LayerX promoted heavy generative AI use under “Token Maxxing” in 2025, and is shifting in 2026 to “Token Economics,” which tracks costs and outcomes while selecting tasks and models. At Bet AI Day 2026 on September 3, CEO Yosuke Fukushima compared the change to self-driving and said, “Last year was Level 3; this year is Level 4.” As AI takes on routine work while people step in for exceptions, companies need systems to manage usage and costs. In an interview on September 16, 2026, Shunsuke Tominari, product manager for LayerX’s “AI Token Advisor,” explained the company’s approach.

“Token Maxxing” in 2025 encouraged employees to expand their AI use so they could learn what it can and cannot do. As usage and costs rise, the next step is to understand usage and optimize models and methods for each task. Tominari described setting a time period or budget for real-world use, then moving to management based on the results. Budgets can be set as a monthly amount, in terms of the labor cost of a certain number of people, or around using AI until a specific project produces results. The key measure is not just token volume, but why AI was used.

Management dashboards and APIs can show models, token counts and spending, but those figures alone do not reveal whether AI is being used well. Tominari gave the example of someone asking a high-performance model, “What will the weather be tomorrow?” An employee spending ¥100,000 a month is not necessarily using AI more effectively than one spending ¥10,000, and simply cutting costs is not always the right goal. Companies need to connect who used which model for what work, what they asked it to do and what outcome they got. They can also analyze prompts and tasks to check whether a cheaper model could deliver the same result.

Choosing models is another part of cost management. People unfamiliar with AI may choose the latest high-performance model for reassurance, although tasks such as text classification and information extraction may not require top-tier performance. At Bet AI Day 2026, LayerX said it was building an AI gateway that assesses requests behind a single API key and routes them to suitable models. The system prompts users to improve when prompt caching is not working and retrieves data directly through SQL when an LLM call is unnecessary. LayerX is building processes for model selection, caching and data retrieval, rather than relying only on appeals for employees to save money.

When AI use is left to individuals, prompts, responses and knowledge that produced results can remain on personal computers or within individual work habits, creating gaps in usage and skill across the organization. Companies can also struggle to see where costs arise and which work the AI supported. LayerX’s theme for the coming year is “from personal AI to organizational AI.” It plans to make shared company knowledge available to AI, align security standards and work procedures, and track usage and costs across the organization.

As “Ambient Agents”—systems that start work in response to events such as incoming email, schedule changes or data updates without waiting for a person’s instruction—spread, employee actions will no longer track token consumption. Malicious emails or large batches of work could trigger more AI use than expected, Tominari said, making usage guardrails and shutdown functions important. Companies also need to design budgets and stop conditions in advance, and forecast costs based on average and peak usage as well as seasonal changes.

Managing usage appropriately does not by itself mean an AI investment has succeeded. LayerX still has no clear answer for measuring return on investment (ROI). Tominari said a simple management dashboard or web app can be built with AI but go unused a month later: “Being able to build something” and “creating value” are different. In sales, if AI prepares meeting notes, next steps and follow-up emails so staff can spend more time negotiating and building customer relationships, that saved time can also count as an investment return.

Tominari said the ultimate measure should be whether the requested work was completed, rather than the number of tokens or the model used. The proposed approach breaks down human work and compares its cost when done by a person with the AI token cost. Since not every task involving human conversation or physical action can be replaced by AI, the initial focus is on digitally completed work, measuring completion counts and costs. Corporate metrics are expected to include completed-work counts, cost per task and comparisons with human costs, alongside usage, models and spending.

LayerX promoted heavy generative AI use under “Token Maxxing” in 2025, and is shifting in 2026 to “Token Economics,” which tracks costs and outcomes while selecting tasks and models. At Bet AI Day 2026 on September 3, CEO Yosuke Fukushima compared the change to self-driving and said, “Last year was Level 3; this year is Level 4.” As AI takes on routine work while people step in for exceptions, companies need systems to manage usage and costs. In an interview on September 16, 2026, Shunsuke Tominari, product manager for LayerX’s “AI Token Advisor,” explained the company’s approach.

“Token Maxxing” in 2025 encouraged employees to expand their AI use so they could learn what it can and cannot do. As usage and costs rise, the next step is to understand usage and optimize models and methods for each task. Tominari described setting a time period or budget for real-world use, then moving to management based on the results. Budgets can be set as a monthly amount, in terms of the labor cost of a certain number of people, or around using AI until a specific project produces results. The key measure is not just token volume, but why AI was used.

Management dashboards and APIs can show models, token counts and spending, but those figures alone do not reveal whether AI is being used well. Tominari gave the example of someone asking a high-performance model, “What will the weather be tomorrow?” An employee spending ¥100,000 a month is not necessarily using AI more effectively than one spending ¥10,000, and simply cutting costs is not always the right goal. Companies need to connect who used which model for what work, what they asked it to do and what outcome they got. They can also analyze prompts and tasks to check whether a cheaper model could deliver the same result.

Choosing models is another part of cost management. People unfamiliar with AI may choose the latest high-performance model for reassurance, although tasks such as text classification and information extraction may not require top-tier performance. At Bet AI Day 2026, LayerX said it was building an AI gateway that assesses requests behind a single API key and routes them to suitable models. The system prompts users to improve when prompt caching is not working and retrieves data directly through SQL when an LLM call is unnecessary. LayerX is building processes for model selection, caching and data retrieval, rather than relying only on appeals for employees to save money.

When AI use is left to individuals, prompts, responses and knowledge that produced results can remain on personal computers or within individual work habits, creating gaps in usage and skill across the organization. Companies can also struggle to see where costs arise and which work the AI supported. LayerX’s theme for the coming year is “from personal AI to organizational AI.” It plans to make shared company knowledge available to AI, align security standards and work procedures, and track usage and costs across the organization.

As “Ambient Agents”—systems that start work in response to events such as incoming email, schedule changes or data updates without waiting for a person’s instruction—spread, employee actions will no longer track token consumption. Malicious emails or large batches of work could trigger more AI use than expected, Tominari said, making usage guardrails and shutdown functions important. Companies also need to design budgets and stop conditions in advance, and forecast costs based on average and peak usage as well as seasonal changes.

Managing usage appropriately does not by itself mean an AI investment has succeeded. LayerX still has no clear answer for measuring return on investment (ROI). Tominari said a simple management dashboard or web app can be built with AI but go unused a month later: “Being able to build something” and “creating value” are different. In sales, if AI prepares meeting notes, next steps and follow-up emails so staff can spend more time negotiating and building customer relationships, that saved time can also count as an investment return.

Tominari said the ultimate measure should be whether the requested work was completed, rather than the number of tokens or the model used. The proposed approach breaks down human work and compares its cost when done by a person with the AI token cost. Since not every task involving human conversation or physical action can be replaced by AI, the initial focus is on digitally completed work, measuring completion counts and costs. Corporate metrics are expected to include completed-work counts, cost per task and comparisons with human costs, alongside usage, models and spending.