Meta is recalibrating the way it pushes artificial intelligence inside its workforce, moving away from directly measuring employees by their AI consumption while simultaneously introducing a powerful new autonomous agent. The change suggests that the company is beginning to distinguish between simply using AI frequently and demonstrating that the technology produces meaningful improvements in productivity, software development, and everyday work.
For much of the generative AI boom, technology companies have been trying to accelerate internal adoption as quickly as possible. Meta has been particularly aggressive in promoting an AI-first culture, encouraging employees to incorporate internal models, coding assistants, and other AI tools into their workflows. But when adoption becomes something that workers believe could influence their performance evaluations, the incentive structure can quickly become distorted.
According to WIRED, Meta has now adjusted its performance-review guidance so that explicit AI-usage classifications are no longer part of the evaluation framework. Employees are instead expected to be judged primarily according to their impact and contributions, regardless of whether those results were produced with extensive AI assistance or through more traditional methods. Meta spokesperson Tracy Clayton told the publication that employee contributions remain the basis for performance evaluations, while engineers have reportedly been informed that AI-adoption dashboards and token consumption will not determine their impact.
The change addresses an unusual side effect of the corporate race to adopt generative AI: employees can begin optimizing for AI activity instead of actual productivity. If workers believe that managers expect them to demonstrate extensive AI usage, generating more prompts and consuming more tokens can become an objective in itself, even when those interactions provide little additional value.
That dynamic reportedly contributed to what became known internally as “tokenmaxxing.” Meta employees could generate large amounts of AI activity by repeatedly interacting with internal systems, and the company even experimented with mechanisms for visualizing adoption. An internal leaderboard reportedly compared AI consumption among employees, with particularly active users receiving playful recognition such as the label “Token Legend,” although the leaderboard was later removed after details about it became public.
The episode illustrates a fundamental measurement problem that many companies will encounter as generative AI becomes embedded in the workplace. AI usage is extremely easy to quantify because organizations can count prompts, tokens, active users, generated code, chatbot conversations, or sessions with coding assistants. Measuring whether those interactions actually improve productivity, however, is considerably more difficult.
A software engineer who generates millions of tokens is not necessarily producing better software than someone who uses AI selectively to solve difficult problems. The same applies outside engineering: an employee who constantly interacts with an AI assistant may still deliver less business value than someone who uses the technology only when it provides a clear advantage. Once raw usage becomes a performance target, organizations risk creating incentives for employees to maximize visible activity rather than meaningful outcomes.
Meta’s adjustment should therefore not be interpreted as a retreat from artificial intelligence. The company remains deeply committed to integrating AI throughout its operations, but its strategy appears to be shifting toward separating employee performance from simplistic consumption metrics. At the same time, Meta is introducing technology that could dramatically increase AI usage without employees having to artificially inflate their numbers.
Central to that effort is an agentic AI system called Hatch, which Meta has reportedly been testing internally. Unlike conventional chatbots that primarily respond to individual prompts, AI agents are designed to perform sequences of actions on behalf of users. Hatch reportedly has the ability to browse the web and interact with applications, potentially allowing employees to delegate more complex workflows rather than simply asking an AI model for information or generated text.
The difference between traditional generative AI and agentic AI is significant because it changes the role of the model from assistant to operator. A conventional chatbot might explain how to complete a task, summarize information, generate code, or draft an email, but the user remains responsible for taking the next steps. An autonomous agent can potentially navigate services, collect information, interact with software, coordinate multiple tools, and complete parts of the workflow itself.
That additional autonomy also introduces substantially greater security and privacy concerns. Some Meta employees have reportedly been reluctant to connect Hatch to personal services such as email and calendars, reflecting concerns about both data access and the possibility that an autonomous system could perform an incorrect or unwanted action. These reservations highlight one of the most important obstacles facing the next generation of AI assistants: users must trust not only what the model says, but also what it is capable of doing.
An inaccurate chatbot response can often be ignored or corrected before anything happens. An autonomous agent with permission to interact with external systems creates a different category of risk because an incorrect decision could result in information being sent, appointments being changed, files being modified, accounts being accessed, or other actions being performed without the user manually executing every step. The consequences of AI errors therefore become increasingly tied to the permissions granted to the system.
For enterprises, this makes identity and access management a critical component of agentic AI deployments. Organizations will need to determine not only what an employee is authorized to access but also what an AI agent acting on behalf of that employee should be permitted to do. Simply allowing an agent to inherit every permission associated with a user account could create unnecessary exposure, particularly when the system can operate across several applications.
The principle of least privilege will consequently become increasingly important. An AI agent should ideally receive only the permissions required to complete a particular task, while sensitive or irreversible actions may require additional confirmation from a human operator. Enterprises may also need detailed audit trails showing which actions were requested by employees, which were proposed by AI systems, and which were ultimately executed autonomously.
Trust is especially relevant inside Meta because the company’s previous internal AI initiatives have reportedly raised concerns among employees. WIRED reported that Meta had operated an internal project involving the collection of employee activity data from work devices for AI training purposes. Although that initiative has reportedly been paused, people familiar with the situation said it affected confidence in some of the company’s AI programs.
Removing AI-consumption metrics from performance evaluations could consequently have an effect beyond simply eliminating an inefficient measurement system. Employees who no longer feel pressured to demonstrate that they are constantly using AI may be more willing to experiment with the technology naturally, selecting it when it genuinely improves their work rather than when they need to satisfy an adoption metric. Counterintuitively, reducing pressure could therefore lead to more meaningful AI adoption.
The arrival of autonomous agents also makes token consumption an increasingly questionable measure of productivity. Agentic systems can require many model interactions to complete a single user request because the system may need to understand the objective, construct a plan, interact with tools, analyze the results, determine the next action, and potentially correct mistakes. A seemingly simple request can therefore generate a large amount of underlying inference activity without the user manually submitting dozens of prompts.
This creates an interesting contradiction for companies that previously celebrated high AI consumption. The era of employees deliberately maximizing token usage may be ending precisely when autonomous systems are beginning to consume significantly more tokens as a natural consequence of how they operate. In an agentic environment, raw token counts become even less useful as indicators of employee productivity.
There is also an important economic dimension. Running increasingly capable models is expensive, and deploying autonomous agents across tens of thousands of employees could generate substantial inference costs. Companies will eventually need to evaluate these systems using metrics closer to conventional business economics, including cost per completed task, time saved, reliability, error rates, automation success rates, and measurable improvements in output.
A system that consumes ten times more tokens but completes a workflow that previously required several hours of human labor could be extremely valuable. Conversely, an AI tool that generates enormous amounts of inference activity while producing marginal improvements could become an expensive technological distraction. Understanding that distinction will be essential as organizations move from experimental deployments toward large-scale production use.
Meta’s evolving strategy may therefore provide an early example of how enterprise AI adoption matures. The first phase of the generative AI boom focused on experimentation, as employees explored what large language models could accomplish. The next phase emphasized adoption, with companies encouraging workers to integrate AI into as many processes as possible and measuring how frequently the technology was being used.
The emerging phase is likely to focus much more heavily on outcomes. Instead of asking how many prompts employees submit or how many tokens they consume, organizations will need to determine whether AI actually reduces development time, improves customer service, accelerates research, eliminates repetitive work, increases software quality, or enables employees to accomplish tasks that previously required significantly more resources.
Agentic AI systems such as Hatch could accelerate that transition because they promise much greater automation while simultaneously creating greater operational risk. If AI systems increasingly perform actions rather than merely generate information, companies will need stronger governance, better observability, more sophisticated permission models, and clearer mechanisms for determining when humans should remain involved.
Meta’s decision to ease off token-based pressure while pushing a more capable AI agent therefore represents more than a minor adjustment to an internal performance system. It reflects a broader challenge facing companies across the technology industry as artificial intelligence moves from novelty to infrastructure: AI adoption itself is not the objective. The meaningful measure is whether the technology allows people and organizations to accomplish more valuable work with greater speed, reliability, and efficiency.
The companies that successfully navigate this transition are unlikely to be those whose employees generate the most tokens. They will be the organizations that learn how to deploy AI where it provides measurable value, establish appropriate safeguards around autonomous systems, and evaluate employees according to the results they produce rather than the amount of computing resources they consume.