There are now three distinct problems converging around enterprise AI, and they are arriving at the same time. Vendor lock-in is tightening. Regulators are beginning to respond. And the people who are supposed to use these tools every day are, largely, not.
The IBM study that surfaced this week surveyed 1,000 senior executives and found that 71% say switching their primary AI vendor or model would be difficult. That number alone is a warning sign for any business that has moved fast on AI adoption without thinking about exit options. The more startling figure is that 91% of respondents said they do not fully understand their own AI dependencies across vendors, models, and infrastructure.
This is not a theoretical risk. The same study found that organizations experience an average of six AI-related disruptions over two years, and 81% said a seven-day outage from a key vendor would cause severe or critical disruption to their operations. Companies with strong control over their AI systems protect 55% more operating profit during such disruptions than those without. Only 7% of organizations have reached that level of control.
The SAP situation adds a concrete example of what lock-in looks like in practice. In April 2026, SAP published a new policy stating that third-party AI tools cannot freely access SAP's data through its standard technical interfaces. The only approved pathways are SAP's own products: its Joule AI assistant, Business Data Cloud, and a new Agent Gateway. Tools like Microsoft Copilot and Salesforce's AI products now need to route through SAP's own layer to reach SAP data, and SAP meters and governs that traffic.
SAP is offering its own AI tools free through the end of 2026, but the 2027 pricing for that metered access has not been disclosed. This is not a coincidence. Companies that build their AI workflows on SAP's free tools this year will be in a weaker negotiating position when those prices are set next year. One analyst described it plainly as a strategy of getting the customer on the highway, knowing a toll booth is coming.
The EU moved on the broader version of this problem just days ago. The European Commission issued preliminary findings that AWS and Microsoft Azure should be classified as gatekeepers under the Digital Markets Act, the EU's main competition law for large tech platforms. Together, AWS and Azure account for roughly 70% of European cloud revenue. If the designation is confirmed, both companies could face obligations around interoperability and data portability, and fines of up to 10% of worldwide turnover for non-compliance. The process is not final, and both companies are contesting it.
GitHub, owned by Microsoft, also changed its pricing this month. It moved from a flat subscription model to one that charges per unit of AI usage, meaning heavy users of its AI coding tools can see bills jump dramatically. One report found some users going from $39 a month to over $800. This is part of a broader shift across the industry away from flat-rate AI access toward metered consumption pricing.
On the workforce side, the numbers are blunt. Forty percent of employees globally now fear losing their job to AI, up from 28% two years ago. Only 17% of workers use AI tools frequently, despite most organizations claiming AI is now integrated into core processes. Research organizations that invest in structured AI training programs see three to four times higher adoption rates than those relying on self-directed learning, yet 42% of employees say their employer expects them to learn on their own.
The pattern across all three problems is the same. Senior leaders moved fast on AI, often choosing tools based on immediate capability without thinking about supplier power, switching costs, or workforce readiness. Now the bills for those decisions are arriving.
For any organization with AI tools embedded in their operations, three things are worth checking now. First, map which vendors your AI workflows actually depend on, because most organizations cannot answer that question clearly. Second, review any contracts with major software platforms, particularly those that run your core business data, to understand what access rights are changing. Third, measure actual employee usage, not just licenses purchased, because overspending on unused tools is already one of the most common AI outcomes.