AI on Deterministic Rails

Narrative

The dominant financial-press theme from early 2025 through June 2026 is a sharpening gap between AI adoption breadth and value depth. McKinsey's 2025 State of AI survey found that 88 percent of organisations used AI in at least one function, yet only 39 percent reported any EBIT impact and just 23 percent were scaling agentic systems beyond one or two functions. BCG's parallel analysis found that 60 percent of enterprises generate no material value despite investment, and only 5 percent are creating substantial value at scale. The MIT-reported figure — widely cited in Fortune and Axios — that 95 percent of generative AI pilots are stalling reflects not model failure but governance absence: no clear owner, no workflow redesign, no production-grade data.

The token-cost shock of early 2026 crystallised the enterprise economics problem. Uber exhausted its full 2026 AI tools budget in four months after rolling out Anthropic's Claude Code to roughly 5,000 engineers; the company's Operations chief admitted there was no clear link between token consumption and useful consumer features. Microsoft cancelled most of its internal Claude Code licences for its Experiences and Devices division in May 2026, steering developers back to GitHub Copilot CLI. Goldman Sachs Research, in its May 2026 'Decoding the Agentic Economy' report, projected that agentic AI will drive a 24-fold increase in token consumption by 2030, reaching 120 quadrillion tokens per month, while also forecasting 60–70 percent annual declines in per-token compute cost. The divergence — falling unit costs, exploding total spend — is the central tension: agentic loops can consume 1,000 times the tokens of a single chatbot query, and budgets set before the Claude Code-driven adoption surge in late 2025 are now structurally obsolete.

Bloomberg Intelligence's June 2026 Generative AI 2026 Outlook, which raised its market forecast to $2.3 trillion by 2032 from $1.8 trillion in its March 2025 edition, identified the shift from training to inference as the primary demand driver and noted that coding agents alone are on track from $1 billion in 2024 to near $100 billion by 2032. The pricing model disruption is already underway: SAP CEO Christian Klein told Bloomberg in March 2026 that the company was shifting to consumption-based pricing; GitHub moved all Copilot plans to usage-based AI Credits billing from June 2026; Bloomberg estimated that subscription-based pricing could decline from 60 percent of software pricing models to 30 percent over the next decade as outcome-based models expand. Incumbent platforms with deep workflow integration — Microsoft, SAP, Workday, ServiceNow — appear structurally better positioned than point-solution SaaS vendors in Bloomberg Intelligence's disruption framework.

The open-weight model shift added a second cost-control lever alongside orchestration. DeepSeek V3, released in December 2024 at a claimed training cost of $6 million, offered inference at roughly 12.5 times the cost advantage over Claude 3.5 Sonnet according to IISS analysis. MindStudio's May 2026 survey of enterprise deployments found that open-weight models including DeepSeek V3, Qwen 3, and Llama 4 Maverick were matching closed frontier models on coding, classification, and structured data extraction tasks, with costs 5–20 times lower and the option of self-hosted private inference for data-sensitive workflows. The harness-over-model thesis gained traction in parallel: practitioners and Bloomberg Intelligence analysts converged on the view that orchestration design — context seeding, subagent delegation, deterministic control flow, error handling — now accounts for more variance in production outcomes than the choice of underlying model.