The Real Future of Work Is a Payment Reform Problem

What if the biggest obstacle to better healthcare and better artificial intelligence is the same mistake: asking people to create value while rewarding someone else for capturing it?

Healthcare reform offers a revealing case study. One model gives doctors and hospitals responsibility for managing care, then asks them to improve outcomes and reduce costs. Another gives insurers a large, predictable payment for each enrolled person and lets them organize the resulting system. The first may produce better care, but the second often attracts more capital, talent, and institutional energy.

Generative AI is arriving at a similar fork in the road. Organizations can treat it as a tool that employees are expected to use for the public good, or they can redesign workflows so that the people who adopt it gain time, authority, income, or professional leverage. The difference will determine whether AI becomes a durable productivity system or merely another layer of unpaid adaptation.

The deeper lesson is simple: reform does not spread because it is beneficial. It spreads when the people who must change can see, and reliably collect, the benefit.

The neglected question behind every reform

Consider two approaches to improving Medicare. Accountable care organizations, or ACOs, give groups of providers responsibility for the cost and quality of care delivered to a population. In principle, this aligns clinical judgment with public value. If physicians prevent avoidable complications, coordinate treatment, and reduce unnecessary spending, everyone should benefit: patients, taxpayers, and the providers themselves.

Yet the financial signal has often been weak. Although many ACOs improve quality and lower spending compared with similar nonparticipating providers, only 63 percent earned performance payments in one reported period. Participation has declined from its peak. The model asks organizations to invest in data systems, care coordinators, new routines, and difficult conversations with patients, while making the reward uncertain and delayed.

Medicare Advantage presents a different incentive structure. Insurer run plans receive a substantial payment for each member, reported at roughly 104 percent of the cost of traditional Medicare. That payment creates a large, visible pool of revenue from which an organization can fund administration, technology, marketing, clinical programs, and expansion. Whatever one thinks of the resulting quality or cost, the business case is clear enough to attract major players.

This contrast is more important than the familiar ideological argument about whether doctors or insurers should control care. It reveals a general design principle: the location of responsibility must be matched by the location of reward.

If a primary care practice is told to manage the long term health of a population but is paid mainly for individual visits, it is being assigned a systems problem with a transaction based budget. If an insurer receives a predictable population payment but can externalize much of the clinical burden, it has a strong incentive to scale enrollment and optimize the payment mechanism.

The result is not necessarily a failure of goodwill. It is a failure of economic architecture. People respond to the scorecard that governs their survival, especially when the requested behavior requires sustained investment.

The question is not merely, “What produces the best outcome?” It is, “Who has to change their behavior, who pays for the change, and who receives the return?”

That three part question is also the key to understanding the future of work.

Why AI adoption will stall without a value capture mechanism

Generative AI can draft documents, summarize information, produce first versions of analyses, answer routine questions, and assist with software development. But these capabilities do not automatically create organizational productivity. They become valuable only when embedded in a workflow, connected to relevant information, checked by accountable people, and linked to a decision that matters.

This distinction is easy to miss because demonstrations make AI look like an individual superpower. A worker enters a prompt and receives a polished answer in seconds. The visible act is faster. The invisible system around it may not be.

A clinician using AI to prepare a visit summary still needs accurate records, privacy safeguards, review protocols, and a way to incorporate the summary into care. A claims analyst using AI to identify suspicious patterns needs access to integrated data, a process for investigation, and authority to act on the result. A manager using AI to write reports may save an hour, but the organization realizes no real gain if the saved hour is consumed by new review requirements or simply filled with more reporting.

This is where the healthcare analogy becomes useful. AI is not primarily a software adoption challenge. It is an incentive and workflow allocation challenge.

Suppose a hospital introduces a generative AI system that reduces the time doctors spend documenting. If the hospital responds by increasing the number of patients each doctor must see, physicians may experience the tool as a speedup imposed on them, not as an improvement they own. If the saved time is used to coordinate care, educate patients, or reduce burnout, the same tool becomes part of a professional renewal strategy.

The technology has not changed. The distribution of value has.

Organizations often make this mistake with knowledge workers. They measure the time AI might save, then assume the savings will become productivity. But time is not the same as value. A saved hour can become profit, quality, learning, rest, additional volume, or administrative burden. The outcome depends on who controls the hour after it has been liberated.

This is the value capture problem. When one group bears the cost of adopting a new system and another group captures the benefit, adoption becomes fragile. Employees must learn new practices, correct model errors, redesign tasks, and absorb uncertainty. Executives may capture the resulting margin. Customers may receive faster service. Shareholders may receive higher returns. The people doing the adaptation may receive only an expanded workload.

Under those conditions, resistance is not irrational. It is information. It signals that the organization has not explained how the transition will produce a fair exchange.

The difference between a tool and an operating model

A useful way to think about AI is to separate three layers that are often collapsed into one.

Layer one is capability. Can the system generate a summary, answer a question, classify a document, or create a draft?

Layer two is workflow. Where does that output enter the real sequence of work? Who reviews it? What happens when it is wrong? Which decisions become faster or better?

Layer three is governance and reward. Who is accountable for the result? How is performance measured? Who receives the gains, and who bears the risks?

Most organizations begin and end at layer one. They purchase access, run experiments, and publish usage statistics. This is similar to giving providers a new responsibility without giving them a dependable payment pathway. The formal program exists, but the operating logic is incomplete.

The strongest AI transformations will therefore resemble well designed population health programs more than software rollouts. They will begin with a defined outcome, assign responsibility to the people closest to the work, provide the resources to change the workflow, and establish a credible way for participants to benefit.

Imagine three companies introducing the same clinical documentation assistant.

The first company measures only how many notes the system generates. Employees learn to produce more text, while errors and duplication increase.

The second measures minutes saved per note. Managers celebrate efficiency, but clinicians are required to fill the freed time with additional appointments. Burnout worsens, and staff quietly avoid the tool.

The third measures a broader result: documentation accuracy, time returned to patient interaction, reduced after hours work, and continuity of care. Clinicians help refine the system and receive a share of the operational benefit through staffing flexibility, professional development, or reduced caseload pressure. Adoption is slower at first, but the workflow becomes durable.

The third organization understands that productivity is not simply an increase in output per worker. It is an improvement in the relationship between effort, quality, and human capacity.

This has implications beyond healthcare. A customer service representative who uses AI to resolve routine issues may be rewarded for handling more tickets. That encourages speed. Or the representative may be given authority to solve more complex problems, with performance measured by resolution quality and customer retention. That encourages judgment. The same technology can either deskill the job or elevate it, depending on the scorecard.

A technology becomes transformative when it changes what the worker is able to own, not merely what the worker is able to produce.

The hidden politics of productivity

Productivity is often presented as a neutral goal, but every productivity program distributes power. If a process becomes more efficient, someone decides what happens to the surplus.

The surplus might support lower prices, higher wages, shorter hours, better service, new investment, or higher executive compensation. There is no natural outcome waiting inside the technology. Institutions choose the outcome through contracts, budgets, targets, and norms.

This is why the debate over artificial intelligence cannot be reduced to whether the technology will replace jobs. The more immediate question is whether it will reorganize bargaining power inside jobs.

A worker with AI assistance may become more productive and less autonomous if every output is monitored and every saved minute is converted into additional assignments. Another worker may become more capable and more valuable if AI removes routine tasks while leaving judgment, relationship building, and decision authority in human hands.

Healthcare payment reform illustrates the same political reality. ACOs may generate better care at lower cost, but that achievement does not guarantee expansion. If the organizations making the investment cannot predictably retain part of the savings, the model will lose momentum. Meanwhile, a model with a clearer revenue stream can grow even when evidence about its comparative quality is mixed.

That is not an argument that predictable payment always produces better outcomes. It is an argument that scalability follows credible value capture more reliably than it follows moral superiority.

For leaders, this suggests a practical test for every AI initiative:

1. What costly or valuable outcome are we trying to change?
2. Which workers must alter their daily behavior?
3. What new burden will adoption place on them?
4. How will they know the change is working?
5. What tangible benefit will return to them if it succeeds?

The last question should not be treated as a motivational bonus. It is part of the system design. The return may be financial, but it can also be autonomy, reduced administrative load, better scheduling, career mobility, or the ability to spend more time on meaningful work. What matters is that the benefit is concrete, visible, and proportionate to the adaptation required.

Designing the next generation of work systems

The most promising organizations will move from a tool centered model to an outcome centered model. They will not ask, “Where can we add AI?” They will ask, “Where is a valuable outcome being blocked by repetitive cognitive work, fragmented information, or delayed feedback?”

That shift produces a more disciplined adoption sequence.

First, identify a population, customer group, or workflow with a measurable problem. In healthcare, this might be patients with multiple chronic conditions who experience fragmented care. In a business, it might be customers whose issues require repeated handoffs.

Second, assign ownership to a group that can actually influence the outcome. Responsibility without authority is as damaging as authority without responsibility. A team cannot be accountable for reducing readmissions if it cannot coordinate follow up, access patient data, or influence scheduling.

Third, fund the transition. Training, review time, data cleanup, and process redesign are not incidental costs. They are the equivalent of the infrastructure required for an ACO to manage a population.

Fourth, measure both the result and the burden. If an AI system improves turnaround time but increases error correction, hidden work, or employee exhaustion, the dashboard is lying by omission.

Fifth, precommit to sharing the gain. If the organization cannot state in advance how success will improve the lives of the people doing the work, it should expect suspicion and slow adoption.

This framework also guards against a common failure: using AI to optimize a broken process. Automating a needless approval chain does not create transformation. It creates a faster needless approval chain. In the same way, paying for more transactions while ignoring fragmented care does not create population health.

Key Takeaways

• Match responsibility with reward. Do not assign teams accountability for outcomes they cannot influence or benefit from.
• Measure workflow impact, not tool usage. The number of prompts, generated documents, or automated tasks says little about quality, cost, or human capacity.
• Treat adoption costs as real investment. Training, review, data integration, and redesign determine whether a technology survives beyond the pilot stage.
• Make value sharing explicit. Return the gains through compensation, autonomy, reduced workload, professional growth, or better service conditions.
• Audit the distribution of surplus. Whenever AI saves time or reduces cost, ask who receives the resulting benefit and who absorbs the new risk.

The future of work will not be decided by the raw power of generative AI alone. It will be decided by the institutional arrangements built around that power. The same is true of healthcare reform. A model can be clinically admirable and economically fragile, or financially scalable and socially disappointing.

The uncomfortable conclusion is that better ideas do not win by being better ideas. They win when their incentives make sustained participation rational for the people who must carry them into reality.

So the defining question for leaders is not, “What can this technology do?” It is more demanding: “What new bargain are we offering the people whose work will make it valuable?”

If the answer is clear, fair, and tied to a meaningful outcome, AI may become more than an efficiency tool. It may become a way to redesign work around judgment, care, and human capacity. If the answer is vague, the technology will likely do what many reforms do: generate impressive evidence in the pilot, produce modest gains in practice, and eventually lose momentum to the systems that made value capture easier.