The New Denominator for AI Assessment: Token Efficiency

Over the past two years, almost every company has been pouring money into AI. Employees are using large language models, and bosses have included "AI usage rate" and "Token consumption" in performance assessments. But what is the result? Meetings have not decreased, approvals have not disappeared, and workforce efficiency data has stayed flat — "AI is in use, individual efficiency has improved, but corporate workforce efficiency has not risen."

According to the China Enterprise Human Resource Efficiency Research Report released by Musun Consulting, nearly 70% of executives and business owners mention "workforce efficiency" all the time every year, but very few enterprises can truly "understand and control" workforce efficiency.

Today, I will thoroughly explain this topic through two drastically contrasting cases: Volkswagen in Germany is cutting production capacity and redundant positions, while Meta has called a halt to its "replacing manpower with AI" initiative.

01 Workforce Efficiency Is Not a "Simple Division"

Workforce efficiency is the abbreviation of Human Resource (HR) Efficiency, whose essence is the "input-output ratio", that is, the output of operation or business volume divided by human resource input. However, when it comes to actual operation, 9 out of 10 enterprises simplify it into a "simple division": total revenue divided by the number of employees in the enterprise, or total profit divided by the total labor cost of the enterprise.

Against the backdrop of the traditional HR profession that leans more towards "qualitative" rather than "quantitative" approaches, the crisis that HR teams need data-based capabilities as their professional moat, and the ongoing digital transformation of enterprises, this type of "simple division" caters to temporary needs, but obscures the essential significance of workforce efficiency for business operations.

I call this calculation method "broad-caliber workforce efficiency", which calculates the overall efficiency of the company. Bosses and HR teams naturally use it to estimate the reasonable staffing size and total labor cost package of the enterprise. But enterprises that only focus on broad-caliber workforce efficiency will vulgarize their understanding of workforce efficiency, with quite dangerous consequences — it will guide managers to do the easiest thing, that is, "reduce the denominator". By laying off staff, cutting salaries and downsizing the workforce, the statements will immediately show impressive figures, but these figures are often false and leave "internal injuries" to the enterprise.

The core of workforce efficiency management lies in "narrow-caliber workforce efficiency", which refers to the "North Star metric" of performance divided by the "core talent pool". In the digital age, 20% of core talents create 80% of the performance, and the leverage effect is extremely obvious. In the AI era, the leverage effect of core talents is even greater. Perhaps only 10% of the original 20% core talents can be retained, and they can leverage 100% of the performance. This fully proves that narrow-caliber workforce efficiency is the real core driving force for growth.

Furthermore, the importance of workforce efficiency also lies in that it can drive financial efficiency. A large-sample study conducted by Musun Consulting on the A-share market in 2020 found that for enterprises with internet attribute tags, every unit change in workforce efficiency brings a 4.33-unit change in financial efficiency in the same direction.

Taking into account the leverage effect of core talents and this rule, what we really should care about is the input-output ratio of human resources on each business stream (the collection of work flows). This means that to break down the narrow-caliber workforce efficiency, we need to go deep into business scenarios, and observe whether human resources are used in the most critical links in the business scenario, delivering performance that meets expectations and outperforms competitors.

This is obviously a business proposition, which can never be completed by HR teams alone. Business owners should work together with HR teams to calculate detailed accounts and drive business operations.

02 The Trap of "Goodhart's Law"

In the AI era, a new type of "simple division" is becoming popular — assessing Token consumption.

First, let's clarify a common sense in economics. British economist Charles Goodhart proposed in an article in 1975 that "any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes". Intuitively, if policy makers try to achieve goals by controlling a certain indicator, they will often make the indicator lose its original reference value. This judgment was initially used to analyze monetary policy, and later widely applied in many fields such as education, enterprise management, and social policies, which is popularly summarized as "Goodhart's Law". How to understand this law? To put it simply — when an indicator is used as an assessment target, it is no longer a good indicator. Back to enterprises' assessment of "AI transformation", it fully conforms to this rule: if you assess the number of lines of code, programmers will split one line into ten lines; if you assess Token consumption, employees will let AI run idle.

This is not a joke. According to media reports, an engineer at a leading large tech company built an internal Token leaderboard in his spare time. Tens of thousands of employees burned trillions of Tokens in 30 days, and the top-ranked employee alone had a monthly computing power cost of more than one million US dollars. Even more absurd, some people have specially commanded AI agents to run "redundant research tasks" for several consecutive hours, with only one purpose — to brush up their ranking.

I am extremely familiar with this kind of thing. It is exactly the same as the old saying I have criticized before — most enterprises are "treating revenue as cost, and treating cost as expense". You force something that should have measured output into a clearly priced "input" that can be artificially inflated, and employees will naturally try their best to expand this input. As a result, people brush up revenue regardless of cost, and after a while, the cost expenditure becomes unclear, completely turning into shared expenses that are inexplicably spent and increasing year by year. In these enterprises that assess "Token consumption", what employees are brushing up is not even Tokens, but their sense of existence; what they are burning is not computing power, but real money.

I have criticized that workforce efficiency management focuses on input-output ratio, not reducing input. But have you noticed that when enterprises are extremely obsessed with a certain output (such as "AI transformation"), they will go to another extreme — completely forget to assess the input-output ratio, and directly assess the input volume. They take for granted that there is a fixed input-output ratio, that is, when Token consumption rises, the enterprise's "AI transformation" will definitely advance, and thus they fall into the trap of Goodhart's Law.

Essentially, assessing "how much AI is used" is still assessing input, except that the "headcount" is replaced by "Token count". Enterprises think manpower is useless so they cut staff and salaries; enterprises think AI is useful so they increase Token input. Their thinking never goes deep into business scenarios to look at the input-output ratio, which is no different from the rigid thinking that only stares at the denominator.

03 It Is Impossible to Talk About Workforce Efficiency Without Adjusting the Organization

The first case is the self-rescue of Volkswagen in Germany after facing "declining efficiency".

In the first half of 2026, Volkswagen's operating profit margin dropped from 4.2% to 3.8%, and its after-tax profit fell by 30.7%; CEO Oliver Blume himself admitted that Volkswagen's fixed costs are about 30% higher than those of similar companies. To save itself, Volkswagen launched the largest restructuring in its century-old history. It plans to optimize about 50,000 positions in Germany by 2030, with a potential global layoff scale of up to 100,000 people. Its production capacity will be cut from 12 million vehicles to 9 million, the number of models will be reduced by up to half, and configuration complexity will be reduced by up to 75%.

Please note a detail from Blume — the next key cost reduction target of Volkswagen is the costs "outside direct vehicle production", namely the management, group-level positions and central functional departments. To put it simply, Volkswagen believes that what is truly bloated in the company is not the workers on the production line, but the middle and back-office teams and middle management that do not directly "generate revenue".

In other words, Volkswagen's problem has never been "too many people", but "too bloated organization", and the staffing is not used in the core business. That 30% disadvantage in fixed costs is essentially the organizational redundancy piled up layer by layer by department walls, insulation layers and process buckets. Volkswagen's current approach is also simple and crude: layoffs, production capacity cuts and model cuts are all "subtraction" operations.

Is there anything wrong with doing subtraction? No. But it would be a big mistake to "only do subtraction". What is cut off may be the front-line employees who generate revenue, while the people who stay are those who only do superficial work in the administrative departments. I call this operation "organizational compression" rather than "organizational refinement" — the enterprise becomes less and less flexible after constant cuts, and workforce efficiency will only further deteriorate.

Fortunately, according to the plan disclosed by Volkswagen, their target seems to be the real organizational redundancy. But in the actual implementation process, will there be deviations? It remains unknown.

The other case is the failed attempt of Meta to "replace manpower with AI".

Earlier this year, Meta planned a restructuring code-named "Project OT (Organizational Transformation)", whose core is to reshape the operation of the entire company with AI agents and more streamlined small teams. The internal team once envisioned to reduce the size of some teams by up to 60% — through layoffs, hiring freezes and performance-based dismissals.

But what was the result? AI failed to deliver the expected workforce efficiency improvement, and the plan was called off. Meta's own data best illustrates this: the number of code changes on its internal platform increased by 220% year-on-year, but the number of new features or improvements actually launched to users only increased by 36%. The extra code mostly piled up like a blocked lake in the processes of approval, review and rework; AI agents also caused many problems, with major technical safety accidents rising by 40% year-on-year, and employees spent 70% more time putting out fires.

220% vs 36%, what a magical set of numbers. This is exactly what I have been saying all the time — AI improves individual efficiency, which does not equal improving organizational efficiency. AI makes every programmer write code faster, but the code has to go through ten more checkpoints before it can be turned into products. You compress the production process from 3 days to 30 minutes, but the circulation process remains 15 days unchanged. How much faster has this company actually become? This is the truth of "pyramid organization + AI" — it is not equal to an agent-based intelligent organization.

AI investment without organizational transformation is essentially replacing a thicker faucet for a bucket with a leak — the water flow is larger, but the water in the bucket may not increase.

At the end of the day, AI is not used to "replace manpower", but to be embedded in business scenarios to amplify core workforce efficiency. A traditional automaker can even see the organizational problems and start to reduce organizational redundancy, which is very likely to move towards organizational refinement and even organizational transformation; but a Silicon Valley high-tech enterprise, with the idea of "replacing manpower with AI", carries the banner of "organizational transformation" while avoiding real "organizational reform". The real business world is really full of absurdities.

04 To Reshape the Organization, We Need a "New Ruler"

After talking so much, two questions arise:

First, should we pay attention to the denominator? If yes, what denominator should we focus on?

Second, if we cannot simply focus on the denominator, nor only assess Token consumption, what should we assess?

My answer is very simple: first, penetrate into business scenarios and focus on the input-output ratio; second, in the AI-enabled organization, include "Token consumption" in the denominator of efficiency, and shift the ruler from "input" to "input-output ratio".

Following this logic, the "new ruler" for workforce efficiency in the AI era is not difficult to understand.

The first dimension is "Token Efficiency". Divide financial or business data by the total human-machine input — "labor cost + total AI input".

First, let's clarify the concept of "total human-machine input". It is the total consumption of an enterprise running a human-machine team, which is the denominator that AI-enabled enterprises should really focus on. AI investment is divided into two parts: one is the investment in creating agent employees, and the other is the investment in running agent employees. The former is priced according to GPU/TPU computing power duration and the labor cost of training and development, while the latter can be measured by "Token consumption". These two parts, plus the daily labor cost, can be converted using "Token" as the unit.

In this way, we can get the total input of the enterprise's "theoretical Token consumption", and naturally we can measure the "Token Efficiency". Once this indicator comes out, the question of "whether AI really improves efficiency" will turn from a slogan into a mathematical calculation.

If we break down the "theoretical Token consumption" into each business scenario, it will naturally become clearer "which business scenarios have achieved AI efficiency improvement". Volkswagen seems to be able to accurately lock in organizational redundancy; Meta's embarrassment will no longer appear.

The second dimension is "efficiency improvement in key decision-making cycles", which is the core driving indicator after going deep into business scenarios. This indicator does not look at how much AI you use, but only looks at whether you can make decisions faster.

I suggest enterprises select three typical decisions — one budget addition, one cross-department resource allocation, and one new product launch, to measure how many days it takes from proposal to implementation respectively. If you have been using AI for a whole year, but these three figures have not changed at all, it means that AI has only entered the tool layer, not the management layer.

The third dimension is "AI Efficiency Improvement Sharing Index". This indicator is very clever, as it not only focuses on the magnitude of efficiency improvement, but also focuses on the incentive feedback after efficiency improvement. My point is: if the efficiency improvement brought by "AI transformation" does not give feedback to employees and does not motivate them, this kind of efficiency improvement is not sustainable.

Specifically, this indicator asks: how much of the business increment created by AI has been turned into incentives and distributed to employees? If this number is zero, then AI only has one identity in the enterprise — an expense.

These three indicators are interlinked. If Token Efficiency is rising, the key decision-making cycle will definitely be shortened, and the AI Efficiency Improvement Sharing Index will also rise. A single "Token Efficiency" may even make enterprises fall into the trap of Goodhart's Law, but if all three linked indicators are rising, the enterprise will definitely jump out of this trap. You can see that this is somewhat similar to the double-entry bookkeeping in finance.

05 Grab Four Things Immediately in the Third and Fourth Quarters!

Now that 2026 has entered the third and fourth quarters, how to set the workforce efficiency plan for next year? I provide four actionable starting points for the heads of the business management team, functional departments and business departments, which can be used immediately.

First, change the ruler this quarter.

Remove "AI usage rate" and "Token consumption" from employees' KPIs, and replace them with Token Efficiency. Shift the assessment from "how much is used" to "how much revenue or profit can 1 dollar of AI input and 1 dollar of human-machine team input generate in each business scenario". HR and CFO can sit together and finish this adjustment in a week.

Second, select three key decisions and measure the baseline first.

Record the current cycle days of budget addition, cross-department resource allocation, and new product launch respectively. Measure them again at the end of the fourth quarter to see whether AI really speeds up decision-making. Remember, this action is cheaper than buying any amount of computing power, but it can immediately show whether AI has entered the tool layer or the management layer.

Third, cut a number of "non-value-generating establishments".

Learn from Volkswagen, don't only stare at people, but focus on business scenarios. In each business scenario, check your product lines, configuration items, and management levels to see which ones are "treating revenue as cost and treating cost as expense". If there are such cases, cut them firmly. The economic winter has already made many enterprises shiver, but their fear has not yet overcome their habit of sticking to old rules. Our workforce efficiency research report has long pointed out that 43.5% of enterprises still have a flattening index less than 1, with a lot of redundant "supervisor" positions set up; only a little over 30% of enterprises have more than 30% of their employees working on the front line to "generate revenue". These are the parts that should be adjusted. If you don't cut them, are you keeping them for the New Year?

Fourth, calculate an "input-output account" for AI, not just an "input account".

Stop asking "how many Tokens have we used", and ask "how much quantifiable output has this year's AI input and human-machine team input brought back" and "in each business scenario, how much quantifiable output has the human-machine team input brought back". If you can't figure this out, it means you are not at the stage of talking about AI efficiency improvement yet,