After AI Boosts Efficiency, Why Performance Systems Must Be Redesigned

Recently, while working on an enterprise AI deployment project, I ran into a genuinely interesting question.

The client is a B2B company. In the past, a salesperson spent much of the day prospecting, researching companies, finding key contacts, drafting outreach emails, preparing for meetings, writing up meeting notes, updating the CRM, and scheduling the next round of follow-ups based on each account's status.

After we broke the entire sales pipeline apart, we found that a substantial share of this work was already a candidate for agents to take over. A prospecting agent can find and qualify leads automatically. A research agent can fill in company background and contact details. An outreach agent can produce a first draft of the sales email. A meeting agent can write up meeting notes. A CRM agent can update account status from meeting outcomes and flag the next action.

Measured across the whole process, the repetitive work a salesperson actually has to do personally could fall to 30–40% of what it was.

At that point in the discussion, management raised a very natural question:

A salesperson used to be able to reach 30 accounts a day. Now that we have AI, shouldn't that be 100?

The question sounds entirely reasonable. When a company invests in AI, it expects higher output per person. But from the employee's side, another question follows immediately: if I spend time learning the AI, tuning the agents, organizing my own methods, and feeding the system continuously, and the only outcome is that my salary stays the same while my workload triples, why should I push AI forward?

This is the question I find myself tracking more and more across enterprise AI projects.

Companies typically start the AI conversation around tools, models, agents, automation rates, and how many person-hours can be saved. But once AI actually enters the business process, its impact does not stay at the tool layer. When the mode of production changes, job definitions, performance evaluation, incentives, and even managers' sense of "what a great employee is really worth" all begin to shift.

The performance system is simply where this chain of changes finally lands in management mechanics.

---

1. AI Lowers the Cost of Actions, and the Foundation of Many Traditional KPIs Is Disappearing

Many traditional KPIs rest on an implicit assumption: a person's time is finite, and every completed action requires real human effort.

So it was natural to use "number of actions" as a proxy for an employee's effort. How many calls a salesperson makes, how many emails they send, how many clients they visit each day; how many conversations a support agent answers, what their average response time is; how many requirements a programmer completes, how many features they ship; how many images a designer produces in a day, how many assets in a month; how many contracts legal reviews; whether a manager holds meetings on time, writes weekly reports, and summarizes project status.

These metrics made sense in the past because completing the action itself consumed a lot of time.

Once agents start taking over execution, that underlying condition is changing. With AI, a salesperson can generate 300 outreach emails a day without much difficulty, but the business value of those 300 emails is not necessarily higher than that of 50 carefully screened emails with a better customer fit. A designer can generate 100 images a day, but only a dozen or so may be worth testing in production. A coding agent can generate a lot of code quickly, but code volume is a poor proxy for engineering value; system stability, maintainability, sound architecture, and release quality still matter more.

Support is the same. AI can handle large volumes of standard questions at once, but complex complaints, high-value customers, and edge cases still demand more careful judgment.

Behind these changes is a common pattern:

Once AI dramatically lowers the production cost of an action, the volume of that action becomes an increasingly poor representation of the value an employee actually creates.

This is where many enterprise AI efficiency projects go off track. A person used to complete 30 operations a day; an agent expands that capacity to 100; so the company moves the KPI from 30 to 100. On the surface, productivity tripled. But if conversion did not rise in step, customer satisfaction did not change, and profit did not increase, then what the company gained was more action, not a proportional increase in business value.

So as execution becomes cheaper, companies have to answer a question again:

What is actually still scarce and valuable about a role?

---

2. Agents Are Taking Apart the Job

During recent enterprise AI process mapping, I've gotten into the habit of doing one thing first: break the role apart, then discuss which parts are suitable to hand to AI.

Almost every knowledge role contains two fundamentally different kinds of work at the same time.

The first can be called delegable work. It is usually high-frequency and repetitive, with relatively clear inputs and outputs, rules and historical data available, verifiable results, and a manageable cost of a single error.

In sales, prospecting, data enrichment, CRM cleanup, first-draft outreach emails, and meeting notes all fit this description. In support, order lookups, standard replies, ticket classification, and information extraction are similar. Programmers are increasingly handing boilerplate code, unit tests, documentation, and standardized refactors to coding agents. In legal, contract template comparison, risk-item extraction, and standard clause checks offer the same opening; in finance, data cleanup, reconciliation, and standard anomaly detection do too.

These tasks used to consume a large share of people's time. As agent capabilities keep improving, they will keep being absorbed by the system.

What remains of the role will concentrate more and more on human strengths: defining goals, handling ambiguity, making value trade-offs, owning high-risk judgment, resolving exceptions, inventing new methods, maintaining key customer relationships, coordinating across teams, and being accountable for final outcomes.

As a result, the real work chain in many future roles will evolve into something like this:

A person defines the goal and boundaries → agents gather, analyze, generate, execute, and monitor → the person handles exceptions, reviews risk, and makes the key calls → the system records outcomes → the person keeps adjusting rules, knowledge, and ways of working.

This means the employee's role changes too. Many knowledge workers will gradually move from being pure task executors to being managers, trainers, and decision-makers for agents.

In the past, assessing someone's professional ability usually meant asking "can this person do this work?" In the future, we will also need to ask: can they tell whether an agent did it correctly? Do they know when an agent can be trusted? Can they take over quickly when something goes wrong? Can they use feedback and rules to make the agent do better next time?

Once these questions enter the job requirements, the performance system naturally has to change with them.

---

3. In the Future, One Employee May "Manage" Several Agents at Once

This shift is especially vivid in sales.

A mature sales workflow will likely run several agents at once: a prospecting agent finds and qualifies leads, a research agent compiles company background, key contacts, and recent activity, an outreach agent generates emails, LinkedIn messages, and follow-ups, a meeting agent writes up notes and extracts needs, objections, and next steps, and a CRM agent updates account status and flags follow-ups automatically.

At that stage, the work that actually consumes the salesperson's time shifts. They have to judge which accounts are worth the investment, decide on a communication strategy, join key meetings, handle complex objections, build trust, and ultimately drive key decisions and close deals.

The job title may still be "sales," but the mode of production has changed structurally.

That's why I've become more inclined to call out one new capability on its own:

AI Collaboration Capability: Agent Management Capability

An employee's productivity in the future will depend heavily on whether they can break a complex goal into work for different agents, give agents the right context, knowledge, and permissions, judge which outputs can be used as-is and which must be verified by a person, and keep improving how the agents work.

This capability is a lot like traditional management. A good manager does not do all of the team's work personally; they decompose goals, assign tasks, check progress, handle exceptions, and develop team members. As agents spread, this management capability will trickle down from management roles into a large number of ordinary knowledge roles.

One salesperson might manage five agents. One programmer might use code-generation, testing, review, documentation, and research agents at the same time. One operations person might use content, data analysis, monitoring, and auto-execution agents at once.

In the future, many knowledge workers will need a new baseline capability:

Know the business, and manage your own digital workforce.

---

4. Employee Value Will Shift from "Individual Output" to "Leverage Output"

Once employees start managing agents, a thornier question appears in performance: should we look only at how much work the person did themselves, or should we also count the impact they had on the whole system?

I've seen this pattern in support, engineering, and sales. To protect customer information, I've simplified the specific numbers and business details below.

A company has two senior support agents. Employee A is extremely capable: they can handle 150 complex cases a day, know the business cold, and can make quick calls on many edge cases. But they rarely write up their experience, don't maintain the knowledge base, rarely give proactive feedback on agent errors, and keep many of their techniques in their own head.

Employee B can only handle 100 complex cases a day on their own, but they consistently document high-frequency issues, turn their judgment into rules, add to the knowledge base, label the agent's failure cases, and keep refining prompts and processes. In the end, this work raised the accuracy of the agent used by the entire 30-person support team by 5%.

Under a traditional performance system, Employee A easily becomes the star, because their individual throughput is higher. But from the standpoint of total organizational value, Employee B's reach may be wider.

This is where two concepts need to be separated.

Direct output is how much work an employee completed themselves.

Leverage output is how much an employee makes systems, agents, and other employees work better.

Traditional organizations are very good at measuring direct output. Who signed how much, who handled how many customers, who shipped how many requirements — these are relatively easy to record. Leverage output has always been hard to measure, and agents will amplify the problem.

A single excellent rule, once it enters an agent, may be executed thousands of times. A judgment method one employee wrote down used to reach maybe a dozen people through training; now it can go straight into the system and be called by dozens of employees and multiple agents at once.

So a growing part of a strong employee's value in the future will come from a new question:

How many people and how many agents can my experience make better?

Once a company starts taking that question seriously, the object of performance measurement shifts: from a person's individual output toward the organizational leverage a person can generate.

---

5. Why Strong Employees May Naturally Resist "Training the AI"

Push further, and you hit a very real organizational game.

A strong employee contributes their experience to the company. The company writes it into the knowledge base, rules, and agents. Then newcomers gain some of the ability that used to belong only to veterans, the team's overall efficiency rises, and the company hands the employee even more work — while their income, bonus, and promotion prospects barely change.

From a personal-interest standpoint, their motivation to keep contributing know-how naturally declines.

This is why, as many companies push AI, managers see behavior that is hard to understand: employees are reluctant to give the AI serious feedback, experts won't write SOPs, people prefer to solve problems privately, nobody wants to systematically label agent errors, and the most valuable business experience stays in individual heads.

Managers tend to chalk this up to "employees resisting change," but from a mechanism-design standpoint the cause is usually straightforward:

The company designed a system in which hoarding knowledge better serves the employee's own interest.

This becomes even more pronounced in the AI era.

In the past, when an expert wrote their experience into an SOP, it was hard to calculate how much value that knowledge created. Today, if their experience enters an agent and is continuously called by the whole organization, the reach of that contribution can be tens or hundreds of times larger.

Correspondingly, companies need to start answering questions they rarely took seriously before: an employee contributed a rule that raised the whole team's efficiency — how do we value that? An employee found a critical failure mode that lowered the agent's error rate — should that count in performance? Someone spent two weeks optimizing a process, and from then on the whole team saves 500 hours a month — how should that contribution show up in bonuses and promotions?

If these questions go unanswered for long, employees draw a rational conclusion: keeping my experience to myself is safer for me.

---

6. After AI Boosts Efficiency, Companies Must Reopen the Question of How to Split the Gains

Suppose a team saves 1,000 hours of human time in a year through agents. Where should those 1,000 hours ultimately go?

The company of course needs part of the benefit: have one person serve more customers, have the same-size team carry more business, lower unit service cost, raise margins. That is the operating return a company expects from investing in AI.

But if every freed-up hour is immediately replaced with more of the same work, employees will quickly form a very clear belief: the value of AI is that I get to do more work.

Once that belief takes hold, employees' drive to push AI will certainly decline.

I favor a "three-part efficiency dividend" approach.

The first part goes to the business result: scaling output per person, lowering unit cost, raising margins. The second part is reinvested into high-value work, so employees turn their time toward customer relationships, innovation, new products, complex problems, strategic judgment, and exploring new business. The third part must return to employees — through bonuses, promotions, expanded responsibility, less mechanical labor, or an explicit system-contribution metric — so they share in the gains from efficiency.

A healthier loop looks like this:

Employees contribute experience → agents get better → team efficiency rises → the business earns returns → employees receive corresponding rewards → employees keep contributing experience.

If that last link is missing for long, the loop is hard to sustain.

So once AI-driven efficiency reaches real scale, "how to split the efficiency dividend" shifts from a cultural question into a very concrete organizational-mechanism question.

---

7. Future Performance Systems Need a "System Contribution" Dimension

Traditional business-result metrics will of course remain. Revenue, profit, renewal rate, project outcomes, and customer satisfaction are still the most important underlying metrics.

What really changes is that companies need to understand anew: how an employee created those results.

Start with a simple framework:

Broken down further, a future performance system needs to cover at least four kinds of metrics.

The first is outcome metrics — revenue, GMV, profit, renewals, project outcomes, customer satisfaction. These remain the bottom-line business results.

The second is scale and leverage metrics — the business scale one employee manages, customers served per person, agent automation ratio, and business value generated per unit of human time. These measure how much business one employee can move.

The third is high-value behavior metrics — key customer conversations, new strategy tests, creative exploration, complex problem solving, and high-risk judgment. There may not be many of these, but they often directly shape the final result.

The fourth is the system contribution metric that the AI era especially needs to add — effective rule contributions, prompt and agent optimization, knowledge-base contributions, labeling of high-value failure cases, SOP improvements, design of new automated workflows, agent accuracy gains, and improvements to the team's overall efficiency.

New forms of box-ticking creep in easily here too. If a company starts measuring "how many prompts did you submit this month" or "how many knowledge-base entries did you write," it will quickly slide back to being action-volume driven.

So system contribution, too, needs to be tied to actual results wherever possible. A single rule that cuts an agent's error rate by 10% is usually worth far more than 100 knowledge-base entries that the system never called.

---

8. Future Star Employees Will Look Increasingly Like Coaches

You can already see this shift across many roles.

Take programmers. A top programmer used to demonstrate professional value by "being able to write the hardest code." As coding agents mature, more of their value in the future may concentrate on defining architecture, breaking down tasks, setting engineering constraints, reviewing, identifying risk, handling edge cases, and establishing engineering standards. Code still matters, but how much code you personally wrote will gradually move from a core metric to a secondary one.

Sales is the same. A star salesperson's biggest value used to be closing the biggest accounts themselves. In the future, beyond closing deals, they will need to distill their screening logic, communication methods, and objection handling into agents and the team, so their successful approach can be replicated.

A designer's value will likewise expand from a single piece of work to visual direction, template systems, asset libraries, reviewing agent output, and creative exploration. A senior lawyer, beyond reviewing contracts themselves, will need to define risk rules, make clear which clauses must escalate to human judgment, and continuously maintain the boundaries and knowledge of the contract-review agent.

So I think many senior employees in the future will converge on a new combination of roles:

Domain Expert + Coach + AI Supervisor

They remain a domain expert first, with deep enough business judgment. At the same time, like a coach, they turn personal experience into team capability. And one step further, like an AI supervisor, they manage the agents' rules, quality, permissions, and boundaries.

This is one of the most important contributions of a strong employee in the future:

Turn individual capability into a capability the whole organization can keep calling.

---

9. Future Managers Will Also Manage a "Digital Workforce"

As this shift spreads to the organizational level, what managers face changes too.

A manager used to oversee 10 people. In the future, the actual production unit they manage might be:

10 people + 30 agents.

This brings a whole new set of management questions: which tasks should people do and which can be given to agents? What data can an agent read, and how much execution authority does it have? Which operations require human approval? Who is accountable when an agent makes a mistake? At what accuracy level should an agent be stopped? Which agents are worth continuing to iterate and which no longer have business value? Which business experience should be distilled into the system? How do agents work with each other, and where should the handoff between people and agents happen?

These questions are in fact very close to traditional hiring, training, delegation, evaluation, promotion, and removal.

The only difference is that the organization gains a new kind of production resource:

Digital Workforce

In the past, companies mainly managed human resources. In the future, they will manage human and digital labor at the same time, and the two will be highly coupled.

Human capability determines whether agents can be trained correctly, and agent capability determines whether people can free up time for higher-value work. The organization's ultimate efficiency depends on how the company designs the division of labor, permissions, responsibilities, and collaboration between the two.

---

10. The Way We Calculate Productivity Per Person Will Change Too

All of this ultimately comes back to the books.

Companies used to like using:

Revenue ÷ headcount

to measure output per person.

That metric is still worth watching, but when a company uses many agents, counting headcount alone increasingly fails to explain real production efficiency.

Companies may need to add another metric:

Business output ÷ human work hours

Further:

Business output ÷ (labor cost + AI / agent cost)

Further still, companies may need to build separate operating books for agents.

How much human time did each agent save? How much incremental value did it create? How much manual rework did its errors cause? What is the model-inference cost? How much human effort is needed to maintain the knowledge base and processes? Which agents look cutting-edge but ultimately just add token cost? Which agents automate only a tiny step yet happen to free up the scarcest expert time in the organization?

These metrics matter far more than "how many agents did the company launch."

AI management ultimately has to come back to business results. The number of agents is only a process metric; what is worth measuring is whether it actually produced better business outcomes.

---

Finally, What Companies Really Need to Answer Again: How Do We Value One Person?

Companies used to measure an employee's value mainly by how much work they completed themselves. Once agents enter the company, that yardstick becomes increasingly incomplete.

An employee might complete only 100 tasks themselves, yet write a set of rules that makes 100 other people 10% more efficient. A programmer might personally write less and less code, yet establish engineering standards that make the output of multiple coding agents markedly more stable. A salesperson might not rank first individually, yet distill a customer-screening and objection-handling method that lifts the whole team's conversion rate. A senior support agent might handle fewer cases than others, yet through rules, the knowledge base, and failure-case labeling, keep raising the accuracy of the agent the whole team uses.

These kinds of value used to be hard to record. Now that agents exist, they matter more and more — and have more and more opportunity to be quantified by the system.

So once enterprise AI transformation goes deep, the questions inevitably reach the layer of roles, performance, incentives, and organizational structure. Technical teams can build the agents. Business teams can get the processes running. But in the end, management still has to solve a harder problem:

When an employee's experience can be copied into agents, when one person's judgment can be continuously called by the whole organization, and when a strong employee begins to influence dozens of people and dozens of digital workers at once, how should we measure their contribution?

In the future, companies will manage people and digital labor at the same time, and the performance system must answer a question that rarely came up before:

When an employee can make ten agents — or even the entire organization — work better, how should that value be counted?