- Agent-based workflows are five to 30 times more computationally intensive than chatbot queries, according to McKinsey, making architectural choices the primary cost lever as inference spend scales.
- Deloitte’s November 2025 TMT Predictions forecast the autonomous enterprise market growing from $8.5 billion in 2026 to between $35 billion and $45 billion by 2030, a spread Deloitte ties directly to whether orchestration is managed well.
- Gartner projects more than 40% of agentic AI projects will be cancelled by end of 2027, citing unclear value propositions and governance failures as the primary causes.
Multi-agent system adoption among large enterprises quadrupled in a single year, according to KPMG’s Q3 2026 AI Pulse Survey, yet the same survey found roughly 93% of participants blew past their AI budgets. Scale is arriving faster than the controls to manage it.
Deployment Numbers Jump
The KPMG survey found that 62% of companies with $1 billion or more in annual revenue are currently deploying or developing AI agents, up from 53% the previous quarter. Multi-agent systems, where several agents collaborate on a task, reached 25% adoption in Q3 2026, up from 6% across all of 2025.
OpenAI has been building persistent, integrated agent capabilities aimed at enterprise workflows, adding further supply-side pressure on teams still working out their deployment strategies.
The Cost Reality
Per-token inference costs have fallen sharply: Stanford’s AI Index found GPT-3.5-level capability dropped more than 99% in price, from $20 to $0.07 per million tokens, between late 2022 and late 2024 alone. That drop has not translated cleanly into margin. McKinsey research suggests only a minority of companies attribute any positive EBIT impact from their AI programmes, a figure that sits uncomfortably against the deployment acceleration the KPMG survey records.
Architectural choices are the primary lever. Model tiering by task complexity, context optimisation, prompt caching and batch APIs each reduce per-task cost without requiring renegotiated vendor contracts. For teams facing Gartner’s projected fivefold rise in agent inference costs by 2028, that architectural work is not discretionary.
Governance Gaps
Autonomous agents create an access-control problem most enterprises have not resolved. These systems operate across multiple applications, handle sensitive data and make decisions with limited human review, often holding credentials that outlast the specific task they were granted for.
A 2026 Delinea Identity Security Report titled “The AI Enforcement Gap” puts a precise number on the disconnect: 99.7% of IT and security leaders report having a formal policy for AI agent data access and isolation, but those policies are frequently not enforced through technical controls. Policy without enforcement means agents can retain network and data access beyond their intended scope, an attack surface that requires no sophistication to exploit.
Data architecture compounds the problem. Real-time autonomous AI cannot run effectively on legacy, siloed environments, and many organisations carry decades of fragmented infrastructure. Deloitte found that only one in five companies has a mature governance model for autonomous AI agents, a gap that is consistently underestimated at the planning stage. In regulated industries, the distance between policy and technical enforcement is a compliance risk, not just a cost one.
ROI or Cancellation
Gartner projects more than 40% of agentic AI projects will be cancelled by end of 2027, primarily because of unclear value propositions and governance failures as scaling costs mount. McKinsey’s 2026 report makes the same point from the other direction: deploying agents without defined outcomes is itself a cost.
Some deployments have produced concrete returns. JPMorgan’s COIN system automates legal document review and has delivered measurable efficiency gains. KPMG’s Q3 2026 survey found 59% of business leaders expect measurable ROI on their AI investments within 12 months, though a separate KPMG report on broader AI adoption found only 24% of organizations are achieving ROI at scale across multiple use cases, exposing a persistent gap between individual project payoff and enterprise-wide returns.
The $10 billion spread between Deloitte’s lower and upper 2030 market estimates is essentially what the firm assigns to getting orchestration right.



