From Autonomy Economics, the reference entry for the discipline
Autonomy does not require full deployment to trigger repricing; markets respond to directionality. Autonomous capability does not need to be fully deployed for its economic implications to matter; it only needs to become credible in commercially meaningful activities. Exposure appears first in structured, decomposable work tied to HBP, in interaction-dependent models tied to HBI, and in models where capability scarcity is beginning to weaken through CSD.
Margin pressure often precedes demand erosion as the perceived value of human-bounded tasks declines. Commercial terms often adjust slowly, but the basis of value can weaken earlier. In that setting, markets may respond to direction rather than completeness.
The cost of model inference falls. Token prices decline. Context windows widen. Open-weight models reduce dependency on proprietary APIs. At the same time, usage rises. Firms automate more routines, run agents continuously, and embed decision logic across surfaces. This mirrors the pattern seen in cloud computing: unit cost falls while aggregate spend grows. The key change is that execution cost, interaction efficiency, and access to capability become more measurable and, in some settings, easier to compare across firms and models.
Howard Marks, co-founder of Oaktree Capital Management and a widely read investor known for his market memos, frames the issue in practical terms. In his February 2026 memo AI Hurtles Ahead, written as a follow-up to his December 2025 memo Is It a Bubble?, he argues that the central economic question is whether AI does the work, and that whether it truly understands matters less. He also stresses that the commercial outcome depends on whether organisations know how to organise and direct it effectively (Marks, 2026).[15]
The relevance here lies in execution. A system does not need to replicate human understanding in full to affect the economics of a revenue line. It only needs to carry forward commercially meaningful stages of activity, mediate commercially meaningful parts of interaction, or broaden access to capability with enough reliability for the commercial basis of value to come under pressure. At the same time, firms that structure activity effectively can use the same capability to expand what they deliver.
Clients observe faster turnaround, reduced staffing, and automated progression across stages of activity. They begin to question fee structures built on human progression, human-bounded interaction, or scarce capability access. An early signal appeared when KPMG pressed its auditor Grant Thornton UK for a reduction in audit fees, arguing that AI-enabled efficiency gains should be reflected in lower prices, resulting in a fee reduction from $416,000 to $357,000 for the 2025 audit (Foley, 2026).[23] Bloomberg reported a related pattern at PwC, where Dan Priest, the firm’s Chief AI Officer, said AI had improved the efficiency of some systems integration work by roughly 30 per cent and that part of that gain had appeared as price reductions to clients (Davalos and Ford, 2025).[24] A change in how part of the underlying activity could be carried forward was enough to affect the commercial basis of the engagement.
Once efficiency becomes measurable and sufficiently comparable across firms, it enters competition more directly. Commercial structures adjust, positioning changes, and automation becomes more expected. In many settings, operational advantage becomes harder to sustain on process alone, even though integration quality, workflow design, and human-system coordination remain important differentiators. As more firms reach similar levels of execution capability, differentiation moves away from how activity is carried forward towards how that capability is organised and applied.
A contemporary illustration appears in Salesforce’s launch of Headless 360 (Salesforce, 2026),[25] which exposes the full platform as APIs, MCP tools, and CLI commands so agents can operate it without the browser. The significance is commercial as well as technical. It suggests that the graphical interface may no longer hold the same place as the primary locus of value in enterprise software. The platform is being repositioned as agent-accessible infrastructure as well as a user-facing application. This is consistent with the broader pattern described here. As execution becomes more measurable and interface dependence weakens, defensibility moves towards data, workflow logic, governance, trust layers, and the orchestration of capability across surfaces. The accompanying move towards consumption-based pricing for Agentforce also reflects the weakening of seat-based assumptions once agents as well as users increasingly mediate execution.