Generated by Codex with GPT 5.6 Sol XHigh

This summary covers The Economist’s August 22nd, 2026 Bartleby column, listed in the contents as On AI and ROI and published under the headline Returns policy.

Corporate enthusiasm for artificial intelligence has moved from “use it everywhere” to a harder question: is it creating value? Early adoption campaigns often rewarded visible activity, such as token consumption or rankings of the heaviest users. Those measures can help establish whether employees are experimenting with a new technology, but they say little about whether the work is faster, better or more profitable.

The article argues that companies need a balanced set of measures covering three things: use of the technology, the outcomes it produces and the organisation’s ability to adapt. None is sufficient alone. The challenge is not to find one perfect return-on-investment figure, but to understand how AI changes work over time.

Usage is a starting point

Input measures still have a role while a company is trying to encourage adoption. Different jobs will consume different amounts of AI, but persistent non-use may show that employees lack suitable tools, training or incentives. Usage data can therefore reveal whether an experiment is reaching the people it is intended to help.

Yet activity is not value. AI may save employees an hour or two without improving the company’s results; the freed time might simply become personal leisure. A surge in output can also hide a fall in quality. One study found that a health-and-nutrition data set produced an average of four academic papers a year from 2014 to 2021, but 190 papers in the first nine months of 2024 after generative AI became widely available. More papers are not necessarily more useful science.

The same warning applies to software development. Google evaluates engineering work through speed, ease and quality. A shorter code-review cycle or smoother onboarding process is valuable only if the resulting software remains sound. Counting prompts, tokens or completed tasks without checking the standard of the output rewards motion rather than progress.

Returns arrive unevenly

Outcome measures should track changes that matter to the business, such as team productivity, customer satisfaction, cost, revenue and quality. Designing them is difficult because AI’s effects can spread across a workflow. A faster team may create a bottleneck elsewhere, while small time savings may be dispersed among many employees and never appear clearly in company accounts.

Benefits may also take time to emerge. Workers must learn new tools, managers must redesign processes and established performance systems may stop fitting the new way of working. Research on AI adoption by American manufacturers points to a “J-curve”: productivity can fall during the disruptive learning period before it improves. A short-term assessment may therefore reject a useful technology just as the organisation is beginning to absorb it.

Even the meaning of a financial return requires a choice. A company can turn saved time into immediate cost reductions by cutting jobs, but that can damage morale and remove people who could have been redeployed to more productive work. Another approach is to measure how much hiring AI allows the company to avoid as it grows. Revenue gains are harder to attribute and require credible baseline data, controlled comparisons and a culture of A/B testing.

Measure the capacity to learn

Because AI is evolving quickly, firms also need organisation-based measures. These might track whether employees are satisfied with implementation, whether managers are removing workflow obstacles and whether the company is building expertise in areas that AI is unlikely to automate. Such indicators do not prove that an investment has paid off, but they reveal whether the business is becoming capable of using the technology well.

The practical lesson is to treat measurement as a portfolio. Usage metrics show whether adoption is happening. Outcome metrics test whether it improves performance. Organisational metrics show whether the company can keep learning as the technology changes. Focusing only on usage permits expensive activity without accountability; demanding immediate, rigid returns can halt experimentation before its value becomes visible. Good measurement keeps both discipline and learning in view.