Randomized experiments show that AI can shorten some business tasks and make others worse, so a cost saving depends on which task is automated and who does it. In a Boston Consulting Group (BCG) experiment with 758 consultants, people using GPT-4 finished 12.2 percent more tasks and finished them 25.1 percent more quickly on tasks within the tool's abilities, yet on one task chosen to be beyond it they were 19 percentage points less likely to reach a correct answer. None of the studies below measured savings in dollars, so each result is a time or quality figure that still has to be priced.
Key takeaways
- In the BCG experiment, AI users were faster and produced better work on 18 tasks inside the tool's range, and worse on one task outside it.
- In a five-month trial with 640 Kenyan entrepreneurs, an AI business mentor produced no average gain, with stronger firms gaining and weaker firms losing.
- A study of 5,179 customer support agents found about 14 percent more issues resolved per hour, with little effect on the most experienced staff.
- Time saved is not cost saved: review, rework, tool fees, and error costs have to be counted.
Why time saved is not the same as money saved
Business cases for AI usually start from hours. A task that takes 25 percent less time looks like a 25 percent cost cut, but the hours only turn into savings if the freed time is used elsewhere, if headcount or overtime changes, and if the output does not need extra checking. Field experiments are useful here because they randomly assign who gets the tool, which separates the effect of AI from differences between workers. They still measure speed and quality, not invoices, and the participants are not a random sample of firms.
The consulting experiment: faster inside the frontier, worse outside it
The study, a paper on the jagged technological frontier by Dell'Acqua and colleagues (full text), was published in Organization Science on March 11, 2026. It was preregistered and covered 758 BCG consultants, about 7 percent of the firm's individual-contributor consulting workforce. After a baseline task, each person was randomly placed in one of three groups: no AI, GPT-4 access, or GPT-4 access with a short prompting overview.
On 18 realistic tasks that the authors judged to be within the tool's range, AI users completed 12.2 percent more tasks, finished 25.1 percent more quickly on average, and delivered work of significantly higher quality. Consultants in the bottom half of baseline skill gained the most, and the top half also improved. On a single managerial task designed to sit outside the frontier, the control group was correct about 84.5 percent of the time, against 60 percent and 70.6 percent in the two AI groups. The authors call the uneven pattern a "jagged frontier" and note that one such task is a limitation of the design. Several authors work at BCG, which also helped design the tasks, as the author affiliations in the paper show.
The entrepreneur trial: the same tool, different results by firm
A trial by researchers at the University of California, Berkeley's Haas School and Harvard Business School tested a GPT-4 business mentor delivered over WhatsApp. It ran for five months with 640 entrepreneurs in Kenya, whose businesses included fast food outlets, poultry farms, and cybershops. Half received a standard business guide and half received the AI mentor.
According to the Haas summary of the study, there was no positive effect on average. Entrepreneurs who were above the median at the start saw profits or revenue rise by about 15 percent, while those below the median saw revenue fall by about 8 percent. Both groups asked a similar number of questions, but weaker firms asked about harder problems such as drought, competition, and capital, where the AI's advice was less useful. For a small business, the lesson is that a cost-saving tool can lose money when applied to the wrong problem.
A third data point: customer support
The National Bureau of Economic Research paper by Brynjolfsson, Li, and Raymond followed 5,179 customer support agents as an AI assistant was introduced in stages. Issues resolved per hour rose about 14 percent on average, about 34 percent for novice and low-skilled workers, and little for experienced ones. The authors also report better customer sentiment and higher employee retention.
Pricing a time saving: an illustrative example
The figures below are invented for illustration and are not drawn from any study. They show how a 25.1 percent speedup can shrink once checking time is counted.
| Step | Hours |
|---|---|
| 100 tasks at four hours each, without AI | 400 |
| Same tasks with a 25.1 percent speedup | 299.6 |
| Gross time saved | 100.4 |
| Review of each AI output, half an hour per task | minus 50 |
| Net time saved before tool fees and rework | 50.4 |
Pricing the 50.4 net hours depends on the loaded hourly cost of the staff involved, and the result drops further if the tool carries a subscription or if some outputs must be redone.
Questions for an internal test
- Which tasks resemble the 18 inside-the-frontier tasks, and which resemble the one outside?
- Is there a comparison group, or only a before-and-after count?
- Are errors measured, and is anyone responsible for catching them?
- Do weaker and stronger performers get the same benefit?
- Is the saving converted into real cost, such as fewer overtime hours or more work completed?
FAQ
Do these studies prove AI cuts business costs?
No. They show time and quality effects for specific tasks and groups. Dollar costs were not measured, and the results do not automatically apply to other industries or to newer tools.
Why did AI make the consultants worse on one task?
The task was chosen by the researchers to fall outside what the tool could do reliably. The paper reports the drop of 19 percentage points in correct answers and draws the general point that AI help is uneven across tasks that look equally hard.
Does AI help beginners or experts more?
In the BCG and customer support studies, less experienced workers gained most. In the Kenya trial, stronger firms gained and weaker firms lost, so the answer depends on the task and the setting.
What should a small business measure first?
Time per task, error rate, and rework, for the same tasks done with and without the tool, before converting any of it into a cost figure.
Related reading
- The full cost of implementing AI
- AI and complex problem solving
- AI in cleaning services
- AI hardware and its costs
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Field experimental evidence of the effects of AI on knowledge worker productivity and quality (jagged frontier study), Dell'Acqua et al., Organization Science, 2026
- News article on the Kenya entrepreneur study (Otis, Clarke, Delecourt, Holtz, and Koning), Haas School of Business, UC Berkeley, April 1, 2024
- Generative AI at Work, Brynjolfsson, Li, and Raymond, National Bureau of Economic Research Working Paper 31161; Quarterly Journal of Economics, 2025