM intelligence where the data stays inside one platform, it’s defensible.
Microsoft Copilot is useful for document-level productivity — drafting, summarizing, formatting — within the Office suite. For knowledge work that stays inside Microsoft’s ecosystem, it reduces real friction in daily operations.
The danger isn’t using vendor AI for commodity tasks. The danger is building your AI strategy around a vendor’s roadmap instead of your own.
The pattern: vendor AI works when the task is small, the scope is narrow, and you’re already committed to the platform. It falls short when you need AI that crosses system boundaries, learns from data the vendor doesn’t control, or becomes a core capability you can’t afford to lose.
The danger isn’t using vendor AI for commodity tasks. The danger is building your AI strategy around a vendor’s roadmap instead of your own.
How do you adopt AI without getting locked in?
Build on open-source frameworks, use API-agnostic architectures, keep your training data portable, and treat model providers as interchangeable utilities — not strategic partners. The infrastructure cost is lower, the switching cost is near zero, and the capability ceiling is set by the market, not one vendor’s product team.
Use open-source AI frameworks. Frameworks like the Vercel AI SDK, LangChain, and LlamaIndex provide AI orchestration that works across model providers. Build on these, and switching from one AI provider to another becomes a configuration change instead of a migration project. Your AI investment stays portable. The work product you build carries forward regardless of which model provider offers the best price or capability next quarter.
Self-host where the data is sensitive. For AI that touches proprietary data — customer intelligence, pricing logic, competitive analysis — self-hosted models eliminate the risk that your data improves a vendor’s model or gets exposed through their infrastructure. Open-weight alternatives (Llama, Mistral, and others) make this feasible at a fraction of what it cost even two years ago.
Keep your AI logic in code you own. When your AI configurations live in your own code repositories instead of vendor configuration screens, they’re portable by default. Switching providers becomes a simple configuration change, not a project to rebuild months of accumulated work. The AI logic your team builds becomes an appreciating asset in your codebase — not a sunk cost inside someone else’s platform.
Design for multi-model from day one. The AI model market moves faster than any software market in history. The leading model shifts every few months, and the price difference between providers can be 5-10x for comparable capability. Your AI infrastructure should route different tasks to different models. No single provider is best at everything, and no single provider’s pricing is best for every use case. Use Claude for complex reasoning, GPT for high-volume commodity tasks, open-weight models for sensitive data that can’t leave your infrastructure. An AI-native architecture makes this possible by treating model providers as interchangeable utilities behind a common interface. Switching — or splitting workloads across providers — becomes a configuration change, not a migration project.
Own your training data. If your AI improves through feedback loops — and it should — that feedback data must live in your infrastructure. When the training data is yours, the intelligence is yours, regardless of which model processes it.
This is the core idea behind the SaaSpocalypse: own the infrastructure that creates value. A recent Retool survey found 35% of enterprises have already replaced at least one SaaS tool. They aren’t replacing it with another vendor’s platform. They’re building on open infrastructure they control.
The market is already splitting
The data shows two divergent paths forming, and the gap between them is widening every quarter.
Path A: Vendor-dependent AI. Companies adopt AI through their existing SaaS vendors, or they consolidate on a single AI provider’s enterprise subscription and wire everything through that provider’s connectors. Both flavors feel productive in the first year. Within 18-24 months, the AI logic, training data, and organizational dependency create switching costs that rival the original SaaS lock-in. The two flavors of Path A are converging: SaaS vendors want to own the AI layer, AI providers want to own the platform layer, and companies on either track end up locked in.
Path B: AI-sovereign infrastructure. Companies build AI capabilities on open frameworks, self-hosted models, and API-agnostic architectures. The upfront investment is higher, the learning curve is steeper, and the first six months show less visible progress. But the long-term cost is lower, the capability ceiling is higher, and the switching cost stays near zero.
Global AI spending hit $2.52 trillion in 2025 (Gartner), and it’s splitting along these lines. The 72% of enterprises piloting agentic AI (Mayfield, 2026) include both paths. But the companies on Path A are discovering what we spent 16 years watching in SaaS: the vendor’s interests and your interests diverge the moment the contract is signed.
Deloitte’s projection — over 40% of agentic AI projects failing by 2027 — will disproportionately hit Path A. Not because vendor AI is incompetent, but because the infrastructure underneath can’t support what AI needs to do. The same systems that created the SaaS lock-in problem — siloed data, rigid workflows, limited access to your own information — are the same systems companies are now layering AI onto. Adding AI doesn’t fix the underlying constraints. It deepens the dependency on an architecture that was designed to keep you paying, not to keep you capable.
The companies that get this right in 2026-2027 will own their AI capabilities the way they should have owned their CRM data and their CMS code and their analytics infrastructure all along. The companies that don’t will be writing another round of vendor lock-in recovery plans in 2028 — with “AI” in the subject line instead of “SaaS.”
Find out where your website stands on AI readiness. Get your free assessment — our AI evaluates your site’s tech stack, performance, and AI readiness in 60 seconds.