How AI Is Changing the Meaning of Productivity
Artificial intelligence (AI) is technology that performs tasks associated with human capabilities such as understanding language, recognizing patterns, and generating content. It is changing the meaning of productivity: rather than measuring only how much a person produces, organizations increasingly need to consider how people and AI systems divide work, improve quality, and create value together. The effects are already visible in research: a 2023 study of customer-support agents found that access to a generative AI assistant increased issues resolved per hour by 14% on average, with larger gains among less experienced workers. Yet productivity gains depend on the task, worker, and workplace—not simply on using AI. This article examines AI’s effects on output, the changing mix of human and machine work, measurement challenges, risks, and practical ways to evaluate results.
AI productivity: From faster task completion to broader value
Productivity traditionally refers to the relationship between outputs and the inputs used to produce them. The Organisation for Economic Co-operation and Development (OECD) describes labour productivity as the amount of output produced per unit of labour input, commonly measured as output per hour worked. AI complicates this familiar measure because it can change the quantity, quality, and type of work produced, as well as the time workers spend on it.
A useful distinction is between task-level productivity—how quickly or effectively someone completes a particular activity—and organization-level productivity, which concerns the value generated across a firm or economy relative to all relevant inputs. AI may save minutes on drafting or data analysis without increasing overall productivity if employees spend that time correcting errors, handling more low-value tasks, or waiting for approvals. Conversely, AI can create value that simple output counts miss, such as clearer customer communication or faster access to expertise.
The evidence suggests meaningful but uneven benefits. In a controlled field study published in the Quarterly Journal of Economics, researchers Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond found that a generative AI assistant raised customer-support productivity by 14% on average, measured by issues resolved per hour. The effect was concentrated among novice and lower-skilled agents, while the most experienced workers saw little benefit. This points to a central feature of AI-enabled productivity: results depend on how well the system’s capabilities match workers’ needs and the task at hand.
Task productivity and workflow productivity
Task productivity measures performance on a discrete activity, such as summarizing a document, writing a first draft, or classifying a support request. Workflow productivity measures whether a sequence of connected activities becomes faster or better overall. A generated draft may shorten writing time, for example, but the complete workflow also includes fact-checking, revisions, approvals, and publication.
The distinction matters because AI tools can shift work rather than eliminate it. A 2023 study by Shakked Noy and Whitney Zhang, published in Science, found that participants using ChatGPT completed selected professional writing tasks 40% faster on average and produced work rated 18% higher in quality. Those results apply to the study’s specific tasks and participants; they do not establish that every writing workflow will see the same gains. Organizations should therefore measure end-to-end outcomes, not just the speed of the first AI-assisted step.
Economic productivity and the implementation gap
Economic productivity concerns output across a business, industry, or economy. Improvements at the individual task level may take time to appear in broader statistics because organizations must integrate tools into processes, train staff, manage risks, and redesign roles. This lag is sometimes discussed as a productivity paradox: promising technology can spread before its benefits become visible in aggregate measures.
The OECD’s work on AI and productivity emphasizes that adoption alone is not enough; complementary investments in skills, data, management, and organizational change help determine whether technology raises performance. The practical implication is that buying an AI tool is not itself evidence of higher productivity. The relevant question is whether the tool improves valuable output after accounting for implementation costs and the human work required to use it safely.
Human–AI collaboration: How work and expertise are redistributed
Human–AI collaboration is the arrangement in which people and AI systems contribute different capabilities to a shared task. AI can generate, retrieve, classify, or recommend; people can set goals, supply context, exercise judgment, and take responsibility for consequences. The division of work varies by occupation and task, so AI is better understood as changing some activities within jobs than as affecting all jobs uniformly.
The International Labour Organization’s 2023 analysis of generative AI and employment concluded that, for many occupations, automation and augmentation are both possible, with augmentation more likely to be the dominant effect in many cases. That does not mean every worker benefits equally. The distribution of gains depends on who controls the tools, how work is reorganized, whether workers receive training, and whether saved time translates into better outcomes or simply higher workloads.
Automation: Delegating bounded, repeatable tasks
Automation occurs when a system performs a task with limited human intervention. AI can automate or partially automate activities such as sorting routine requests, extracting information from forms, and generating standard summaries. These uses are most suitable when inputs and acceptable outputs are clear, errors are detectable, and the consequences of mistakes are manageable.
Automation can free people to focus on complex work, but it can also create new oversight duties. A system that drafts a response may reduce initial writing time while requiring a person to verify accuracy, privacy, and tone. Productivity assessment should include both the work delegated to the system and the monitoring or correction work that remains.
Augmentation: Extending human capability
Augmentation means using AI to assist rather than replace a person’s contribution. Examples include asking a model to explain unfamiliar code, generate alternative approaches, or summarize research that a professional then evaluates. In the customer-support study by Brynjolfsson, Li, and Raymond, the larger gains for less experienced agents suggest that AI assistance can help spread effective practices and make expertise more accessible.
Augmentation works best when users can judge the system’s output and retain authority over consequential decisions. It may be less effective when a task requires tacit knowledge, sensitive interpersonal judgment, or information the system cannot access. The goal is not to maximize the number of AI-generated steps, but to improve the combined performance of the person–tool system.
Skill change and the novice–expert divide
AI changes skill requirements by lowering the effort needed for some tasks while increasing the importance of others, including problem framing, verification, domain knowledge, and communication. This can help novices perform more effectively, but it can also make it harder for them to develop expertise if they rely on generated answers without understanding the underlying work.
The uneven results in the customer-support field study illustrate why productivity should be examined across experience levels, not only as an average. Managers can pair AI access with training, review processes, and opportunities to practice unaided skills. That helps ensure that short-term gains do not come at the cost of long-term capability.
Productivity measurement: What AI gains do—and do not—count
AI productivity measurement is the structured assessment of how AI affects output, quality, time, costs, and other outcomes relative to the resources used. A sound evaluation compares the AI-assisted process with a credible baseline and accounts for implementation, verification, training, and maintenance. It should also identify who benefits and whether performance changes are sustained.
No single metric captures every form of value. Output per hour may be useful for a stable, repeatable process, while a professional service may also require measures of accuracy, client satisfaction, rework, and risk. The United States Bureau of Labor Statistics notes that productivity indexes compare output with input over time; for AI adoption, organizations often need more detailed operational measures alongside those broader indicators.
Output, quality, and rework
Output measures the quantity of completed work, while quality measures whether that work meets relevant standards. Rework captures the effort needed to correct or redo it. Tracking all three can reveal whether faster production represents genuine improvement or merely shifts effort to later stages.
For instance, an AI-assisted document process could be evaluated by drafts completed per hour, factual error rates, reviewer edits, and time to final approval. The Science study by Noy and Zhang is notable because it assessed both task time and evaluator-rated quality, rather than treating speed alone as productivity. Its findings show the value of using more than one outcome measure, while remaining specific to the tasks tested.
Total cost, adoption, and realized value
Realized productivity value is the improvement that remains after accounting for the full cost of using AI. Costs may include subscriptions, computing, integration, staff training, data preparation, security controls, and human review. An attractive demonstration may not translate into net gains if the tool is expensive to maintain or requires extensive correction.
Organizations should distinguish between potential productivity—the benefit a system could deliver under ideal conditions—and realized productivity, which reflects actual use at scale. Adoption rates, task suitability, reliability, and workflow fit all affect the gap between the two. A small pilot can identify promise, but long-term tracking is needed to determine whether improvements persist.
Worker and customer outcomes
Productivity is not only a matter of organizational output. Worker outcomes include workload, autonomy, learning, and job quality; customer outcomes may include response times, accuracy, accessibility, and satisfaction. These measures help show whether efficiency gains improve service or simply intensify work.
A balanced evaluation can combine operational measures with surveys, error audits, and feedback from affected groups. Results should be compared across roles and experience levels, since an average gain can conceal losses for particular workers or customers. This is especially important when AI changes performance expectations or decisions about staffing.
AI productivity risks: Reliability, fairness, and work intensification
AI productivity risks are the potential harms or hidden costs that reduce the value of AI-assisted work. These include inaccurate output, bias, privacy breaches, security vulnerabilities, overreliance, and pressure to increase work pace. Risk management is not separate from productivity: errors and harmful decisions can create direct costs, damage trust, and undermine the quality of output.
The National Institute of Standards and Technology’s AI Risk Management Framework organizes risk work around governing, mapping, measuring, and managing AI risks. For employers, that approach supports a practical principle: evaluate the system in its actual context, set clear human responsibilities, and monitor performance after deployment rather than assuming that a successful test guarantees safe or productive use.
Errors, verification, and accountability
Generative AI can produce fluent but incorrect information, making verification essential in tasks where mistakes matter. The appropriate level of review depends on the stakes: a brainstorming suggestion may need less scrutiny than a medical, legal, financial, or employment-related recommendation. Organizations should specify which outputs require human approval and who is accountable for the final decision.
Verification time belongs in the productivity calculation. If a system produces a result quickly but requires extensive checking, its net advantage may be small. If reviewers are unable to identify plausible errors, the risk can exceed any measured time saving.
Bias, privacy, and security
Bias can arise when training data, system design, or workplace use produces systematically different outcomes for different groups. Privacy and security risks can emerge when users submit confidential, personal, or proprietary information to tools without appropriate safeguards. These issues can damage people and organizations while creating compliance, remediation, and reputational costs.
Practical safeguards include approved-tool lists, data-handling rules, testing across relevant user groups, access controls, and regular audits. These controls should be proportionate to the task and its consequences. They can support productivity by making responsible use dependable rather than leaving workers to guess what is permitted.
Work intensification and uneven distribution of gains
Work intensification occurs when efficiency tools lead to higher expectations, faster pacing, or additional tasks without a corresponding improvement in working conditions. AI may allow an organization to handle more requests, but that does not automatically mean employees experience less pressure or share in the gains.
The International Labour Organization’s analysis highlights that generative AI’s effects vary by occupation and task. Employers can make distribution visible by tracking workload, hours, staffing, training access, and worker feedback alongside output. Consultation with employees can also help identify where automation removes tedious work and where it creates new demands.
Putting AI productivity into practice: A measured approach
A measured approach treats AI adoption as a work-design and evaluation process, not simply a software rollout. It begins with a specific problem, tests whether AI is appropriate, and compares results against a baseline. This reduces the chance of mistaking novelty or increased activity for durable productivity improvement.
Start with the task and establish a baseline
Define the task, its inputs and outputs, the people involved, and the cost of errors. Record current performance before introducing AI, including time, volume, quality, rework, and relevant user outcomes. Select tasks where the system’s strengths fit the work and where results can be checked.
Run a controlled pilot and measure the whole workflow
A pilot should compare AI-assisted work with the existing process under reasonably similar conditions. Measure the full sequence, including prompt preparation, review, corrections, escalation, and final delivery. Where possible, examine results by task type and worker experience so that differences are not concealed by a single average.
The studies by Noy and Zhang and by Brynjolfsson, Li, and Raymond offer useful examples of task-specific evaluation: both examined measurable work outcomes, and the latter identified different effects by worker experience. Their results are evidence for careful testing, not universal forecasts for every organization.
Train workers, set guardrails, and redesign roles
Training should cover tool capabilities, limitations, verification, privacy, and escalation. Clear guardrails should identify approved uses, restricted data, required reviews, and decision owners. As teams learn what works, workflows and roles may need redesign so that people focus on judgment, service, and tasks where human knowledge adds the most value.
Review results and share gains responsibly
After deployment, review performance regularly for accuracy, cost, adoption, security, and worker and customer impacts. Update or stop a use case if benefits fail to outweigh its costs or risks. Organizations should also decide how gains will be shared—through improved service, reduced repetitive work, worker development, shorter turnaround times, or other clearly stated outcomes.
Conclusion: Productivity becomes a human–AI outcome
AI is changing productivity from a narrow focus on individual speed toward a broader assessment of output, quality, workflow, costs, and human outcomes. Evidence from professional writing and customer support shows that AI can improve performance on defined tasks, but the size and distribution of gains vary. Automation and augmentation redistribute work differently; economic gains depend on implementation; and measurement must include verification, rework, risk, and worker experience.
The most useful next step is to choose a specific workflow, establish a baseline, run a measured pilot, and evaluate the full human–AI process. Organizations, workers, and policymakers should ask not only whether AI makes work faster, but whether it makes valuable work better, safer, and more sustainable—and who benefits when it does.
Sources: Organisation for Economic Co-operation and Development, “GDP and productivity,” https://www.oecd.org/sdd/productivity-stats/; Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond, “Generative AI at Work,” The Quarterly Journal of Economics, 2025, https://academic.oup.com/qje/article/140/2/889/7990658; Shakked Noy and Whitney Zhang, “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence,” Science, 2023, https://www.science.org/doi/10.1126/science.adh2586; Organisation for Economic Co-operation and Development, “The impact of Artificial Intelligence on productivity, distribution and growth,” https://www.oecd.org/economy/the-impact-of-artificial-intelligence-on-productivity-distribution-and-growth.htm; International Labour Organization, “Generative AI and Jobs: A global analysis of potential effects on job quantity and quality,” 2023, https://www.ilo.org/publications/generative-ai-and-jobs-global-analysis-potential-effects-job-quantity-and; U.S. Bureau of Labor Statistics, “Productivity,” https://www.bls.gov/productivity/; National Institute of Standards and Technology, “Artificial Intelligence Risk Management Framework (AI RMF 1.0),” 2023, https://www.nist.gov/itl/ai-risk-management-framework.
