Why Buying AI Licenses Does Not Make a Company More Productive

When a company buys access to generative AI tools, employees can use software that drafts, summarizes, sorts information into categories, or answers questions. The license gives employees a tool to use at work. Productivity measures how much a company produces for the labor, time, and other resources it spends. Buying licenses alone does not show that the company produces more with those resources.

AI tools improve performance on some tasks. Company-wide gains depend on whether employees use them for suitable work, can check the results, and use any time saved to improve the business. To judge a purchase, compare the full cost and effort of doing real work with and without AI. Then check whether AI improves output or quality. Also check for improvements in service and cost.

Faster tasks do not automatically mean higher productivity

A tool can help an employee finish one task faster without increasing the company’s total output. The employee might use the saved time to handle more work, improve quality, reduce a backlog, or finish the same workload with less overtime. Each outcome has value, but only some directly increase measured output per unit of labor.

Research shows that AI can produce substantial gains on specific tasks. In an experiment involving professional writing, participants using ChatGPT completed tasks 40% faster and received output quality ratings 18% higher than participants without it. A separate study of customer-support agents found that AI access increased issues resolved per hour by 14% on average, with larger gains among novice and lower-skilled agents. Those results apply to specific tasks and study conditions. They do not predict what will happen when a company buys licenses for employees across the business. See the writing-task experiment and the customer-support study.

A field experiment across 66 firms and 7,137 knowledge workers shows how task-level results can differ from company-wide results. Workers with access to generative AI spent less time on email, but researchers detected no change in the quantity or mix of their tasks. The study tracked how workers spent their time, leaving unanswered whether companies used the time saved to produce more, lower costs, or increase revenue. A local time saving alone therefore does not establish a company-wide productivity gain. (The six-month field experiment)

AI helps when the task fits its capabilities

Employees need to know which work suits the tool and where its output needs close review. In a study of 758 junior consultants using GPT-4, participants completed tasks within the model’s demonstrated capabilities faster and with better results. On a task outside those capabilities, AI users performed 23% worse than participants without GPT-4. The researchers called the boundary between tasks a “jagged frontier”: the tool’s competence changes from one kind of task to another, and that boundary is not always obvious to the user. (The GPT-4 consulting experiment)

Because the tool performs unevenly across tasks, a company needs to choose which work employees give it. An AI assistant might prepare a first draft or summarize a set of documents, while an employee checks facts, handles exceptions, and approves the final result. A company should test the tool on representative tasks and include review and correction time in the comparison. A faster first draft is not a gain if the reviewer spends the saved time finding errors or rebuilding the work.

Employees need a working method, not just access

A license does not tell employees which information to use, how to prompt the tool, or what to do when its answer conflicts with company rules. Managers and the people responsible for work procedures need to decide which tasks use AI, what information employees may provide, and who checks important results. Employees also need access to current business information and enough training to recognize when an answer needs verification.

A U.S. Census Bureau study of businesses found that firms adopting AI often made other changes as well, including training staff, developing new workflows, investing in cloud services or storage, and changing data-management practices. The study documents these accompanying changes but does not quantify how much each contributes to productivity. Buying access still leaves companies to make these supporting decisions. (The Census Bureau study of firm AI use)

For example, a company that wants employees to use AI to draft customer responses should specify which records the tool can draw on and who approves replies involving unusual terms or sensitive cases. Without those choices, employees may avoid the tool, use it inconsistently, or spend extra time checking whether its drafts are safe to send.

Measure the result across the whole task

A company should compare all the work needed to produce an acceptable result. Logins, prompts, and draft speed do not prove that productivity improved. Count the time employees spend preparing information and checking answers. Include the time they spend correcting mistakes and handing work to colleagues. Then choose a business measure that matters for the task. Options include completed cases and turnaround time, as well as quality and cost.

To test AI, start with a task employees do regularly. Record what they produce, how long it takes, and the quality standard the result must meet. Then compare those results with AI-assisted work under normal review conditions. If employees finish faster but the company sees no change in output, service, quality, or cost, the license has not yet demonstrated a productivity gain.

Reported business results call for the same care. A 2026 survey of executives in the United States, United Kingdom, Germany, and Australia found that 89% reported no AI-related labor-productivity effect over the previous three years. That figure comes from executives’ survey responses, not an audit that measures cause and effect. It reports what businesses said they experienced. It does not prove that AI tools cannot improve performance. (The 2026 executive survey summary)

What I'm building

Delegate tasks. Get software.

Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.

Take a look at vroni.com