By Rajesh Beri | July 16, 2026
Here's a question that should make every enterprise AI leader uncomfortable: How many of your deployed "AI agents" can actually complete a multi-step workflow without human intervention?
If you're like the majority of enterprises surveyed in VentureBeat's latest Pulse Research, the answer is: almost none.
Across 101 enterprises with 100+ employees, 71% admit that a quarter or fewer of their deployed "agents" are true multi-step orchestrated workflows. Only 10% have crossed the halfway mark. The rest? Single-prompt chatbot wrappers dressed up with agent branding — a phenomenon VentureBeat calls "the chatbot trap."
This finding arrives as enterprises pour unprecedented money into agent orchestration platforms. Anthropic's Claude leads at 40% primary platform adoption — more than double any rival. Microsoft trails at 18%, OpenAI at 13%. The platforms are being adopted. The infrastructure is being built. The agents, however, largely don't exist yet.
And the clock is ticking. Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 due to escalating costs, unclear business value, or inadequate risk controls. The firm estimates only about 130 of the thousands of agentic AI vendors are real — the rest are engaging in "agent washing," rebranding existing chatbots, RPA bots, and AI assistants without any substantial agentic capabilities.
The enterprise AI world has a deployment problem, not a platform problem. And most organizations don't even realize they're stuck.
The $2.5 Trillion Mismatch
The scale of the ambition-reality gap becomes staggering when you follow the money.
Global AI spending will hit $2.52 trillion in 2026 — a 44% increase year-over-year, according to Gartner. Enterprises will more than double their spending on generative AI models and AI agents, adding $6 billion in spending this year alone. Data center spending will grow 55.8%, surpassing $788 billion. The four largest hyperscalers have collectively committed $725 billion to AI infrastructure in 2026, a 77% surge from $410 billion in 2025.
Against this backdrop, the VentureBeat data is damning:
- 62% of enterprises say only 1–25% of their "agents" do true orchestration — most are basic assistants
- 9% say literally 0% — every single deployment is a chatbot or prompt wrapper
- Only 3% report 76–100% of their agents running as advanced, largely autonomous systems
The survey's methodology adds credibility to these self-reported numbers. The sample includes 101 respondents from organizations with 100+ employees, drawn from a single June 2026 wave. Eighty-one percent are recommenders, influencers, or final decision-makers for AI solutions. CIOs, CTOs, and CISOs make up 13% of respondents. These aren't junior engineers inflating their team's capabilities — these are senior leaders admitting the gap exists.
The trap is not evenly distributed. Splitting the sample by organization size, 77% of smaller enterprises say a quarter or fewer of their agents do true multi-step work, compared to 62% of larger ones. The mid-market is running the least mature agents — and, as we'll see, on the least instrumented budgets.
Why Enterprises Chose Their Platforms — and Why They're Already Planning to Leave
The VentureBeat data reveals a paradox at the heart of enterprise agent strategy: organizations are consolidating rapidly onto model-provider platforms while simultaneously planning to escape them.
Platform market share (primary agent orchestration platform):
| Platform | Share | Notes |
|---|---|---|
| Anthropic Claude Platform | 40% | More than 2x any rival |
| Microsoft AI Foundry / Copilot Studio | 18% | Strong enterprise integrations |
| OpenAI Agents SDK / Responses API | 13% | Developer momentum |
| Google Enterprise Agent Platform | 8% | Plus 2% on Amazon Bedrock |
| LangChain / LangGraph | 6% | Open frameworks marginal |
| Custom in-house | 5% | Growing category |
| Not orchestrating yet | 3% | Declining fast |
Anthropic's dominance mirrors what the survey calls "model gravity" — the single largest factor driving platform selection at 21%. Enterprises pick the orchestration environment closest to the frontier model they've standardized on. Anthropic's Claude leads because enterprises that want the best base model take the orchestration layer that ships with it.
But the second-tier factors tell a different story. Flexibility across models and tools (17%) and ease of development (17%) rank nearly as high — signals that enterprises want optionality even as they consolidate. Security and permissions (14%) and total cost of ownership (11%) round out the buying logic. Raw performance (latency/memory) sits dead last at 4%.
The most revealing finding: 96% of respondents plan to change their orchestration approach within the next 12 months. Satisfaction scores sit at 3.94 out of 5 — provisional acceptance, not endorsement. Enterprises tolerate these platforms more than they love them.
Their planned moves cluster into three near-tied priorities: increasing investment in custom, in-house orchestration control planes (25%), standardizing on a single centralized framework (24%), and expanding agents from sandbox into production (23%). The message: fewer frameworks, more production exposure, and more ownership of the control layer.
The Hybrid Control Plane: Enterprise AI's Architectural Bet
The most architecturally significant finding in the VentureBeat research is how enterprises expect to structure agent control by end of 2026.
51% expect a hybrid control plane — combining provider-native orchestration with external control layers they own. Only 6% plan to hand full control to a provider-managed agent service. Every architecture that keeps control at least partly outside the provider sums to 88%.
The reason is straightforward: vendor lock-in is the #1 risk enterprises fear from provider-resident control, cited by 35% of respondents. Security and permissioning limitations follow at 28%, with inflexibility across models and tools at 21%.
This marks a shift from earlier data. In VentureBeat's April–May wave (n=145), only 34% expected a hybrid control plane, and 12% expected full provider control. By June, hybrid jumped to 51% and full provider control dropped to 6%. The direction is unambiguous: enterprises are moving toward keeping control.
The concern has matured too. In the April–May wave, security and permissioning limitations led at 32%, with lock-in second at 24%. By June, the two traded places. The worry has evolved from "can these platforms be secured?" to "can they be replaced?"
This echoes what I wrote about model-agnostic architecture earlier this month: the AI model market is fragmenting so fast that vendor lock-in has become the single highest-risk decision in enterprise technology. The hybrid control plane is the architectural hedge enterprises are building to protect against exactly that risk.
Framework #1: The 5-Level Agent Maturity Self-Assessment
Based on the VentureBeat data, Gartner's agentic AI hype cycle research, and IDC findings that 88% of AI agent POCs never reach production, I've developed a framework for enterprises to honestly assess where their "agents" actually sit.
Level 1: Prompt Wrapper (9% of enterprises report all deployments here)
Characteristics: Single-prompt, stateless interactions. No tool use, no memory, no multi-step capability. Essentially a branded chatbot interface over an LLM API.
Indicators:
- Agent responds to one query at a time with no context carryover
- No API integrations or external tool calls
- No workflow state management
- User must manually chain steps together
- "Agent" is functionally equivalent to a ChatGPT conversation
Enterprise reality: This is where Gartner's "agent washing" lives. If your vendor rebranded an existing chatbot or FAQ bot as an "agent" without adding multi-step capability, tool use, or autonomous decision-making — you're here.
Level 2: Stateful Assistant (Most enterprises cluster between Levels 1-2)
Characteristics: Multi-turn conversations with memory. May include basic retrieval-augmented generation (RAG). Can maintain context across a session but cannot independently take actions.
Indicators:
- Remembers conversation history within a session
- Can pull from a knowledge base or document store
- Still fundamentally reactive — waits for user input at each step
- No autonomous decision-making or branching logic
- Cannot write to external systems
Level 3: Tool-Using Agent
Characteristics: Can call external APIs, query databases, or trigger actions. Makes decisions about which tools to use based on user intent. Beginning of genuine agentic capability.
Indicators:
- Integrates with 2+ external systems (CRM, ERP, databases)
- Makes tool selection decisions autonomously
- Can read from AND write to external systems
- Handles error states and retries
- Still operates within a single task context
Level 4: Multi-Step Orchestrated Workflow (Only 10% of enterprises past halfway)
Characteristics: Completes complex, multi-step tasks with branching logic. Maintains state across steps. Can handle exceptions and route to human review when confidence is low.
Indicators:
- Executes workflows with 5+ sequential or parallel steps
- Manages state across minutes or hours of execution
- Includes conditional branching and exception handling
- Has defined escalation paths and human-in-the-loop checkpoints
- Measurable task completion reliability >90%
Level 5: Multi-Agent Autonomous System (3% of enterprises report this)
Characteristics: Multiple specialized agents coordinate to accomplish complex goals. Supervisor agents delegate to specialist agents. System operates with minimal human intervention.
Indicators:
- 3+ specialized agents working in coordination
- Supervisor/orchestrator pattern with delegation logic
- Cross-agent state management and communication
- Autonomous goal decomposition and planning
- Real-time cost control and circuit breakers
How to Score Your Portfolio
For every deployed "agent" in your organization, assign a level. Then calculate your Agent Maturity Index (AMI):
AMI = (Sum of all agent levels) / (Number of agents × 5) × 100
- AMI below 30: You're in the chatbot trap. Most of your agents are Levels 1-2. Redirect budget from new agent deployments to upgrading existing ones.
- AMI 30-50: Transitioning. You have some genuine agents but most are still assistants. Focus on moving Level 2-3 agents to Level 4 before launching new ones.
- AMI 50-70: Production-grade. A meaningful portion of your portfolio does real multi-step work. Invest in monitoring, cost control, and the hybrid control plane.
- AMI above 70: Advanced. You're in the 3% running largely autonomous systems. Your challenge is fiscal control and governance, not capability.
The Token Burn Problem Nobody's Solving
Even enterprises that have moved beyond chatbot wrappers face a different trap: they can't control what their agents cost.
The VentureBeat survey finds:
- 32% rely entirely on native platform controls — built-in budget caps and throttling
- 27% have reactive monitoring only — no real-time kill switch for runaway agents
- 23% build custom gateway middleware to intercept runaway runs
- 19% use dynamic routing to offload heavy work to cheaper models
More than a quarter of enterprises have no real-time way to stop an agent before a budget-breaking bill arrives. They learn about it from the logs afterward.
This isn't hypothetical. Portal26 reports that Microsoft began canceling internal Claude Code licenses after runaway token bills made consumption unsustainable at scale. Yarken research found that token unit prices are falling 90% by 2030, but token consumption is growing 50 to 100 times faster — meaning the total bill keeps rising even as per-token costs drop.
BCG published new guidance this week on enterprise AI token cost management, recommending governance frameworks, RoAI (Return on AI) measurement, and workflow-level attribution. EY's agentic AI token cost analysis highlights the "governance burden" — the incremental investment required to keep agents safe, auditable, and compliant through guardrails, cyber protections, and human-in-the-loop reviews.
A company-size split makes the problem worse: roughly one in three enterprises under 2,500 employees exercises only reactive control of agent spend, versus 20% of larger enterprises. The mid-market is running the least mature agents on the least instrumented budgets.
Framework #2: Agent Orchestration Platform Decision Matrix
Based on the VentureBeat findings and the enterprise priorities they reveal, here is a decision framework for evaluating your agent orchestration platform.
Step 1: Score Your Current Platform (1-5 Scale)
| Criterion | Weight | What to Evaluate |
|---|---|---|
| Model Quality | 25% | Does the platform's native model match your reliability and capability needs? Task completion rate on your specific use cases? |
| Flexibility | 20% | Can you swap models without rewriting agents? Multi-provider support? Open standards (MCP, A2A)? |
| Security & Permissions | 20% | Granular tool-level permissions? SOC 2/HIPAA compliance? Data residency controls? Audit logging? |
| Total Cost of Ownership | 15% | Token pricing + platform fees + engineering time. Real-time cost attribution? Budget caps? Kill switches? |
| Lock-in Risk | 10% | Portable agent definitions? Standard APIs? Can you migrate in <90 days? |
| Ecosystem Maturity | 10% | Pre-built integrations? Community? Enterprise support SLAs? |
Step 2: Map Your Control Plane Architecture
Based on the VentureBeat finding that 88% of enterprises want control outside the provider, determine which pattern fits:
Pattern A: Provider-Native (6% of enterprises)
- All orchestration runs inside the model provider's platform
- Lowest engineering overhead
- Highest lock-in risk
- Best for: Small teams, single-vendor strategy, non-critical workloads
Pattern B: Hybrid Control (51% of enterprises — recommended)
- Model provider handles execution and runtime
- You own the control plane: routing, permissions, cost governance, monitoring
- Custom gateway sits between your applications and the provider
- Best for: Most enterprises. Balances capability with portability
Pattern C: Abstraction Layer (15% of enterprises)
- External orchestration platform abstracts away model providers entirely
- Maximum portability, maximum engineering overhead
- Best for: Multi-cloud mandates, regulated industries with vendor concentration limits
Pattern D: Full Custom (22% of enterprises)
- Build your own orchestration from scratch
- Maximum control, maximum cost
- Best for: Organizations with 50+ agents in production and dedicated platform engineering teams
Step 3: Build Your Fiscal Control Stack
Based on the VentureBeat finding that 27% lack any real-time cost control:
- Token-level attribution: Every agent call tagged by team, project, use case, and model
- Budget caps per agent: Hard limits that halt execution, not just alert
- Circuit breakers: Automatic kill switches for recursive loops or anomalous consumption patterns
- Model routing: Tiered routing that sends simple tasks to cheaper models (GPT-4o-mini, Haiku) and reserves frontier models for complex reasoning
- Weekly cost review: Dashboard showing cost-per-task-completed, not just total token spend
The Ode Signal: Even AI Labs Know the Gap Is Real
The timing of the chatbot trap data is notable. Just yesterday, Anthropic, Blackstone, and Hellman & Friedman launched Ode — a $1.5 billion enterprise AI implementation company. Ode deploys roughly 100 specialized forward-deployed engineers directly into enterprises, built on the premise that the next trillion-dollar AI business is implementation, not models.
This is an AI model company admitting — with a $1.5 billion bet — that its own customers can't close the gap between buying Claude and actually deploying agents that work. The deployment war I covered last week documented $9 billion flowing into forward-deployed engineering units in just 60 days. TCS alone plans 8,900 forward-deployed AI engineers.
The pattern is unmistakable: the AI industry has realized that selling frontier models to enterprises produces chatbot wrappers, not agents. Moving from Level 1-2 to Level 4-5 requires human engineers embedded on-site, rebuilding workflows from the inside.
What to Do Monday Morning
If the VentureBeat data is even directionally correct — and its methodology is solid enough to trust — most enterprise AI leaders need to have an uncomfortable conversation with their teams this week.
1. Audit your agent portfolio honestly. Use the 5-level maturity assessment above. Count how many of your deployed "agents" can actually complete a multi-step workflow without a human manually advancing the process. If the answer is fewer than 25%, you're in the majority — but that doesn't mean you're safe.
2. Stop launching new Level 1-2 agents. Every new chatbot wrapper deployed under the "agent" label obscures the real maturity gap and dilutes your team's focus. Redirect that energy to upgrading your 3-5 most promising existing deployments to Level 4.
3. Build your fiscal control stack before you scale. If you're in the 27% without a real-time kill switch for runaway agents, fix that this quarter. It doesn't matter how sophisticated your orchestration platform is if a recursive agent loop can burn $50,000 before anyone notices.
4. Plan for the hybrid control plane. If you're building on a model-provider platform (and 80% of enterprises are), start designing the control layer you own — routing, permissions, cost governance, monitoring — before production scale forces the issue. The 88% of enterprises that want control outside the provider aren't being paranoid. They're being realistic about a market that changes model providers faster than enterprises change architectures.
5. Measure agent maturity, not agent count. Your board doesn't need to hear that you deployed 50 AI agents this quarter. They need to hear that your AMI moved from 25 to 40, that your Level 4 agents now handle 30% of customer escalations autonomously, and that your cost-per-task-completed dropped 15%. Maturity metrics, not vanity metrics.
The Bottom Line
The enterprise AI industry has a naming problem that's become a strategy problem. When 71% of deployed "agents" are chatbot wrappers, when 96% of organizations plan to change their orchestration approach within a year, and when even Anthropic bets $1.5 billion that implementation — not models — is the bottleneck, the message is clear: the orchestration layer is being built well ahead of the orchestrated portfolio it's meant to run.
The chatbot trap isn't permanent. Enterprises are investing in the right infrastructure — workflow tooling leads spending at 34%, security and permissions at 25%, scaling infrastructure at 20%. The platforms are consolidating. The control plane architecture is maturing. The fiscal governance layer is emerging.
But the gap between what enterprises call their AI agents and what those agents actually do is the largest unacknowledged risk in enterprise AI today. The companies that close it — by upgrading real capability, not relabeling chatbots — will define the next era of AI-driven business. The rest will spend the next 18 months explaining to their boards why $2.5 trillion in industry AI spending produced a portfolio of expensive chatbots wearing agent costumes.
