VentureBeat Pulse Research surveyed 145 qualified respondents in June 2026, spanning the period when Anthropic’s Claude Fable 5 was taken offline by a U.S. export-control order on June 12. The survey finds that two-thirds of enterprises had already hedged their AI model strategy before the blackout, and the Fable 5 disruption illustrated why such hedging has become mainstream.
What happened and why it matters
Claude Fable 5 launched on June 9 and was notable for both its capabilities and its price (reported at $10 per million input tokens and $50 per million output tokens). On June 12 the U.S. issued an emergency export-control directive prohibiting access by foreign nationals; because Anthropic could not verify nationality in real time, it suspended access for all customers. The model later returned with tighter safeguards. The incident made clear the operational risk of vendor dependency: a model a company relies on can be made unavailable overnight by external action.
Beyond vendor dependency, the survey highlights a wider problem: many enterprises lack automated monitoring to detect when an AI system in production drifts, behaves unsafely, or fails.
Key numbers from the survey
- 51% of respondents run a hybrid posture: closed frontier models for general reasoning and open-weight models deployed on their own infrastructure for specialized execution. An additional 16% are moving core workflows entirely off closed APIs. The remaining 33% relied exclusively on closed ecosystems when the blackout occurred.
- Only 1 in 10 enterprises (14 out of 145) have automated monitoring and alerting that would detect model drift, misbehavior, or production failure.
- 32% expect to catch most issues only “eventually,” 19% would likely learn of a failure from end users first, and 8% report no systematic visibility into production AI behavior.
- 79% of enterprises have already experienced a real financial or operational hit from autonomous agents. Of these, 49% named shadow AI (unauthorized, departmental agentic pipelines charged to corporate cards) as the most severe issue; 25% cited infinite-loop bills from uncaught recursive workflows; 6% reported agents that degraded production databases via unthrottled queries.
The Control Gap
VentureBeat labels the gap between rapid AI deployment and weak governance the “Control Gap.” Across independent survey questions the same pattern emerged: deployment is running ahead of governance, visibility, and cost control. The Fable 5 shutdown acted as a stress test of that gap.
Organizational barriers to governance
The top-cited barrier to governing AI across platforms is the absence of a single owner or accountable team (32%). Vendor opacity is second at 25%, missing tooling at 16%, and lack of talent appears last at 5%.
- Only 38% say a central team currently governs AI behavior across their platforms. 21% report ownership is unclear or contested, and 17% say no role holds formal accountability.
- 85% operate two or more platforms that each claim to be the “primary” AI layer (ERP, ITSM, productivity suites, data platforms, etc.); 36% describe an open contest among four or more such platforms. Only 8% have consolidated to a single platform.
Respondents commonly suggested one corrective: assign a single accountable owner and build a control plane that abstracts cost, drift, and model choice away from end users.
Cost control, tokens and agent economics
Per-token inference costs are falling 70–80% per year, but agentic workloads consume 100–500x the tokens of the LLM tools they replaced, so total consumption can explode. Examples cited in the survey and reporting include Uber burning through its 2026 AI coding budget in four months after widespread Claude Code adoption, and Microsoft canceling many internal Claude Code licenses in favor of its own tooling.
Experts at VentureBeat events recommended right-sizing models and employing semantic routing so the platform sends only those requests that truly require frontier-scale reasoning to expensive models, while routing commodity work to smaller, specialized models.
Operational approaches and human review
While many firms rely on human review as a control — for example Liberty IT and Morgan Stanley described human sign-off layered on observability and governance — human review alone scales poorly as agentic workloads multiply output volumes. The leaders quoted stress that human accountability should sit on top of automated observability, identity, and governance tooling.
Bottom line: replaceability outpaces ownership
The survey paints a picture of enterprises moving quickly on AI but with weak controls beneath:
- 58% are adding more AI initiatives than they retire.
- 85% run multiple competing AI platforms.
- Three times as many enterprises rely on human review to catch a failing production model as have automated monitoring in place.
- 79% have already paid for an agent control failure, most often unauthorized agent spending on corporate cards outside IT oversight.
On model dependency, enterprises have adapted: two-thirds hedge their model strategy (51% hybrid, 16% moving off closed APIs). The Fable 5 shutdown validated the value of that posture, allowing hedged organizations to route around a model that became unavailable by government order.
However, internal governance problems—most importantly the lack of a single accountable owner (32%) and the 17% reporting no formal accountability—remain unresolved. Assigning an owner requires no vendor purchase yet many companies still have not done so.
VentureBeat plans a Q3 follow-up to measure whether organizations assigned owners and installed automated monitoring after June’s events, or simply added a second model and moved on.
Data and methodology: VentureBeat Pulse Research surveyed 145 respondents at organizations with 100 or more employees in June 2026; fielding spanned the Fable 5 blackout that began June 12. The sample is self-selected and directional: 41% work in technology/software, 20% are consultants or advisors, and responders skew senior and technical (CIO/CTO/CISOs 18%, directors of engineering/IT 14%, enterprise architects 12%). More than half were from companies with 10,000+ employees. The full methodology is in the report.
VentureBeat’s VB Transform event (July 14–15, Hotel Nia, Menlo Park) will center on themes from this report: agent orchestration, governance, and cost control. Disclosure: VentureBeat’s June 24 AI Impact event was sponsored by Red Hat and Intel; sponsors had no input into the Pulse Research design, findings, or editorial coverage.
