What happens to leadership, management, developers, marketing, sales, creative work, and the structure of business itself — Part 3 of the Internet 2027 series.

The first and second articles in this series examined changes taking place at the interface between organizations and the outside world. The Internet is increasingly serving people and machines simultaneously, while AI systems move beyond crawling information toward interpreting, comparing, recommending, and acting upon it. That transition makes reliable organizational knowledge more important because automation magnifies both the value of accurate information and the consequences of bad information.

Once those pieces come together, the transformation inevitably moves inside the company. If a business can establish reliable knowledge, connect its systems, and deploy AI capable of interpreting and acting on that information, then many activities historically performed by people no longer need to occur in exactly the same way. This is where the discussion often becomes unnecessarily polarized, with one side predicting rapid elimination of large parts of the workforce while the other treats AI as little more than a better productivity application.

The evidence supports a more consequential but less sensational conclusion. AI is already changing meaningful portions of work, yet simply giving employees access to AI does not automatically redesign a company. The larger transformation begins when leaders stop asking only how AI can make an existing task faster and begin asking why the task, workflow, department boundary, approval structure, or management layer exists in its current form.

Using AI is not the same as operating differently

Organizational AI use is already widespread. Stanford’s 2026 AI Index summarized 2025 survey data indicating that roughly 88 percent of organizations reported using AI in at least one business function, although the depth of those deployments varied and agent use remained considerably less mature — fewer than 10 percent of organizations reported having fully scaled AI in any single business function (Stanford HAI). That distinction between access and structural adoption is critical because a company can use AI extensively while preserving almost every important element of its previous operating model.

An employee can summarize email without changing the company’s communication architecture. A marketer can generate campaign copy while the campaign still passes through the same approval process. A developer can use an AI coding assistant while the underlying requirements, technical debt, testing procedures, and deployment constraints remain unchanged. A manager can summarize every meeting faster while the organization continues to hold the same unnecessary meetings.

One of the strongest pieces of evidence in the research comes from a six-month randomized experiment involving 7,137 workers across 66 firms who were given access to a generative AI assistant integrated into their everyday work tools. Employees with AI access saved time in communication-related work, including approximately two hours per week on email in the study, yet researchers did not observe a broad recomposition of work simply because AI tools were available (National Bureau of Economic Research).

That finding should temper both enthusiasm and fear. Giving employees AI does not automatically produce a dramatically smaller organization, but it also does not demonstrate the limit of what AI can change. It demonstrates that transformation requires redesign rather than procurement.

The workflow, not the job title, is the real unit of change

Companies are generally organized around functions: marketing, sales, finance, technology, operations, customer service, and leadership. Customers experience something different. A single customer journey may begin with marketing, move into sales, require pricing from finance, depend upon availability from operations, touch technology during a transaction, and eventually involve customer service. Every departmental boundary creates transfers of information, responsibility, and context.

Historically, people handled much of that movement manually. An employee sent an email, someone updated a spreadsheet, a manager collected status reports, a meeting reconciled conflicting interpretations, and another employee moved information from one system into another. These activities became so normal that businesses often stopped recognizing them as the cost of disconnected processes and instead began treating them as work in their own right.

AI is particularly well suited to many of the underlying activities: retrieval, summarization, classification, drafting, comparison, routing, monitoring, and standardized decision logic. The research therefore anticipates a gradual move toward humans supervising broader portfolios of automated execution when workflows are sufficiently explicit, measurable, and governable. It also cautions that deep agent use and true operating-model redesign remain substantially less mature than broad AI adoption.

This is why starting with the question “How many people can AI replace?” usually leads the organization in the wrong direction. The better questions are what outcome the process exists to produce, which steps genuinely require human judgment, which depend on authoritative knowledge, which can be standardized, where exceptions occur, which systems should perform each action, and who remains accountable for the result. Only after the workflow is understood does staffing become a meaningful question.

The organization begins to resemble its digital infrastructure

There is a useful parallel between the evolution of the website and the evolution of the company. The traditional website consisted largely of pages created individually for humans, with information assembled into each presentation. The emerging website increasingly resembles a governed system of knowledge, media, identity, transactions, permissions, and capabilities that can be exposed through several different interfaces.

Traditional companies often operate like the old website. Knowledge is distributed among departments and individuals, employees manually assemble context, and work moves through layers until someone has enough information and authority to act. Much of the organization effectively exists to compensate for information that systems cannot understand or move on their own.

The emerging model increasingly makes reliable knowledge explicit, connects systems, formalizes rules and permissions, automates predictable execution, and uses AI to interpret less-structured information within defined boundaries. Humans remain essential, but their work moves toward direction, judgment, exceptions, relationships, creativity, and accountability rather than serving primarily as the connective tissue between systems.

The transformation is therefore not accurately described as replacing a human worker with an AI worker. It is better understood as converting a human-operated process into a governed system that combines human and machine execution according to what each is suited to do.

AI changes tasks before it changes occupations

The research is substantially stronger on task transformation than on sweeping employment predictions. Across every role examined, administrative and first-draft production work becomes easier to automate or augment, while judgment, relationships, accountability, system design, evidence, and distinctive expertise remain comparatively scarce. The magnitude differs dramatically across occupations and tasks, and no reviewed evidence supports universal near-term elimination of managers, developers, marketers, salespeople, or creative professionals.

Most jobs contain several different types of work. Administrative work includes reporting, scheduling, documentation, coordination, formatting, and information collection. Production work includes writing, coding, analysis, design, proposals, campaigns, and other outputs. Judgment work involves ambiguity, prioritization, risk, tradeoffs, and deciding what should happen when the obvious rule does not apply. Relationship work includes leadership, trust, negotiation, persuasion, conflict resolution, customer understanding, and accountability between people.

AI affects those categories differently, which is why job titles are usually too crude a unit for determining automation potential. A manager can lose half of the administrative burden while becoming more valuable as a leader, a developer can produce substantially more code while spending more time on architecture and validation, and a salesperson can automate research while becoming more responsible for diagnosis and negotiation. The occupation may remain recognizable even as its center of gravity moves substantially.

The CEO becomes less dependent on the information pyramid

Traditional organizational hierarchies exist partly because senior leaders cannot personally consume everything happening inside a business. Employees report to managers, managers consolidate information for directors, directors summarize it for executives, and executives eventually present the most important issues to the chief executive. Every layer performs leadership functions, but it also performs information compression.

AI can reduce some of that dependency by continuously synthesizing operating metrics, competitive information, customer feedback, project status, financial trends, and organizational signals. That does not make the chief executive role easier in the dimensions that matter most; it simply makes access to analysis less scarce. The harder responsibility remains deciding what deserves attention, what tradeoffs are acceptable, and where the organization should go.

The research anticipates a larger span of accessible analysis and execution for CEOs and founders, making the design and governance of an AI-enabled operating system more explicit elements of leadership. The role therefore moves further toward enterprise architecture in the broad business sense: determining objectives, capital allocation, proprietary capabilities, acceptable risk, human versus machine authority, and the accountability structure behind automated decisions.

These responsibilities have always belonged to leadership, but AI increases their leverage. When one strategic decision can be implemented through a larger volume of automated execution, weak judgment becomes more expensive and good judgment becomes more valuable.

Senior leaders have to understand the system, not only the department

AI also makes functional boundaries harder to defend as independent worlds. Marketing increasingly depends on identity systems, structured data, automation, AI-generated media, customer information, and technical infrastructure. Sales depends on data quality, pricing systems, digital discovery, content, and automated qualification. Operations depends on software, predictive systems, and customer communications, while technology decisions increasingly shape customer experience, marketing effectiveness, financial controls, and business strategy.

The research therefore anticipates fewer coordination steps in well-instrumented environments and a shift toward leaders managing portfolios of people, software, and agents rather than only traditional human teams. At the same time, executive alignment, stakeholder negotiation, crisis leadership, talent sponsorship, organizational judgment, and accountability remain human-intensive responsibilities.

This does not require every executive to become a programmer. It does require senior leadership to understand enough about interconnected systems to recognize when optimizing one department creates a problem elsewhere. The company increasingly competes through the effectiveness of the whole operating system rather than through the isolated performance of individual functions — the same premise behind HCG’s view that most AI disappointment traces back to the operating model, not the technology.

Middle management faces a split between administration and leadership

Few organizational roles attract more AI speculation than middle management because so much managerial time is currently consumed by coordination. Deloitte’s 2025 Global Human Capital Trends research reported managers spending roughly 40 percent of their time on current problems and administrative work and only about 13 percent on people development (Deloitte). That creates an obvious opportunity for automation in scheduling, reporting, status collection, routine follow-up, meeting synthesis, work routing, documentation, and capacity analysis.

The mistake is to treat those activities as the full definition of management. Strong managers resolve ambiguous priorities, coach employees, handle conflict, interpret organizational context, make local tradeoffs, identify when the standard process does not fit the situation, and maintain accountability without escalating every decision. Those activities are fundamentally different from assembling a weekly report.

The evidence therefore supports distinguishing administrative management from leadership management. Administrative coordination can compress substantially as systems improve, allowing capable managers to spend more time on judgment, coaching, change leadership, exception handling, and organizational sensemaking. The same research warns that simply removing layers can create centralized bottlenecks or shadow management when the underlying coordination problem has not actually been solved.

A more likely transformation is therefore not the disappearance of management but a higher standard for what management is expected to contribute. Managers who primarily move information may become increasingly exposed, while managers who improve the performance of people and systems may become more valuable.

Developers move up the abstraction stack

Software development offers one of the clearest demonstrations of why the ability to generate an output is not equivalent to owning the professional responsibility behind it. AI can already produce code, tests, scripts, queries, documentation, refactors, and debugging suggestions that once required considerably more manual effort. That capability will almost certainly continue improving.

Yet software is not simply accumulated code. Someone still has to determine what should be built, understand how systems relate, design the data model, establish security boundaries, evaluate reliability, handle failure, and accept responsibility when the system does not behave as expected. The easier code becomes to generate, the more important these higher-level responsibilities become.

Evidence on developer productivity illustrates the unevenness of the transition. Some controlled studies show substantial gains on bounded coding tasks, while METR’s early-2025 study of experienced open-source developers working in mature repositories found participants were 19 percent slower when AI tools were available, even though they had estimated after the fact that the tools had sped them up (METR). The result does not mean AI makes developers slower generally; it demonstrates that productivity depends on the task, the codebase, the expertise of the developer, and the friction introduced by the tool.

The emerging developer therefore moves upward in abstraction. Memorizing syntax and manually producing boilerplate decline as differentiators, while architecture, integration, debugging, security, system ownership, context engineering, evaluation, and product understanding become more valuable. A developer may personally write less code while becoming responsible for a substantially larger body of generated software, which is not a reduction in responsibility but an expansion of it.

Marketing stops being primarily an asset-production function

Marketing may experience one of the fastest visible changes because such a large portion of the traditional workflow involves producing and adapting assets. Writing emails, creating advertising variants, resizing imagery, translating copy, assembling landing pages, generating reports, producing social content, and creating personalized versions are precisely the kinds of activities generative systems can perform quickly.

This does not make marketing less important. It exposes the difference between producing marketing material and understanding markets. A machine can generate fifty headlines, but someone still has to understand which promise matters to the audience. A system can produce a thousand personalized variants, but someone must determine what information should influence personalization and what crosses an ethical or strategic line. AI can generate claims, while the business remains responsible for whether those claims are true.

The research therefore anticipates marketing becoming a governed content and experimentation system in which humans concentrate more heavily on insight, positioning, quality thresholds, authority, and strategic judgment. Commodity production becomes less scarce, while customer understanding, differentiation, editorial judgment, experimentation, creative direction, claims governance, and brand stewardship increase in relative value.

This connects directly to the truth-infrastructure problem discussed in Article 2. When an organization can generate exponentially more communications, weak underlying knowledge becomes exponentially more dangerous. The ability to create more does not remove the obligation to know what deserves to be said.

Sales loses information asymmetry but not the need for trust

Sales historically benefited from controlling access to information. Salespeople knew the product, pricing, implementation details, competitive differences, limitations, and common objections, while buyers often needed to engage with sales simply to understand the category. Search reduced that asymmetry, and AI reduces it further by giving buyers increasingly sophisticated tools for comparing vendors, summarizing reviews, analyzing proposals, preparing questions, and identifying alternatives before speaking with anyone.

This shifts the salesperson’s contribution away from information delivery. Basic prospect research, follow-up writing, CRM notes, meeting preparation, standard product questions, and proposal assembly can increasingly be automated or augmented. The research does not, however, establish that complex relationship selling is disappearing; high-stakes discovery, stakeholder alignment, negotiation, political navigation, commercial accountability, and trust remain highly contextual.

The salesperson therefore increasingly becomes a decision partner rather than an information gatekeeper. If the customer already understands the product, sales must understand the customer. If AI can generate the comparison, the human must explain why the differences matter in this particular organization. When the buying decision is consequential, the value of credibility and judgment may actually increase because the buyer has easier access to generic information.

This raises the bar for sales rather than eliminating it. Commodity information becomes cheaper, while diagnosis, domain expertise, negotiation, relationships, and accountable commercial judgment become more valuable.

Creative work moves from production scarcity to judgment scarcity

Creative professionals face perhaps the most visible form of AI disruption because generative systems can now create imagery, layouts, voiceovers, storyboards, animations, concepts, translations, and increasingly sophisticated video. When production required scarce technical skill and significant time, the ability to produce was itself part of the creative advantage.

As generation becomes abundant, the source of differentiation shifts. A system can create twenty visual concepts in seconds, but someone still has to recognize which concept communicates the right idea, whether it is genuinely original, whether it fits the brand, and whether it deserves to exist. The research therefore anticipates smaller teams exploring more variants while scarcity moves toward selection, coherence, taste, and defensible originality.

This makes creative direction more operationally important rather than less. When anyone can generate polished imagery, polish alone cannot establish distinction. When every company can produce competent video, the quality of the concept, story, point of view, and editorial judgment becomes more valuable.

The creative professional increasingly becomes a director of possibility rather than only an executor of production. Craft remains important where craft itself contributes to the value, but speed of manual execution becomes a weaker moat as generative tools improve.

Customer service demonstrates how AI can distribute expertise

Customer service provides one of the strongest field examples of augmentation because the role combines repetitive information retrieval with human communication. In a study covering 5,179 support agents, access to a generative AI assistant increased issues resolved per hour by 14 percent on average and by 34 percent among novice and lower-skilled workers, with minimal impact on already-experienced agents (National Bureau of Economic Research).

The importance of that result goes beyond throughput. Experienced service employees accumulate patterns over time: they recognize which questions to ask, which unusual combinations indicate a deeper problem, and how to communicate effectively in difficult situations. AI can distribute part of that accumulated knowledge to less-experienced employees much earlier in their development.

That benefit raises an important unresolved question. If AI performs more of the routine work through which junior employees traditionally gained experience, how will future experts develop the judgment required when AI fails? The research identifies this apprenticeship problem as an unresolved organizational issue across several professions.

AI can shorten the path to competent performance, but companies still need mechanisms for developing genuine expertise. An organization that eliminates every novice task may eventually discover that it has also eliminated part of the process through which experts were created.

Operations increasingly moves toward exception management

Operational work follows a similar pattern. Many processes require people to review information, route requests, update records, confirm status, notify customers, and escalate unusual cases because the underlying systems are unable to manage the sequence independently. As software becomes more capable, predictable portions of those workflows can increasingly execute automatically.

The research finds case summarization, routing, standard communication, transaction processing, information lookup, and classification increasingly augmentable or automatable, while complex exceptions, fraud and risk, policy interpretation, root-cause analysis, and empathy-sensitive situations remain materially human.

That changes the operating skill profile. Memorizing procedures becomes less valuable when a system can retrieve the correct procedure instantly, while recognizing that the documented procedure does not fit the situation becomes more valuable. Manual monitoring becomes less important as systems monitor continuously, while diagnosing why a process is failing becomes more important.

The future operator therefore increasingly manages exceptions rather than routine transactions. This model works only when the underlying knowledge is dependable, because automating an incorrect inventory balance, obsolete policy, or broken workflow merely accelerates the error.

Greater leverage does not translate neatly into fewer people

If one employee can supervise significantly more automated execution, it is reasonable to expect some functions to operate with fewer people. The current evidence does not, however, support precise universal staffing ratios because several competing effects occur at the same time.

Automation reduces labor required for some work, but lower production costs also make previously uneconomic work possible. Companies can test more ideas, produce more variations, offer more personalized service, create more software, and support more customers. Every additional automated system also creates requirements for integration, quality control, governance, security, and exception handling.

The organization therefore does not necessarily shrink in direct proportion to tasks automated. It gains leverage, and management decides how to use that leverage. Some companies will reduce headcount, others will grow output without adding equivalent staff, and others may reinvest the capacity into new products, service levels, experimentation, or growth.

The research remains appropriately cautious: task automation should not be mechanically translated into job elimination, and net employment effects remain uncertain. The more defensible conclusion is that organizations capable of redesigning suitable workflows can increase the amount of execution each human directs.

The capacity inversion changes what is scarce

For most of business history, execution capacity was expensive because human time was expensive. Producing twice as many analyses, campaigns, software features, customer responses, or customized proposals usually required substantially more labor. AI begins weakening that relationship by making some forms of digital execution dramatically cheaper.

The evidence supports this at the task level, though not as a universal organizational multiplier. Controlled studies demonstrate significant improvements in some writing, support, consulting, marketing, and coding activities, while equally credible studies show slowdowns or quality problems when tasks fall outside the effective capability frontier.

The direction still matters because scarcity moves as execution becomes easier. The harder questions become what should be executed, why it matters, which information deserves trust, how quality will be measured, what machines may decide, where humans must intervene, and who remains accountable.

The research identifies a remarkably consistent set of capabilities rising in relative value: judgment under uncertainty, authority and evidence, trust, domain expertise, proprietary knowledge, originality, relationships, system design, integration, evaluation criteria, governance, and canonical knowledge. Routine first drafts, basic retrieval, boilerplate coding, administrative status aggregation, generic variants, manual formatting, information gatekeeping, and undifferentiated content become less scarce where AI can perform them to the required standard.

Executives ultimately care about outcomes, not AI activity

The changing economics of execution also expose a weakness in the way many companies currently discuss AI. Executives do not create lasting value by accumulating AI licenses, generating more assets, deploying more models, or announcing more pilots. Those activities matter only if they change business performance.

Senior decision-makers ultimately care about outcomes: growth, margin, speed, customer experience, operational control, resilience, risk, decision quality, and competitive advantage. Technology matters because it enables those outcomes, not because the technology itself is strategically interesting.

This distinction becomes especially important as virtually every provider begins describing its capabilities with the same vocabulary of AI, automation, personalization, analytics, and agents. The useful executive question becomes less about what technology a provider possesses and more about what becomes possible because that provider knows how to apply it.

The same logic should govern internal adoption. Companies should measure cycle time, quality, conversion, throughput, reliability, customer outcomes, cost, and risk rather than celebrating AI activity as if activity were transformation. The research specifically recommends tracking realized workflow outcomes rather than AI usage counts.

The biggest mistake may be automating what should have been redesigned

Every major technological shift goes through a stage in which new tools are used to preserve old processes. Early websites resembled printed brochures, early mobile sites compressed desktop experiences onto smaller screens, and early software often reproduced paper forms without questioning whether the original workflow made sense.

AI adoption is following the same pattern. Employees copy information into chat windows, generate responses, and copy them back into legacy systems. Meetings are summarized automatically even when the underlying meeting may not have been necessary. Marketing teams use AI to create more content inside publishing systems already overwhelmed with content, while developers accelerate coding without resolving poor requirements.

These uses can still deliver value, but they should be recognized as transitional. The larger opportunity emerges when organizations ask why information must be re-entered, why five departments maintain different versions of the same data, why a weekly report exists, why a customer must complete a particular form, or why a manager needs to ask for status instead of the workflow reporting its own state.

Those questions are not fundamentally about artificial intelligence. They are questions about business design that AI suddenly makes much harder to ignore.

An AI-native company is an operating model, not a tool stack

The phrase “AI-native organization” becomes meaningful only when it describes something deeper than widespread access to AI. A useful definition is an organization whose workflows, knowledge, controls, roles, and systems have been intentionally designed around the relative strengths and limitations of humans and machines.

Such an organization knows which information is authoritative, which tasks can be delegated, which decisions require approval, how automated performance is measured, where exceptions are routed, how actions are audited, and who remains accountable. It uses AI because a process is suited to AI rather than because leadership established an abstract adoption target.

The research is clear that most organizations have not reached that state. AI use is widespread, but deep workflow redesign and mature agent deployment remain substantially less common. That gap may represent one of the largest competitive opportunities entering 2027.

The advantage may not belong to the company that adopts AI first. It may belong to the organization that first learns how to redesign itself around what machines do well without surrendering the judgment, trust, creativity, relationships, and accountability that remain distinctively human.

Human accountability becomes more important as machine execution grows

The final paradox running through this series is that increasing automation does not eliminate human accountability. Someone still decides what the system should optimize, which information it may trust, what level of error is acceptable, which models or tools are authorized, when human approval is required, and how the company responds when the automated outcome is wrong.

Automation can move execution away from humans, but responsibility does not automatically move with it. Customers, employees, shareholders, regulators, and partners will still expect someone to answer for consequential decisions even when software performed much of the underlying work.

This may become one of the defining characteristics of high-value leadership in an AI-enabled organization. The leader will not necessarily be the person manually completing the greatest volume of work; the leader will be the person capable of directing a much larger amount of execution while understanding where human judgment belongs and accepting responsibility for the consequences.

That does not make the company after AI a less human organization by definition. It may instead make the value of human participation clearer by reducing the amount of time people spend moving information between disconnected systems and increasing the time available for judgment, relationships, creativity, leadership, and problem solving.

The three articles in this series converge on the same conclusion. The Internet is no longer just for humans, the information moving through it cannot be trusted merely because machines can read it, and a company does not become intelligent simply because employees have AI tools. Competitive advantage increasingly comes from connecting all three: establishing trusted knowledge, representing it clearly to people and machines, and converting it into governed execution that produces better outcomes.

The old systems are still functioning and should not be discarded simply because something new has arrived. But the new architecture is already emerging, and businesses now have a limited period in which they can learn to operate across both worlds deliberately. The question is no longer whether this transition will affect the organization. It is whether leadership will design for it intentionally or allow the transition to reshape the business by accident.

Sources