Enterprise AI Is Learning To Charge For Work, And Owning The Outcomes Becomes The Contest

Enterprise AI is shifting from a cost per token model to a focus on cost per successful task. OpenAI's CFO proposes "useful intelligence per dollar" as the new metric, emphasizing work accomplished over usage. This change is expected to be complete by 2026, with enterprises utilizing the cheapest capable models for routine tasks and reserving more powerful models for more demanding tasks.

Token counting, once the primary measure, is no longer sufficient, as an agent that closes deals while consuming fewer tokens will be more valuable.

Companies such as Sierra and Salesforce are already adopting this new pricing model. Sierra charges a set rate per resolved conversation, while Salesforce reports "Agentic Work Units" based on work performed rather than occupied seats. Analysts at Constellation Research note that value-based pricing historically benefits vendors more than customers. The key question now is whether the outcome is achieved successfully when the billing occurs.

As enterprise AI pricing moves towards outcome-based models, the product evolves from being the user interface to becoming the agent that delivers the outcome. AI founders view the valuable object as the harness around the AI system, consisting of skills, documentation, and rules that enable the best experts to run workflows and extract value. In this perspective, the future application of AI lies in the harness itself, not just the API or dashboard.

One example of this new approach is Claudeforce, announced in August 2026, which integrates Anthropic's Claude as the default reasoning engine across Salesforce. This harness provides 37 prebuilt sales skills, allowing users to perform tasks like pipeline reviews or deal preparations without even opening the Salesforce app UI. Salesforce reported that Agentforce's annual recurring revenue had grown to over $1.5 billion, a 240% increase year-on-year, and the CEO emphasized that the software apocalypse talk should cease.

The possibility of one universal harness that covers the entire enterprise faces a challenge that has persisted for decades: centralizing an organization's knowledge in a single place has proven difficult at scale. Each system of record excels in a specific domain but stops there. Salesforce owns sales, SAP manages procurement, Workday handles HR, and none of them have become the company's central brain.

Claudeforce is designed to be an agentic expression of the CRM, not the broader enterprise, and this trend is likely to extend to other systems of record that employ their own agents.

Organizational boundaries, rather than technical limitations, will determine the extent to which enterprise AI can be integrated. Companies rely on their existing org charts, and AI is currently bolted onto individual silos instead of reworking the cross-domain workflows. The universal enterprise harness will require a significant organizational reorganization, which is why the near-term winners will be those that successfully own the harness and deliver outcomes within a bounded, high-value domain.

This narrower problem offers a more achievable path to success for new entrants while incumbents continue to reorganize around agents.