From Token Sales to Result Delivery: How AI Is Now Backed by KPI

After Nvidia CEO Jensen Huang urged firms to count token consumption toward engineers' KPIs at GTC, large tech firms began staging "token‑burning contests"—Meta’s internal competition and Amazon staff inflating Agent usage—to superficially boost AI usage metrics. The article argues that when enterprises cannot measure AI's true value, they default to counting calls and tokens instead of outcomes.

McKinsey data show that 88% of companies regularly use AI in at least one function, yet fewer than 6% achieve "AI high‑performance" status (EBIT improvement >5%). Massive budgets, compute, and engineering time often yield only a busy‑looking bill, raising the core question: is the spend worth it?

To address this distortion, a new business model—Result‑as‑a‑Service (RaaS)—emerges, shifting from selling tools to selling verified business results. The article highlights two pioneers: Silicon Valley’s Sierra and China’s Lingxi Technology.

Sierra, founded by OpenAI board chair Bret Taylor and former Google exec Clay Bavor, raised $950 million in May, valuing the company at over $15 billion—more than 100 times its $150 million ARR. Sierra’s customers include over 40% of the Fortune 50, demonstrating that RaaS has moved beyond pitch decks to large‑scale commercial validation.

Lingxi focuses on insurance and automotive sales, where its AI agents handle the entire sales funnel—from initial interest detection to final contract signing. Since 2019, Lingxi turned a negative margin into company‑wide profitability in 2024 and achieved scalable profit and positive cash flow in 2025. A leading insurer that integrated Lingxi’s agent reported over ¥20 billion in new premium, a result that would normally require an 800‑to‑1,000‑person sales team.

Both firms share a technical premise: AI must be responsible for the end result, not just generate text. They augment large models with execution, memory, and evaluation layers, and continuously retrain the system in production. However, their focus diverges—Sierra prioritises correctness for customer‑experience agents, deploying a supervisory model that orchestrates more than fifteen heterogeneous models, while Lingxi must optimise conversion, requiring a reward‑punishment mechanism and causal learning.

Lingxi tackles attribution by collecting "causal‑complete" full‑stack data: user state, page views, AI strategy, and subsequent feedback, building a domain‑specific causal knowledge graph that constrains model outputs. Post‑training reinforces behaviours that lead to conversions and penalises ineffective tactics, supplemented by counterfactual reasoning that evaluates "what‑if" scenarios for missed sales.

This feedback loop creates a self‑evolving system: because Lingxi charges for results, it receives direct outcome signals, enabling continuous evaluation, attribution, and strategy optimisation. The accumulated knowledge—real‑world rules, success cases, and failure lessons—feeds back into the model and knowledge graph, forming a reinforcing flywheel of data, AI‑encoded know‑how, and iterative improvement.

RaaS’s moat deepens as foundational models improve; the service layer—result delivery, risk transfer, and embedded business knowledge—cannot be erased by model upgrades. As customers grow accustomed to paying for outcomes, they begin demanding the same from all AI vendors, reshaping competitive dynamics.

Finally, Morgan Stanley’s 2026 report frames AI as the sixth technological revolution, noting that historically, infrastructure sellers profit first while application‑layer value accrues later. The article concludes that firms willing to shoulder result risk and turn AI into a KPI‑backed digital workforce will capture the next decade’s market, echoing past patterns where the true value lies in turning tools into finished work.