AI in the Workplace: Jobs, Productivity, Skills, and the New Operating Model

AI in the workplace is changing tasks faster than it is eliminating entire occupations. Companies are using generative AI to draft, search, summarize, analyze and support decisions; more advanced systems can also carry out multi-step workflows. The near-term result is not a single wave of replacement. It is a broad redesign of who does each task, which decisions require human judgment and how organizations measure useful work.

The evidence is mixed in an important way. Worker-level studies show meaningful gains in selected tasks, while economy-wide productivity is improving much more gradually. Employment projections point to strong growth in several technical and analytical occupations alongside declines in some routine office and service roles. At the same time, exposure to AI does not prove that a job will disappear.

This guide explains what AI at work actually means, what the latest labor and productivity data can and cannot tell us, which skills matter and how the operating model changes when teams use both people and AI systems.

Why It Matters

AI adoption is becoming a management and workforce question, not only a technology purchase. The decisions companies make now—what to automate, what to augment, what to measure and where to keep human review—will influence productivity, hiring, job quality and risk.

For workers, the relevant question is usually not “Will AI take my job?” but “Which parts of my job are changing, and what new responsibilities follow?” For leaders, the question is not whether an AI tool can produce an impressive demo. It is whether the tool improves a real workflow after implementation costs, errors, oversight and employee adoption are included.

That distinction matters because public discussion often mixes four different signals: what AI can theoretically do, what workers are actually using it for, whether output has increased and whether employment has changed. Those measures are related, but they are not interchangeable.

What AI in the Workplace Actually Means

Workplace AI is an umbrella term for software that recognizes patterns, generates content, recommends decisions or takes actions within a business process. It ranges from narrow systems that classify documents to generative tools that produce text and code, and to agents that can use software tools across several steps.

Most workplace deployments fall into four practical categories:

• Decision support: AI organizes information, detects patterns or recommends an option while a person makes the decision.
• Content and knowledge assistance: generative AI drafts, summarizes, searches, translates or explains information.
• Task automation: software completes a bounded, repeatable task such as routing a request or extracting fields from a document.
• Workflow execution: an AI agent plans and completes multiple steps using connected tools, with human approval at defined checkpoints.

The unit of change is usually the task, not the job title. A customer service representative may use AI to retrieve account information and draft a response while still handling exceptions and sensitive conversations. A software developer may use AI to generate routine code while remaining responsible for architecture, security and review. A manager may receive an AI-generated analysis but still own the decision and its consequences.

This task-level view avoids two common mistakes: assuming that any exposure means full automation, and assuming that a job is unaffected because some of its duties cannot be automated.

How Much AI Is Actually Being Used at Work?

Adoption is rising, but the answer depends on what is being measured. The OECD’s 2026 skills report says reported business use of AI in countries with available data increased from roughly 7 percent in 2021 to 20 percent in 2025. The same report notes that larger firms are generally more likely to adopt AI, while smaller businesses often face barriers involving cost, expertise and infrastructure.

Household surveys capture another layer. The NBER study on rapid generative AI adoption found that use spread quickly after consumer tools became available. But use on some workdays does not mean that a tool is embedded across an entire company or that it controls a complete workflow.

Anthropic’s 2026 labor-market research offers a third measure: “observed exposure.” It combines occupational tasks from O*NET, theoretical estimates of what large language models can speed up and anonymized patterns of Claude usage. The resulting measure asks which theoretically feasible tasks are actually appearing in professional AI use.

The Anthropic analysis found that computer programmers had the highest observed coverage, at about 75 percent of tasks. Data entry keyers were at 67 percent, while 30 percent of workers were in occupations with no task coverage above the study’s threshold. Yet the authors also found no clear rise in unemployment among the most exposed occupations.

These figures should be read as signals, not a census of all workplace AI. They are partly based on one provider’s usage, depend on task classifications and do not show that every covered task is fully automated. They are useful because they narrow the gap between capability and actual use, but they do not by themselves measure jobs lost or value created.

How AI Is Changing the Job Market

The U.S. labor market is not moving in one direction. AI can reduce demand for certain tasks, raise demand for complementary skills, create new tasks and lower costs in ways that increase demand elsewhere. All four effects can occur within the same occupation.

Exposure Is Not the Same as Replacement

An occupation can be highly exposed because AI can assist with many of its tasks, while employment still grows. Exposure describes overlap between a technology and job activities. Replacement depends on reliability, cost, demand, regulation, workflow design and whether the remaining human tasks are essential.

The OECD’s analysis of AI and jobs found that high-skilled occupations were among those most exposed to recent AI progress, yet those workers had experienced employment gains relative to lower-skilled workers over the preceding decade. The OECD cautions that exposure can lead to augmentation, automation or new tasks; it is not a probability of job loss.

Where Employment Is Projected to Grow

The BLS 2024–2034 projections estimate that total U.S. employment will increase by about 5.2 million jobs, or 3.1 percent. In AI- and information-technology-related work, BLS projects employment growth of 33.5 percent for data scientists, 28.5 percent for information security analysts and 15.8 percent for software developers.

These projections do not mean AI is the only cause of growth. Demand for software, cybersecurity, data services and digital infrastructure has multiple drivers. They do show that the same technological shift associated with automation can also increase demand for workers who build, secure, evaluate and apply digital systems.

Where Employment Is Projected to Decline

BLS expects AI-related productivity gains to limit demand in some routine office and administrative work. Its AI and employment projection summary projects customer service representative employment to decline 5.5 percent, or about 153,700 jobs, between 2024 and 2034. It also projects declines for procurement clerks, credit authorizers and several administrative-support occupations.

Even a declining occupation can still have many openings as people retire or move to other work. BLS projects hundreds of thousands of annual openings for customer service representatives from replacement needs. A negative employment projection therefore does not mean there will be no hiring.

What the Early Evidence Says

Anthropic found a slight relationship between observed AI exposure and weaker BLS job-growth projections: each 10-percentage-point increase in coverage was associated with a 0.6-percentage-point reduction in projected growth. That is a correlation between two measures, not proof that AI caused BLS to lower a forecast.

The more defensible conclusion is that hiring and job design may change first in occupations with many digital, language-based tasks. The timing and scale remain uncertain, especially for entry-level work, where AI may remove some routine assignments that historically helped new workers learn.

What the Productivity Evidence Actually Shows

AI productivity should be evaluated at three levels: the worker, the organization and the overall economy. Strong results at one level do not automatically appear at the next.

Worker-Level Gains Can Be Real but Narrow

In a large customer-support field study, access to a generative AI assistant increased issues resolved per hour by about 14 percent on average. The NBER research found larger improvements for less-experienced workers, suggesting that AI can help distribute useful practices and reduce the time needed to reach proficiency.

That result applies to a particular workflow with measurable outputs and access to organizational knowledge. It does not establish a universal productivity rate for every role. Gains may be smaller when tasks are ambiguous, errors are costly, source information is weak or workers spend substantial time checking outputs.

Company-Level Value Requires Workflow Change

A worker may draft a document faster without changing the organization’s output. The saved time creates value only if it is redirected to additional work, faster service, better quality or lower cost. Companies also have to account for licenses, integration, data preparation, training, monitoring, security and rework.

This is why adoption data and return on investment can diverge. A tool can be widely available but lightly used. It can save time without increasing capacity. It can also improve speed while creating new review work. Measuring generative AI ROI therefore requires a baseline, a defined business outcome and the full cost of operating the system—not only a count of users or prompts.

Economy-Wide Productivity Moves More Slowly

The BLS preliminary second-quarter 2026 release reported that nonfarm business labor productivity was 2.2 percent higher than one year earlier, with output up 2.5 percent and hours worked up 0.2 percent. Manufacturing productivity increased 0.9 percent over the same period.

Those figures measure output per hour across broad sectors. They do not isolate AI’s contribution. National productivity also reflects capital investment, industry mix, management practices, demand and many other technologies. It can take years for companies to redesign processes and spread new tools widely enough to affect aggregate statistics.

The micro-versus-macro gap is therefore not evidence that AI has no value. It is evidence that a successful experiment is only the beginning of organizational change.

The Skills Workers Need in an AI-Augmented Economy

Most employees do not need to become machine-learning engineers. They do need enough AI literacy to understand what a tool is doing, recognize when it may fail and use its output responsibly within their field.

The most transferable skill groups are:

• AI and data literacy: understanding inputs, outputs, limitations, privacy and basic measurement.
• Domain expertise: knowing what a correct, useful and compliant result looks like in a specific job.
• Verification: checking sources, calculations, assumptions and generated content before action.
• Workflow judgment: deciding which tasks to automate, which to augment and which to keep human-led.
• Communication and collaboration: giving context, explaining decisions and working across technical and operational teams.
• Adaptability: learning new tools while retaining responsibility for outcomes.

The World Economic Forum’s Future of Jobs Report 2025 found that 63 percent of surveyed employers considered skill gaps a major barrier to business transformation, while 85 percent planned to prioritize upskilling. The figures reflect employer expectations rather than guaranteed outcomes, but they show why implementation cannot be separated from workforce development.

Training should be role-specific. A finance team needs different examples, risk controls and evaluation criteria than a marketing or engineering team. Generic prompting lessons are unlikely to produce durable value without practice on real workflows and clear standards for review.

The New Operating Model: Copilots, Agents and Human Oversight

The operating model changes as AI moves from suggesting work to taking action. The difference is less about labels and more about authority.

Copilots Assist a Person

A copilot helps a worker draft, search, summarize or analyze. The person initiates the task, evaluates the output and decides what happens next. This model is appropriate when judgment is central, the process changes frequently or the cost of an incorrect action is high.

Agents Execute Bounded Workflows

An agent can pursue a goal across several steps, use connected software and adapt when a step fails. Useful agents operate within defined permissions, escalation rules and monitoring. They are not simply chatbots with a new name.

Morning Glance’s guide to AI agents explains the planning, tool use, memory and adaptation that distinguish agents from basic assistants.

The Best Design Is Usually Hybrid

Organizations do not have to choose between full manual work and full autonomy. A hybrid workflow can let AI gather information, prepare a recommendation or execute low-risk steps while people approve exceptions and consequential decisions.

The design should answer five questions:

• What outcome is the workflow meant to improve?
• Which steps can the AI perform reliably?
• What data and tools may it access?
• Which decisions require human approval?
• How will errors, overrides and results be recorded?

The NIST AI Risk Management Framework organizes governance around mapping, measuring and managing risk. For generative AI, NIST emphasizes that different uses may require different levels of human review, documentation and oversight.

For a closer look at permissions, data exposure and failure modes, see Morning Glance’s guide to AI agent security risks.

A Real-World Example: Klarna’s AI Assistant

Klarna’s customer-service deployment shows both the potential and the limits of company-reported AI results. In 2024, the company said its OpenAI-powered assistant handled 2.3 million conversations in its first month, equal to two-thirds of customer-service chats. Klarna reported work equivalent to 700 full-time agents, an average resolution time reduced from 11 minutes to 2 minutes and customer-satisfaction results comparable to human agents.

Those numbers come from Klarna’s own announcement, not an independent audit. They demonstrate that a high-volume, digital workflow can be automated at scale, but they do not reveal every implementation cost, exception or long-term employment effect.

The broader lesson is not that every support team can reproduce the result. Klarna had a narrow use case, large conversation volume, connected account systems and clear metrics. Companies evaluating a similar deployment should compare quality, escalation rates, resolution time, cost per case and customer outcomes before and after implementation.

What Leaders and Workers Should Do Now

The practical response is to redesign work deliberately rather than pursue automation as an end in itself.

1. Map tasks, not just job titles. Break a workflow into research, judgment, communication, execution and review. Use O*NET or internal process documentation to identify what people actually do.
2. Select a measurable use case. Start where volume is meaningful, outcomes are observable and errors can be detected. Define the baseline before introducing the tool.
3. Keep humans at the right control points. Require approval for high-impact decisions and build clear escalation paths for uncertainty, exceptions and complaints.
4. Measure the complete result. Track time, throughput, quality, rework, adoption, risk incidents and operating cost. Do not treat logins or generated outputs as business value.
5. Train for the workflow. Teach employees how to use, verify and challenge the system in the context of their own responsibilities.
6. Review job quality. Monitor whether AI reduces low-value work or instead increases surveillance, work intensity and cognitive load.
7. Update governance as use expands. Permissions that are appropriate for drafting may be inadequate when the same system can send messages, change records or move money.

Workers can use the same task-level approach. Identify repetitive work that AI can accelerate, strengthen the domain knowledge needed to evaluate outputs and keep examples of higher-value contributions—such as decisions, relationships, problem-solving and improvements to the process.

What Happens Next

The next phase of workplace AI will be shaped less by access to models and more by implementation. Many companies already have similar underlying tools. The advantage will come from better data, clearer workflows, stronger evaluation, appropriate oversight and employees who understand both the technology and the work.

Employment effects are likely to remain uneven. Digitized occupations with repeatable language or data tasks may change sooner, while physical work and roles built around responsibility, trust or complex human interaction may change more slowly. Even within one occupation, some tasks will be automated, some augmented and some newly created.

The most useful indicators to watch are not model announcements alone. Watch adoption inside real workflows, measured quality, entry-level hiring, occupational transitions, training participation and whether productivity gains persist after full operating costs are included.

The Bottom Line

AI in the workplace is already changing how work is organized, but the evidence does not support a simple story of universal replacement or instant productivity. The clearest near-term shift is task redesign: AI handles more drafting, retrieval, analysis and routine execution while people remain responsible for context, judgment, exceptions and outcomes.

Companies that measure real business results, train workers and set clear boundaries are more likely to capture value. Workers who learn to direct and verify AI while deepening their domain expertise will be better positioned as roles evolve.

For continuing coverage of workplace automation, labor data and new AI systems, follow Morning Glance’s AI & Automation coverage.

Get the Morning Glance briefing — AI at Work context without the noise.

FAQs

What does AI in the workplace mean?

AI in the workplace means using artificial intelligence to support decisions, generate or organize information, automate bounded tasks or execute parts of a workflow. Most deployments change a collection of tasks rather than replacing an entire occupation at once.

Will AI replace jobs or mainly change them?

It will do both, depending on the occupation and workflow. Some routine roles are projected to decline, while technical and analytical roles are projected to grow. Exposure alone does not determine the outcome because demand, cost, reliability, regulation and the importance of remaining human tasks also matter.

Does AI make workers more productive?

It can. Controlled and field studies have found meaningful gains in selected tasks, including customer support. Results vary by use case, and company-wide value depends on adoption, workflow redesign, output quality, oversight and the full cost of the system.

What skills are most important for working with AI?

Workers need AI and data literacy, domain knowledge, verification skills, workflow judgment, communication and adaptability. Advanced coding is important for specialist roles, but most employees will create value by applying and evaluating AI within their existing field.

What is the difference between an AI copilot and an AI agent?

A copilot assists a person who remains in control of each task. An agent can plan and execute multiple steps using connected tools within defined permissions. Both require oversight, but agents need stronger access controls, monitoring and escalation rules because they can take actions.

How should a company begin using AI at work?

Start with a measurable, bounded workflow rather than an entire job. Establish a baseline, define the human approval points, test quality and risk, train the affected employees and compare the complete operating result before expanding.