AI Pricing Shifts from Compute to Outcomes: Fin Charges $0.99 Per Resolution

Customer service platform Fin charges $0.99 per resolved inquiry, while U.S. AI documentation platform Mintlify bills 25 credits per answered question — service companies are increasingly pricing AI for what it accomplishes, not how it operates. In September 2026, Mintlify shifted from token-based pricing to fixed credits per completed task, charging 250 credits when its automation feature updates documentation. Tasks where the AI fails to answer or finds nothing to fix in the documentation aren't charged. This lets customers forecast costs based on the output they receive rather than how much computation happened behind the scenes. Zendesk has likewise adopted billing based on inquiries AI resolves without human agent intervention.

How is AI cost coming down?

When an agent works longer or repeats a task, token usage can rise accordingly, but higher token usage doesn't guarantee better results. What companies should track alongside usage is revenue growth, reduced processing time, and customer issues resolved. Model providers have released lower-cost models, and "model routing" — directing simple questions to smaller, faster models and complex analysis to high-performance models — is also being used.

• OpenAI says its 'GPT-6.1 Sol' delivers performance close to its flagship 'GPT-6 Astra' in agentic coding, computer use, and specialized tasks, while standard input and output token prices are about one-fifth as much. Input costs $2 per million tokens and output $10, compared to $10 and $50 for Astra.
• Anthropic priced 'Claude Opus 5.5' at $4 per million input tokens and $20 per million output tokens — 20% lower than the previous Opus 5, with cache-read pricing for reused input cut by 60%. Anthropic said typical tasks under default settings cost 40% less than with the previous model, emphasizing total task cost over per-token pricing alone.
• Amazon Web Services' (AWS) Bedrock 'Intelligent Prompt Routing' compares expected performance across models within the same family for each request and routes it to the appropriate one. AWS states this can cut costs by up to 30% while maintaining accuracy — a figure provided by AWS itself.
• Samsung Electronics' 'Hybrid AI' processes tasks requiring fast response and privacy protection on-device, while handling large-scale generation and reasoning tasks that need heavy computation in the cloud. Samsung also featured efficient AI as a key research topic at 'Samsung AI Forum 2026,' held on September 30, 2026.

How should AI performance be measured?

A 2025 study published in the Quarterly Journal of Economics (QJE), "Generative AI at Work," analyzed 5,172 customer service agents. Agents using AI assistance resolved 15% more customer issues per hour on average, with effects varying significantly across workers. The study measured productivity not by how often agents queried AI or how much text was generated, but by the number of problems resolved in a given time.

Outcome-based pricing doesn't guarantee effectiveness just because it's labeled "outcome-based." Fin counts a case as resolved not only when a customer confirms satisfaction with the answer, but also when no further help is requested after a response. Cases can also be billed when a predefined procedure is completed and then handed off to a person or another workflow.

How is SK Telecom transforming internal AX?

According to global AX trend data released by SK Telecom on October 28, 2026, the company defines AI agents as work partners collaborating with employees and is pursuing internal AI transformation by jointly designing roles for people and AI. The goal is a "multi-role" approach in which a single employee collaborates with multiple AI agents across planning, development, and design. To support this, the company is changing the following:

• Consolidating previously fragmented internal AI development platforms
• Strengthening integration with business systems
• Overhauling data access policies and security frameworks

In its previously announced 'AX Innovation 2.0,' SK Telecom laid out plans to expand company-wide an 'AX Sandbox' that redesigns existing work processes around AI. The company said a roughly three-month pilot at some units within its AI in-house corporation (CIC) showed potential improvements in planning turnaround time and decision-making speed, though the latest reference materials did not present company-wide productivity figures.

SK Telecom CEO Jeong Jae-heon presented the "J-curve of AX" in a special lecture at Seoul National University on September 1, 2026. The idea is that productivity initially dips due to the burden of learning and redesigning work during early adoption, but can rise sharply once that phase is overcome. CEO Jeong stressed that companies need the execution capability to redesign their business and ways of working from an AI-first perspective.

Implications for AX at Korean Companies

When budgeting for AI, it makes more sense to base calculations on the total cost of completing a given task rather than per-token price lists. Try factoring in the design choices of which model handles which request and where — but also watch whether cost-cutting measures compromise answer accuracy or work quality.

If adopting outcome-based pricing services like Fin, Zendesk, or Mintlify, check before signing whether their definition of "resolved" matches the outcomes your organization actually wants. Whether cases with no follow-up inquiry count as resolved, and whether cases handed off to humans are still billed, can make the same per-case rate deliver very different real-world value.

The domestic cases of SK Telecom and Samsung Electronics featured here addressed organizational structure and processing design before pricing. SK Telecom redefined AI's roles and data access policies, while Samsung Electronics divided processing locations by function. Companies looking to scale internal agents should first determine which department's tasks the AI will handle, how far its data access extends, and at what stage human judgment is required — before focusing on the number of agents. Given the early dip in the J-curve, it also helps to set evaluation timelines in advance so success or failure isn't judged solely on metrics right after rollout.

There's one thing worth checking this week: whether your current AI tool's performance reports count query volume and token usage, or processing time, rework, and resolved cases.

Source: AI Shifts From Adoption to Outcomes: Companies Pursue the 'AX Shift'