Enterprise AI Controls Lag as 20% Cannot Halt Runaway Agent Spending
VB Pulse finds 20% of enterprises cannot halt runaway AI agent spending in real-time, while 85% use multiple orchestration tools to retain tight control.
Summary
VB Pulse’s August 20, 2026 survey of 107 enterprises finds the median uses three orchestration platforms: 85% use at least two, 64% use three and 15% use one. Microsoft AI Foundry/Copilot Studio appears in 70% of stacks, OpenAI’s Agents SDK in 68%, Anthropic’s Claude Platform in 47% and custom in-house orchestration in 22%; Google’s Enterprise Agent Platform, LangChain/LangGraph, Salesforce Agentforce, Amazon Bedrock and LlamaIndex also appear. Microsoft leads current use; Anthropic leads consideration.
By end-2026, 53% expect a hybrid primary control plane, versus 14% provider-managed, 13% custom in-house and 11% provider-independent external. More than two-thirds plan to switch within a year: 15% within three months, 24% in three to six and 28% in six to 12; 43% are exploring Claude Agent SDK, roughly one-third Google, 31% custom and 25% OpenAI. Satisfaction averages 4.17/5 overall, 3.91 for implementation and 3.63 for value. Purchasing weights flexibility at 29%, security/permissions 17%, reliability and execution control 15% each, model gravity 10%, development ease 8%, ownership cost 4%, and latency/memory 2%.
Spending favors monitoring/debugging, 31%, and security/permissions, 30%, over workflow tooling, 19%, which led one month earlier; priorities are task reliability, 30%, multi-step workflows, 27%, developer productivity, 23%, stability, 13%, and UX, 7%. Concerns are security/permission limits, 37%, lock-in, 23%, visibility, 22%, and model/tool inflexibility, 16%. One in five cannot stop runaway spending in real time; controls are native caps, 30%, custom gateways, 25%, dynamic routing, 25%, and reactive logs without kill switches, 21%; reactive-only rates are 18% at 10,000-plus employees and 23% below. True orchestration spans 76% to 100% of systems for 2%, 51% to 75% for 14%, 26% to 50% for 47%, and 1% to 25% for 35%; 3% deploy only chatbots. June Pulse found 71% had at most one-quarter of agents completing autonomous multi-step work and 10% had agents at scale.
Positives
- 85% of surveyed enterprises use at least two orchestration tools, preserving flexibility across vendors, models and agents.
- 53% expect hybrid primary control planes by the end of 2026, exceeding every single-provider or custom alternative.
- Overall platform satisfaction reached 4.17 out of 5 despite weaker implementation and value scores.
- 61% of orchestration spending targets monitoring, debugging, security and permission enforcement.
- 47% report true orchestration across 26% to 50% of their systems, showing progress beyond basic assistants.
Risks & concerns
- One in five enterprises cannot halt runaway AI agent spending in real time.
- 37% cite security and permissioning limitations, while 23% fear vendor lock-in and 22% report inadequate visibility.
- 21% rely solely on reactive logs and lack real-time kill switches for agent spending.
- Value for money scored 3.63 out of 5, below overall satisfaction of 4.17.
- Only 2% report advanced autonomy across 76% to 100% of systems, while 3% still deploy only chatbots.