What does AI really cost when a cheap model fails, retries the task, hands the work to a person, and leaves the customer returning with the same problem? In this episode of Tech Talks Daily, I welcome back Dan O'Connell, CEO of Front, to discuss AI tokenomics and why token consumption is a poor substitute for business value. Our previous conversation focused on Bring Your Own Agent and the need for companies to choose the agents that fit their customer experience. This time, we turn to the economics behind those decisions. Dan describes tokens as the new cloud bill. As businesses deploy agents across a growing number of customer interactions, usage-based charges can rise quickly. That number may show activity, but it does not explain whether the work created a useful result. A high token count can represent productive automation, repeated attempts, unnecessary complexity, or a task that eventually required human intervention. His preferred measure is the total cost per successful outcome. In customer support, that means looking beyond the model's headline price and accounting for every part of the journey. Model retries, agentic retries, human review, handoffs, coordination, response time, and the final customer result all affect the true cost to serve. The idea matters because lower model prices can be misleading. A cheaper model may save money on the first attempt while increasing the number of retries or transfers. A larger model may cost extra per request while completing a difficult task with fewer steps. Leaders therefore need to compare operating results rather than price lists. Dan explains how model routing can help. Simpler tasks may be directed to smaller models, while difficult work can be sent to models with greater capability. The choice depends on the task and the tradeoffs the business is willing to accept. Dan frames those tradeoffs through three dimensions: quality, speed, and cost. In many cases, a company can optimize two while accepting a compromise on the third. That does not mean every basic task should automatically use the cheapest option. A slower response may be acceptable for internal analysis but frustrating during a live customer conversation. A small reduction in accuracy may carry little risk in one workflow and create expensive repeat contacts in another. The orchestration layer needs rules informed by the customer experience, the complexity of the request, and the cost of failure. Customer service gives companies a useful environment for measuring those results because many of the existing measurements already connect work with outcomes. Cost to serve, time to resolution, deflection, accuracy, customer satisfaction, and repeat contact rates can all reveal whether AI is helping. Dan argues that agent performance should be evaluated alongside human performance using comparable standards. Resolution is particularly important. If an AI system provides an answer and the customer returns an hour, a day, or a week later with the same issue, the task was not resolved. The first interaction may look successful in an activity report while creating extra work, frustration, and cost later. Measuring only initial response or deflection can therefore reward the wrong behavior. The conversation also considers what automation means for the people working around customers. Front's research, as discussed in the interview, found that 54 percent of surveyed leaders see issue resolution as support's main purpose today, while 40 percent believe it should remain the main purpose. Dan interprets that difference as an opportunity to automate lower-level work while allowing employees to spend additional time building customer relationships. The research also reports that 75 percent expect support, customer success, and account management to become increasingly unified. Dan believes this will affect team structures and the platforms companies choose. Customers do not think in internal departmental boundaries. They expect one company to remember their history and coordinate the response, whether the work passes between an agent, a support specialist, customer success, or account management. Shared context and coordination therefore become part of the AI cost equation. A handoff without the customer's history can force someone to repeat the problem, reducing satisfaction and extending resolution time. The technical cost of the interaction may appear low while the operational and customer costs continue to grow. Dan recommends treating AI autonomy as a dial. Companies can begin with a person reviewing recommendations, maintain audit logs showing why the system acted, and increase autonomy as evidence builds. The same outcome measurements applied to employees should also apply to agents, including whether the customer received an accurate answer, achieved a lasting resolution, and had a good experience. The larger message is that AI economics cannot be separated from service design. A model, router, agent, or platform earns its place by contributing to an outcome the business and customer value. Usage may help explain the bill, but it cannot prove the result. Should companies continue celebrating rising AI usage, or start asking what each successful outcome cost and whether the customer would agree it was successful? Listen to the episode and share your thoughts.