AI productivity research: task gains most firms haven't turned into profit

AI productivity research finds large gains on well-defined tasks and small gains across whole companies so far. Support agents with an AI assistant resolved 15% more issues per hour¹. Consultants handed a problem beyond GPT-4's abilities got it right less often with AI than without it². Across the U.S. economy, workers' self-reported savings came to 2.2% of work hours in mid-2026³. Staff at firms that redesign the work around the tool report business results more often than staff at firms that only hand out licences⁴.

What AI productivity research found in controlled trials

Each study below compared people with and without AI on the same kind of work and measured what changed.

• 453 college-educated professionals given ChatGPT finished job-specific writing tasks in 40% less time, and graders rated the work 18% higher⁵.
• In randomized trials at Microsoft, Accenture and a Fortune 100 company, 4,867 developers with an AI coding assistant completed 26% more tasks⁶.
• 758 BCG consultants using GPT-4 on tasks inside its abilities completed 12.2% more tasks, finished 25.1% faster and got quality grades 32% higher².
• 5,172 support agents at a Fortune 500 company resolved 15% more issues per hour with an AI chat assistant¹.

The bars measure different things, so they can't be averaged into one figure. None of the studies ran inside a small or mid-sized North American business, so applying them to one is an extrapolation.

Newer and less-skilled staff gain the most

In the support study, less-skilled and newer agents resolved about 30% more issues per hour, while the most experienced agents saw small speed gains and small drops in quality¹. Less experienced developers used the coding assistant more and gained more⁶. MIT Sloan's write-up of an earlier version of that paper put the gain at 27% to 39% for recent hires and juniors and 8% to 13% for senior developers⁷.

The BCG study found the same pattern. Consultants who scored below average on a baseline task improved 43%, and those above average improved 17%². ChatGPT also narrowed the gap between weaker and stronger writers⁵.

The jagged frontier: where AI makes work worse

The BCG study's authors call the uneven edge of AI's abilities the jagged frontier. AI improves some tasks and worsens others, even inside one workflow and at similar difficulty². To test the far side, they gave consultants a task picked to sit outside GPT-4's abilities: recommending which of a client's brands had the most growth potential, using interview notes and financial data².

Graders still rated the AI users' recommendations as better argued, so the wrong answers looked convincing².

In a 2025 trial by the research nonprofit METR, 16 experienced open-source developers took 19% longer on tasks when AI was allowed, and afterward they believed it had sped them up by 20%⁸. METR has since marked those results out of date. Its early-2026 follow-up found signs of a speedup, which METR calls very weak evidence⁸⁹.

The same problem shows up in company knowledge tools, covered in how AI knowledge assistants work and where they break.

Does AI increase productivity across a whole company?

Across whole companies and economies, measured gains are still small. In the second quarter of 2026, 39.2% of employed U.S. adults used generative AI for work in the past week, and self-reported savings equalled 2.2% of all work hours³. Part of the gap is arithmetic: AI assisted only 6.3% of work hours that quarter, so even big savings inside those hours barely move the total³.

A six-month randomized trial across 66 firms and 7,137 workers found that staff using Microsoft 365 Copilot spent two fewer hours a week on email, with no detected change in the amount or mix of their other work¹⁰. Economists Anders Humlum and Emilie Vestergaard linked surveys of about 25,000 Danish workers to official records and found no measurable effect on earnings or hours through 2024¹¹.

In McKinsey's 2026 global survey, 80% of respondents said AI had improved their own productivity, and 37% said it had added anything to their organization's earnings before interest and taxes (EBIT)¹².

Firms that redesign the work capture more of the gains

McKinsey's high performers are the roughly 6% of respondents who attribute 5% or more of EBIT to AI. Nearly three-quarters of them said they had fundamentally redesigned workflows because of AI, compared with one-quarter of everyone else¹².

BCG's 2026 survey of employees at companies with at least US$100 million in revenue found a similar split. Where companies redesigned workflows end to end or built new business models, 67% of respondents saw measurable business improvement from AI. Where companies only rolled out tools, 43% did⁴.

Both surveys show correlation, so they can't separate redesign from good management in general, and BCG's sample leans toward large companies. For the workflows with the most proof so far, see which sales and operations jobs AI has proven it can do.

AI productivity in Canada depends on skills and process

Statistics Canada found firms using AI were 16.8% more productive than firms that didn't¹³. The premium fell to 10.2% once the authors accounted for how productive the firms already were. It fell again to 5.1%, which isn't statistically significant, once they added skills like data analytics and technology training for staff¹³. The data come from 2019 and 2021, before generative AI spread¹³.

In a December 2025 Bank of Canada survey, 70% of business leaders said they used AI personally in a typical work week, while 8% said their firm made significant use of AI in producing goods or delivering services¹⁴. The survey drew 314 responses and isn't designed to represent all Canadian firms¹⁴.

AI ROI for small business: what to do first

Use the research to decide where to look, then measure your own results. If you haven't mapped where AI could fit in your firm, the readiness check walks through it.

1. Pick two or three well-defined, high-volume tasks, ideally ones newer staff handle. Hold back judgment calls like the BCG brand recommendation until you know how the tool fails on your own work.
2. Record a baseline before anyone gets the tool: time per task, volume and error rates. Self-reports mislead, since the METR developers felt faster while taking 19% longer⁸.
3. Redesign the workflow before you scale it. Decide which steps the tool takes, who checks its output and what people do with the time saved. BCG found 66% of non-manager office staff who use AI regularly get limited or no guidance on that⁴.
4. Teach people to check output as well as write prompts. Consultants given GPT-4 and a prompting overview did worst on the task beyond its reach, at 60% correct². Why staff training decides the result covers what to teach.
5. Judge the pilot over months against the baseline. The Copilot trial reported its email savings for the second half of a six-month experiment¹⁰. Keep what beats your baseline and drop the rest.