To control expenses related to AI, large enterprises are limiting the usage quotas of their employees, opting for smaller models that are less costly, and using routing schedulers to direct only the most complex query requests to the most expensive large models for processing.

The startup behind AI Code Agent and Devin has launched a new initiative. The company revealed that by measuring the efficiency of different engineers in using AI, enterprises can allocate more resources to those employees who are most efficient.

Cognition Chief Technology Officer Calvin Giroux stated: "Users' usage limits should not be fixed, but should be dynamically adjusted based on usage efficiency and the value generated."

Cognition can observe the coding sessions of engineers using Devin to estimate the equivalent man-hours required for humans to complete the same amount of work. By then dividing these man-hours by the cost incurred during this session, we can calculate the efficiency index for each engineer.

When an engineer reaches the usage limit, if their efficiency rating is "good" or "met the standard," the system will automatically increase their quota for them.

Giru said, "It's similar to the logic of a credit card. You start with a fixed credit limit, and as you prove that you can use that credit responsibly and repay on time, the limit will gradually increase."

Cognition is billed based on usage, which seems to contradict its purpose of helping customers control the consumption of Token. However, the company believes that in the highly competitive market of AI coding tools, gaining customer trust and establishing long-term customer loyalty is far more valuable than maximizing short-term customer spending.

Jilu stated that unrestrained use of Token ( tokenmaxxing ) 'is not a sustainable long-term approach. Many of our clients are large listed companies that must be accountable to their shareholders and the public market regarding the efficiency of the tools they use.'

Cognition will be launched simultaneously with Devin Coach. This AI system will provide feedback based on engineers' prompts to guide users in improving their efficiency. The assistant can identify 20 to 30 of the most common errors, such as vague instruction descriptions or users selecting models with excessively high costs for tasks.

For example, if a user packs multiple independent tasks into one command, Coach will recognize that these tasks can be processed in parallel, which is faster, and suggest splitting them into multiple sessions.

Coach can also monitor whether employees use Devin to handle personal matters such as vacation planning. Jilu stated that this feature originated from a concern raised by a customer who noticed that some employees were using Devin for non-work-related tasks, although the proportion of such usage is "relatively low."