The New Inequality

Give two people the same AI subscription. One has clean records, useful company data, permission to use that data, a manager who knows what good work looks like, and time to experiment. The other has none of those things. Both can open the same model. Both can type into the same box. On paper, both have access. Economically, they do not.

That difference is the subject of this chapter. The new inequality around artificial intelligence may look different from the old digital divide, where one household had an internet connection and another did not. AI tools are spreading quickly and many are inexpensive. The harder divide is emerging one step later: who can turn access into dependable advantage.

Chapter Eight asked where the productivity gain goes after AI creates it. This chapter moves one step backward. Before a gain can be divided, someone has to be able to produce it. People and organizations begin that race with different skills, data, time, money, authority, and institutional support. AI can reduce some of those differences. It can also magnify others.

Same Tool, Different Starting Line

The simplest mistake in discussions of AI inequality is to treat the technology as a single thing that people either possess or lack. A calculator works fairly close to that way. Once two people have the same calculator, the machine itself is largely equal. General-purpose AI is different because the quality of the result depends heavily on what surrounds the model.

A lawyer with twenty years of experience can recognize when a draft misses the real issue. A new graduate may accept the same draft because it sounds polished. A large company can connect AI to organized internal data, established workflows, testing systems, and specialist teams. A small company may have excellent judgment but scattered records and no technical staff. A worker with permission to use AI on real work can learn rapidly. Another worker may have access to the tool but be prohibited from using the information that would make it useful.

The same AI can produce different value because people bring different resources to it. Economists call resources that make a tool more useful complementary assets. For AI, these might include accurate records, experienced colleagues, training, or permission to use relevant information. Buying the tool does not automatically supply those things.

Access Is Not a Login

It helps to separate basic access from effective access. Basic access means you can reach the tool. Effective access means you can use it reliably for work that matters.

Effective access usually depends on several ordinary things. You need enough knowledge to ask a useful question and judge the answer. You need relevant information that the system is allowed to see. You need a workflow that can absorb the output rather than leaving it in a chat window. You need time to learn where the tool helps and where it fails. In many jobs, you also need organizational permission to act on what it produces.

None of that requires a computer science degree. It requires context. A person who knows the problem, can supply the right information, and can recognize a bad answer may get enormous value from a widely available model. A person without those advantages may receive fluent text and little else.

This is why access is not a login. The economic question reaches beyond who can use AI: who can convert AI into better decisions, more output, lower cost, more time, or a new capability?

The Klarna Advantage

Klarna, the running company in this book, makes the distinction concrete. Its public filings describe AI as part of a broader operating system: a large proprietary data set, a single cloud-based technology platform, substantial engineering and data-science talent, and broad employee use of generative AI.

The surrounding resources matter. A small merchant can buy access to a powerful AI model. That subscription does not also supply years of customer transactions, business programs that work well together, or employees who know how to connect AI to daily work. A larger business may already have those resources. Both can buy similar AI, but one may be better equipped to make it useful.

Klarna is a boundary example rather than proof that scale always wins. Chapter Seven showed the opposite possibility: AI can reduce the amount of organization needed to start useful economic activity. A small firm can now buy capabilities that once required specialists. Lower entry cost and larger returns to integration can be true at the same time. AI can make it easier to begin while still rewarding organizations that connect it deeply to data, talent, and workflow.

The Conversion Gap

I will call the distance between having AI and turning it into reliable value the conversion gap. The phrase is deliberately simple. A tool creates potential. Conversion turns that potential into an outcome.

The gap can appear between companies, workers, professions, schools, and countries. Two firms may buy the same software while only one changes its workflow. Two employees may receive the same assistant while only one is trained to use it on important tasks. Two students may use the same tutor while one has a teacher who can help distinguish understanding from plausible nonsense. Two countries may have internet access while only one has enough language coverage, digital infrastructure, skilled workers, and institutional readiness for broad productive use.

A conversion gap is more difficult to see than a subscription gap. It can exist inside organizations that proudly announce universal access. If senior employees have better data, better training, more discretion, and more experience judging outputs, they may capture much more value from the same system. Formal equality of access can therefore coexist with unequal productive capability.

Permission Is Part of Access

One of the least visible advantages in AI is permission. In many organizations, the most senior people have the broadest authority to experiment. They can approve a new tool, request access to data, redesign a process, or decide that a result is good enough to act on. Junior workers may be closer to the repetitive work that AI could improve, yet have less authority to change the process or even test the tool on real material.

That is often reasonable. Sensitive information cannot simply be pasted into any service, and an organization needs security, privacy, quality, and legal controls. But the controls themselves become part of the distribution of capability. If only a small group is allowed to experiment, that group learns faster. It discovers useful workflows, develops judgment about failure modes, and accumulates experience that later looks like natural expertise. The advantage may begin as a permission decision and end as a skill gap.

A better answer is staged access: widening permission as someone demonstrates the ability to use AI responsibly. A beginner might practice with invented customer records, then work on an approved task under supervision, and later receive more authority. Each step keeps clear limits. The institution’s rules help determine who gets the chance to learn.

Expertise Can Compound

AI is often described as a democratizing technology because it can make specialized tasks easier. That is real. A person who cannot code can build a simple application. A small business owner can draft marketing material without hiring an agency. A student can ask for an explanation at midnight. Those are meaningful expansions of capability.

But expertise can compound with AI rather than disappear. An experienced accountant may use the model faster because the accountant knows which numbers matter. A skilled engineer may delegate routine code and spend more time on architecture. A senior manager may use AI to explore alternatives while relying on years of institutional knowledge to reject the attractive but foolish option.

The result can be uncomfortable because both stories are true. AI can lower the skill required to perform some tasks while raising the value of judgment around those tasks. The person who once had to spend two hours producing a first draft may now spend twenty minutes producing it and the remaining time deciding what the draft should actually say. When execution becomes cheaper, judgment can become more valuable.

Existing expertise is one of the assets that can help convert AI capability into economic advantage, but the outcome depends on how expertise and the tool interact. In the customer-support study introduced in Chapter Seven, the most experienced and highest-skilled agents saw small quality declines as their adherence to AI recommendations increased, and the authors raised a harder question: if the best workers contribute fewer original solutions, where will future model improvements come from? The evidence is showing both equalizing and compounding effects, depending on the job and the way the technology is deployed.

AI Can Also Narrow the Gap

Any serious account of AI inequality has to include the evidence running in the other direction. In a large customer-support study, less experienced and lower-skilled workers gained more from AI assistance than the strongest workers. The tool helped newer employees move down the learning curve much faster. That is a genuine equalizing effect.

The reason is intuitive. A strong worker already carries a library of patterns, examples, and judgment. A novice does not. If the AI can deliver useful pieces of that accumulated experience at the moment they are needed, the novice gains more. In that setting, the technology spreads practical knowledge instead of concentrating it.

This is the most important caution against treating inequality as the inevitable result of AI. The same model can be an equalizer at the task level and a divider at the system level. Whether it narrows or widens a gap depends on what the user lacks, what the tool supplies, and what the surrounding institution allows the user to do with it.

The Firm-Size Divide

Businesses reveal the same tension. AI tools have lowered the cost of capabilities that once required larger staffs, which should help smaller firms. Yet adoption data show that large firms are currently using AI much more often, especially in knowledge-intensive industries.

The explanation is ordinary. Large firms are more likely to have structured data, dedicated technology teams, training budgets, security processes, lawyers who can approve use cases, managers who can redesign workflows, and enough repeated activity to make automation worth integrating. Small firms often have fewer layers, which can make them faster, but they may also lack the time and infrastructure needed to move from experimentation to dependable use.

This creates an access paradox. AI makes sophisticated capability cheaper to obtain while making complementary assets more valuable. A cheap tool can still produce expensive advantage.

The Global Divide

The same question appears between countries. A model available on the internet can cross a border instantly. Productive advantage moves more slowly. Language coverage, education, electricity, digital payments, reliable networks, regulation, company structure, and the mix of occupations all affect how widely AI can create value.

That distinction matters because discussions of AI access often celebrate availability as though distribution has already occurred. Making a model reachable is only the first step. Broad economic benefits require people throughout the economy to have the language, skills, infrastructure, permission, and opportunities needed to use it productively. A country can therefore be connected to the AI economy while much of its population remains outside the productive loop.

The new inequality may therefore look less like a wall and more like a funnel. Many people can enter at the top. Far fewer emerge with the data, training, permission, capital, and market position needed to turn the tool into durable advantage.

Who Gets the Megawatt?

There is another layer beneath skills, data, permission, and capital that is easy to miss because the user never sees it. The cloud has a power cord. Large AI systems run in data centers that require enormous quantities of reliable electricity, along with transmission, substations, cooling, and land. For most users those physical requirements disappear behind an interface. Economically, those requirements remain.

Compute means the computing resources used to train or run AI, including the processing work performed by chips. A data center houses the equipment that supplies those resources. Renting access to AI still depends on physical machines somewhere, and those machines need power.

This changes the inequality question. A country or region may have talented people, internet access, and demand for AI, yet still lack available power at the locations and on the timelines needed for large-scale computing. A company with enough capital can reserve power years ahead, finance new generation, sign long-term contracts, or build near abundant supply. A smaller company usually cannot do those things. The constraint therefore sits upstream of the software.

Electricity scarcity in the United States is regional and temporal rather than one national shortage. The practical bottleneck is often deliverable power: generation, wires, transformers, substations, interconnection capacity, and time. A megawatt that exists somewhere else, or five years from now, cannot power a data center that needs to open next summer.

That makes allocation consequential. When a utility, grid operator, regulator, or market decides which large loads can connect, at what price, and on what schedule, it is doing more than managing electricity. It is helping decide where large-scale AI capability can physically exist. When electricity is scarce, power allocation becomes AI allocation.

The distributional question is broader than data-center owners. New infrastructure can create jobs, tax revenue, and investment, but it can also compete with homes, factories, hospitals, and other businesses for grid capacity or shift the cost of upgrades. The hard question is not whether data centers deserve power. It is who pays for the new capacity, who waits, who receives priority, and who benefits from the intelligence that the electricity makes possible.

The geography is already visible. EPRI estimates that data centers draw roughly a quarter of Virginia’s electricity today and could reach 41% to 59% by 2030, with as many as seven other states above 20%. The digital divide of the AI age may therefore begin at the substation. A society can offer everyone an AI login while allowing the underlying compute, power, and economic leverage to concentrate in a small number of companies and regions. Formal access can expand even as infrastructure advantage deepens.

Current Evidence: Who Can Turn AI Into Advantage

Reviewed September 10, 2026.

The U.S. Census Bureau’s Business Trends and Outlook Survey, which has already appeared in Chapters Five and Seven, shows a clear firm-size gap. In the six months through early May 2026, overall business AI use hovered around 17% to 20%. About 37% of firms with at least 250 employees reported using AI, while fewer than 20% of firms with four or fewer employees did. A companion Census working paper found use rates of roughly 50% to 60% among very large firms in information, professional services, and finance. The tools may be widely available, but integration is not spreading evenly across firm size.

The PwC 2026 Global AI Jobs Barometer introduced in Chapter Eight describes a two-track labor market. Roles it calls professionalised, where AI automates routine work and increases the importance of judgment and expertise, were growing about twice as fast as roles it calls democratised, where AI makes the underlying role easier for non-experts to perform. Advertised salaries in the professionalised group had grown 42% faster since 2021. PwC also reported that jobs demanding specific AI skills were growing 69% compared with 9% for the overall jobs market, with an average advertised wage premium of 62% for AI skills. This is employer-advisory research, not causal proof, but the direction is consistent with a market paying for the ability to turn AI into useful work.

The Anthropic career-stage split introduced in Chapter Eight comes from an April 2026 survey of roughly 81,000 Claude users. Among respondents who identified a beneficiary, 60% of early-career respondents said they personally benefited from AI, compared with 80% of senior professionals. The same survey found large self-reported productivity gains at both the high and low ends of the wage distribution, which is an important counterweight to a simple rich-get-richer story. The sample consists of personal Claude users who chose to respond, and several characteristics were inferred from open-ended answers using AI classifiers, so the figures are signals rather than population estimates.

The strongest equalizing evidence in this chapter comes from the published customer-support study introduced in Chapter Seven by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond. Across 5,172 agents, AI assistance increased issues resolved per hour by about 15% on average. Less skilled and less experienced workers gained much more, with productivity increases around 30% for the least skilled group. Workers with two months of tenure and AI assistance performed about as well as untreated workers with more than six months of experience. The study covered one company and one occupation, but it shows that AI can spread know-how downward when the tool is embedded in a real workflow.

The IMF’s July 2026 working paper offers the broadest warning. Using AI usage data across countries, the authors built a concentration index that asks how narrowly the current value from AI use is distributed across occupations. In developing economies, the index was near 1.0, meaning almost all measured usage-based value was concentrated in a small professional enclave. High-income economies averaged roughly 0.4 to 0.5, with concentration declining in many countries. Income and regulatory readiness were associated with broader diffusion, while the absence of an official English language was associated with slower broadening. The measure is new and based on observed usage rather than total economic output, but it makes the global conversion gap visible.

Electricity is becoming part of the access question as well. EPRI’s 2026 Powering Intelligence scenarios project that data centers could consume 9% to 17% of U.S. electricity by 2030, up from about 4% to 5% today. The International Energy Agency projects global data-center electricity use to more than double to about 945 terawatt-hours by 2030 and says data centers could account for nearly half of U.S. electricity-demand growth through the end of the decade. These are scenario projections with wide ranges, and both point to the same physical constraint: the AI buildout is becoming large enough to change electricity planning rather than merely appear inside it.

Reuters reported on September 1, 2026 that requests from very large electricity users, mostly data centers, exceeded 700 gigawatts across parts of the Midwest, Mid-Atlantic, and South. Those were requests, not confirmed future demand. Some utilities reduced their estimates after requiring deposits.

A deposit requirement tests whether a proposal comes with a financial commitment. It does not, by itself, establish that every project withdrawn was imaginary or could never become viable. For Principle 3, the lesson is to state the conditions for access and check whether applicants meet them. Grid planners still have to assess which projects can proceed and what the power system can support.

Klarna’s 2025 annual report shows what the other end of that gap can look like inside a sophisticated company. The filing says 96% of employees were using generative AI in daily work as of August 2024. It describes a proprietary data set that included 2.6 billion data points collected during 2025 and learning from more than 6.4 billion transactions, a single cloud-based platform, and more than 1,500 engineering and data-science positions at year end, about 53% of the organization. Klarna’s August 2026 second-quarter release added another measure of scale: more than 120 million active consumers and more than 1.2 million merchants.

Those are company-reported facts, not a controlled experiment. They matter because they show the complementary assets surrounding enterprise AI. A broadly available model operating inside that network is economically different from the same model sitting beside a small business with a thin data history and no integration staff.

Accountability Analysis

The Delegation Doctrine does not promise equal outcomes. Its job here is to expose the human decisions that shape opportunity before anyone blames the technology for the entire result.

Principle 1 says accountability does not transfer. If an employer decides who receives AI tools, who may connect them to internal data, who gets training, and who is allowed to act on the output, those are organizational choices. A later capability gap among employees may reflect the technology, but it can also reflect the access architecture the organization created.

Principle 3 requires a boundary. Access without boundaries can expose private data or create reckless automation. Boundaries can also be designed so broadly that only senior employees are allowed to use the tools on meaningful work. The response is to keep the safeguards, make the reason for each boundary explicit, and ask whether training, supervised access, or staged permission can let more people develop competence safely.

Principle 5 becomes unusually literal when electricity is the constraint. Architecture moves accountability upstream. Choices by utilities, grid operators, regulators, and large power buyers about generation, interconnection, priority, and cost allocation can shape where large-scale computing is built and who bears the transition costs. The responsible question is not whether one data center is good or bad. It is whether the allocation rule is visible, contestable, and connected to consequences.

Principle 6 asks where consequences land. If AI increases the productivity of people who already have the most authority while junior workers lose the tasks through which they gain experience, the institution may be creating both a productivity gain and training debt. If small suppliers are expected to match AI-enabled speed without being given the data or integration access available to a dominant platform, the commercial gain and the transition cost may sit in different places.

For organizations, the practical questions are simple. Who gets the tool? Who gets the useful data and context? Who gets training? Who gets permission to act? Who receives credit or compensation when the tool creates value? Those questions do not determine one correct distribution. They make the distribution visible enough to govern.

Find the missing resource

Choose one useful task you cannot yet complete reliably with AI. Identify what is missing: relevant information, practice, permission, a way to check the answer, or dependable equipment and service. Pick one gap you can address and a small task on which to test the improvement. If you manage others, ask who else needs that same opportunity.

Accountability Verdict

The new inequality reaches beyond who has AI. It is increasingly about who can turn AI into dependable advantage.

AI can lower barriers dramatically. It can give a beginner access to expert patterns, a small company access to capabilities it could never afford, and a worker the ability to attempt tasks once reserved for specialists. Those effects are real and worth defending.

At the same time, data, judgment, integration, training, capital, institutional permission, and reliable infrastructure remain valuable when a model becomes cheap. In some settings those surrounding assets become more important because the model makes execution abundant. At large scale, electricity and compute can join that list. The inequality question therefore moves from access to conversion.

The useful goal is to widen the number of people who can participate in the gain without slowing capable people. That means treating training as part of access, preserving paths for beginners to develop judgment, building tools that spread expertise rather than merely harvest it, and noticing when formal access hides a practical barrier. It also means treating compute and electricity infrastructure as part of economic opportunity when public rules determine who can connect, who pays, and who waits.

A cheap AI tool can be broadly available while durable advantage remains concentrated. Watch the conversion gap.

Part Three now moves from economic structure to decisions with direct human stakes. Chapter Ten asks what happens when the machine helps shape diagnosis, treatment, and choices that can determine a person’s health.

Sources and Notes

1. Adam Grundy, Cory Breaux, and Dhanapati Khatiwoda, “Large Firms With at Least 20 Employees Biggest AI Users,” U.S. Census Bureau, May 26, 2026; Kathryn Bonney, Cory Breaux, Emin Dinlersoz, Lucia Foster, John Haltiwanger, and Aditya Pande, “The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks,” CES-WP-26-25, April 2026. https://www.census.gov/library/stories/2026/05/ai-use-businesses.html and https://www.census.gov/library/working-papers/2026/adrm/CES-WP-26-25.html. Accessed September 10, 2026. BTOS data show overall business use around 17% to 20%, 37% among firms with at least 250 employees, and under 20% among firms with four or fewer employees. The working paper reports 50% to 60% use among very large firms in information, professional services, and finance. Source record: DA-CH9-S1. Archive reference: pending.

2. PwC, “AI reshapes global labour market into two distinct paths, rewarding human skills: PwC 2026 Global AI Jobs Barometer,” June 15, 2026, and 2026 Global AI Jobs Barometer full report. https://www.pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-ai-jobs-barometer.html and https://www.pwc.com/gx/en/issues/artificial-intelligence/job-barometer/2026/2026-global-ai-jobs-barometer-full-report.pdf. Accessed September 10, 2026. PwC reports that professionalised roles are growing about twice as fast as democratised roles with 42% faster advertised salary growth, and that jobs requiring specific AI skills grew 69% compared with 9% for the overall jobs market, with an average advertised wage premium of 62%. This is employer-advisory research and is used as a current labor-market signal, not causal proof. Source record: DA-CH9-S2. Archive reference: pending.

3. Maxim Massenkoff and Saffron Huang, “What 81,000 people told us about the economics of AI,” Anthropic Economic Index, April 22, 2026. https://www.anthropic.com/research/81k-economics. Accessed September 10, 2026. Among respondents who identified a beneficiary, 60% of early-career respondents said they personally benefited from AI, compared with 80% of senior professionals. The survey also reports strong perceived productivity gains at both high and low ends of the wage distribution. The sample consists of personal Claude users who chose to respond, and some attributes were inferred from open-ended responses using Claude-powered classifiers. Vendor research is used only as a signal about user-reported experience. The author uses Anthropic products as part of the Living Edition’s external critique process; see the project transparency statement. Source record: DA-CH9-S3. Archive reference: pending.

4. Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, “Generative AI at Work,” The Quarterly Journal of Economics, Volume 140, Issue 2, May 2025, pages 889-942. https://academic.oup.com/qje/article/140/2/889/7990658. Accessed September 10, 2026. In a staggered deployment to 5,172 customer-support agents, AI assistance increased issues resolved per hour by 15% on average, with gains around 30% for the least skilled workers. Agents with two months of tenure and AI assistance performed about as well as untreated agents with more than six months of tenure. The study concerns one company and one occupation and is used here as evidence that AI can narrow some skill gaps when embedded in a real workflow. The authors also report small quality declines among the most experienced and highest-skilled agents as adherence to AI recommendations increased, and they raise the possibility that fewer original contributions from the best workers could weaken future model improvement on new problems. Source record: DA-CH9-S4. Archive reference: pending.

5. Rachel Yuting Fan and Ha Minh Nguyen, “Aggregate Gains from AI and Their Distribution: Global Evidence from Usage Data,” IMF Working Paper 2026/147, July 10, 2026. https://www.imf.org/en/publications/wp/issues/2026/07/10/aggregate-gains-from-ai-and-their-distribution-global-evidence-from-usage-data-577586. Accessed September 10, 2026. Using five waves of Anthropic Economic Index usage data, the authors construct an AI concentration index. They report concentration near 1.0 in developing economies, indicating that almost all measured usage-based value is generated in a small professional enclave, versus averages around 0.4 to 0.5 in high-income economies. Income and regulatory readiness predict broader diffusion, while lacking an official English language is associated with slower broadening. The underlying usage data come from the Anthropic Economic Index; Anthropic is also used in the Living Edition’s external critique workflow. The IMF analysis, rather than Anthropic’s interpretation, is the source for the concentration finding. IMF Working Papers are research in progress and do not necessarily represent IMF views. Source record: DA-CH9-S5. Archive reference: pending.

6. Electric Power Research Institute, “Powering Intelligence 2026: Updated Scenarios of U.S. Data Center Electricity Use and Power Strategies,” February 26, 2026. https://esca.epri.com/products/3002034697/. Accessed September 10, 2026. EPRI projects data centers at 9% to 17% of U.S. electricity consumption by 2030, up from roughly 4% to 5% today, and notes that access to power and land is increasingly shaping where large AI training centers are planned. EPRI also reports that data centers already draw roughly a quarter of Virginia’s electricity and could reach 41% to 59% by 2030, with as many as seven other states above 20%. The scenarios are projections rather than a forecast with a single expected outcome. Source record: DA-CH9-S6. Archive reference: pending.

7. International Energy Agency, “Energy and AI,” executive summary, 2025, current online edition. https://www.iea.org/reports/energy-and-ai/executive-summary. Accessed September 10, 2026. The IEA projects global data-center electricity consumption to rise to about 945 terawatt-hours by 2030, more than double current use, with AI as the largest driver of the increase. In the United States, data centers account for nearly half of projected electricity-demand growth through 2030 in the IEA base case. Source record: DA-CH9-S7. Archive reference: pending.

8. Reuters, “Texas’ halt on powering data centers reflects US reckoning over ‘ghost’ demand,” September 1, 2026. https://www.reuters.com/business/texas-halt-powering-data-centers-reflects-us-reckoning-over-ghost-demand-2026-09-01/. Rechecked September 20, 2026. The figure covers large-user requests, mostly data centers, across parts of three U.S. regions. Source record: DA-CH9-S8. Archive reference: pending.

9. Klarna Group plc, Annual Report 2025; “Klarna Reports Second Quarter 2026 Results,” August 18, 2026, furnished on Form 6-K. https://www.sec.gov/Archives/edgar/data/2003292/000200329226000007/klar-20251231.htm and https://www.sec.gov/Archives/edgar/data/2003292/000162828026057573/klarnareportssecondquart.htm. Accessed September 10, 2026. The annual report reports 96% employee generative-AI use as of August 31, 2024; 2.6 billion data points collected in 2025; learning from more than 6.4 billion transactions; and more than 1,500 engineering and data-science positions, about 53% of the organization. The Q2 release reports more than 120 million active consumers and more than 1.2 million merchants. These company sources show the complementary assets and scale surrounding enterprise AI, not a causal AI effect. Source record: DA-CH9-S9. Archive reference: pending.