Research & Evidence
What Controlled Studies Really Tell Us About AI at Work
Three major workplace studies show real gains from AI—and why those gains are not universal.
Sources in this piece
03- Brynjolfsson, Li & Raymond, Generative AI at Work, Quarterly Journal of Economics (2025)
- Cui et al., The Effects of Generative AI on High-Skilled Work, Management Science (2026)
- Dillon et al., Shifting Work Patterns with Generative AI, NBER Working Paper (2025)
Full citations are listed at the end of the article. Every figure is reported with the original source and its caveats.
AI headlines often make work sound like a simple equation: add a tool, get a productivity boost.
Actual workplace studies tell a more interesting story. AI can help substantially, but the benefit depends on the job, the person, the tool and how it is introduced.
Here are three findings worth understanding before making a business decision.
1. Customer-support workers resolved more issues per hour
A study published in The Quarterly Journal of Economics in 2025 examined the rollout of a generative-AI assistant among 5,172 customer-support agents. Access to assistance increased the number of issues resolved per hour by 15% on average. Less experienced and lower-skilled agents saw larger gains in speed and quality, while the most experienced agents saw smaller speed gains and some quality trade-offs. [1]
The tool suggested responses during conversations; employees remained responsible for the interaction.
What it means: An AI assistant can help people access useful knowledge at the moment they need it. But an average across thousands of agents is not a promise that your sales team, accounting department or service desk will gain 15%.
2. Developers completed more tasks with an AI assistant
A paper published in Management Science in February 2026 combined three randomized field experiments involving 4,867 software developers. Developers granted access to an AI coding assistant completed 26.08% more tasks on average across the pooled analysis. The researchers noted variation and statistical uncertainty across experiments; less experienced developers generally adopted the tool more and benefited more. [2]
What it means: Carefully selected tools can remove friction from defined technical tasks. It does not mean every software project will ship 26% sooner—or that more completed tasks always equal better customer results.
3. Office workers spent less time on email—but organizations did not automatically transform
Another field experiment, reported by researchers in a 2025 NBER working paper, followed 7,137 knowledge workers at 66 firms using an AI tool embedded in their normal email, meeting and writing applications. During the second half of the six-month experiment, the 80% of treated workers who used the tool spent approximately two fewer hours each week on email and reduced after-hours work. Researchers did not detect broader shifts in the quantity or composition of tasks from individual access alone. [3]
What it means: Time savings are real and valuable, including for employees' work experience. But saving time and changing how an organization operates are different accomplishments.
The lesson hiding inside the numbers
Across these studies, the recurring theme is task fit. Support agents receiving relevant suggestions, developers getting code assistance, and office workers working inside familiar applications had clear opportunities to benefit. The intervention matched a specific kind of work.
It is tempting to ask, “Which AI tool is the best?” A more useful sequence is: What is the task? What information does it require? How will we judge accuracy? Who checks the output? Where does the work go next?
That sequence protects against another common mistake: measuring only speed. A faster process that produces inaccurate records, confusing customer messages or more rework is not necessarily an improvement.
A sensible way to apply the evidence
Start with a task that is frequent, bounded and easy to evaluate. Establish how long it takes today and what good output looks like. Pilot a tool with the people doing the work. Watch for quality, errors, employee experience and the impact on the following step.
If the pilot works, improve the surrounding process. If it does not, adjust or stop. Good technology decisions include the freedom not to automate.
At Opskind, we believe businesses deserve that kind of evidence-led adoption: practical, measured and designed around actual operations.
Next step: Explore Opskind's AI & Business Automation and Digital Adoption services if you want to identify a use case, test its value and support your team through the change.
Frequently asked question
How much productivity can AI add to a small business? There is no reliable universal percentage. Published results vary by task and setting. A small business should measure a relevant pilot rather than import an average from a different industry.
Where Opskind helps
If this sounds familiar, this is usually where we start.
Each service stands on its own, so you can read about the one that fits before deciding anything.
Keep reading
Related resources
Research & Evidence
The AI Productivity Paradox: Faster People, Same Business?
AI can make a person faster without making the business better. Here is what the latest evidence reveals.
The Business Case
You Do Not Need AI Everywhere. You Need It Where It Helps.
One of the smartest technology decisions a business can make is choosing what not to automate.
Next step
If this raised something specific, tell us about it.
If this piece raised something specific about your own operation, describe it in a few lines. We will reply with an honest view of what would help first.