How much time does AI really save? What the research says

By Lawrence Goh, Director, Autonex Pte Ltd · Updated 23 Oct 2026

Vendor claims of "50% time savings" are everywhere. Independent research gives a more nuanced picture: real gains on the right tasks, little or nothing — sometimes losses — on the wrong ones. Here are the studies worth knowing before you build a business case.

Summary

Study	Who	Finding
NBER, Brynjolfsson et al. (2023)	≈5,000 customer support agents	+14% issues resolved per hour; +35% for novices; ≈0 for top performers
Harvard Business School & BCG (2023)	758 consultants	25% faster and 12% more tasks completed; 19 percentage points less likely to be correct on a task beyond AI's ability
St. Louis Fed survey (2024)	US workers who use generative AI	Saved an average 5.4% of work hours (≈2.2 hours a 40-hour week)
St. Louis Fed survey (2025)	All US workers, users and non-users	Time savings equal to 1.6% of all work hours
METR randomised trial (2025)	16 experienced open-source developers	19% slower with AI tools — while believing they were 20% faster

1. Customer support: +14% on average, +35% for newcomers

The largest field study to date followed about 5,000 customer support agents at a Fortune 500 software company as it rolled out a generative AI assistant that suggested responses. Issues resolved per hour rose 13.8% on average: agents spent about 9% less time per chat and handled about 14% more chats per hour.

The gains were very uneven. Less experienced and lower-skilled agents improved by 35%; the most experienced agents saw little or no benefit. Agents with two months' experience and the tool performed about as well as agents with six months' experience without it. Attrition also fell.

Lesson for ROI: AI often works by spreading the know-how of your best people to everyone else. Expect larger gains where staff are new or work is standardised; expect less from your experts.

2. Consultants: faster and better — inside the "frontier"

In a Harvard Business School experiment with 758 Boston Consulting Group consultants, those using GPT-4 on realistic consulting tasks completed about 12% more tasks and worked about 25% faster, and their output was judged higher quality. But on a task deliberately chosen to be beyond the AI's capabilities, consultants using it were 19 percentage points less likely to reach the correct answer than those working without it.

The researchers called this the "jagged technological frontier": AI is excellent at some tasks and unreliable at others that look similar, and people find it hard to tell which is which.

Lesson for ROI: map tasks individually. A blanket "AI saves 30%" across a department hides tasks where it saves 60% and tasks where it creates errors.

3. Worker surveys: modest gains across whole jobs

Economists at the Federal Reserve Bank of St. Louis surveyed a nationally representative sample of US workers. In late 2024, workers who used generative AI reported saving an average of 5.4% of their work hours — about 2.2 hours in a 40-hour week. A third of users saved an hour or less; about a fifth saved four hours or more. Daily users saved far more than occasional users.

Their 2025 update found generative AI use at work had grown to 37% of adults aged 18–64, and that time savings across all workers, including non-users, equalled about 1.6% of all work hours.

Lesson for ROI: a 5% saving across a whole job is consistent with a 25–50% saving on the 10–20% of tasks AI is good at. That is why our calculator asks about specific tasks, not total hours.

4. Experienced developers: slower, but convinced they were faster

METR, a research nonprofit, ran a randomised controlled trial with 16 experienced open-source developers working on 246 real issues in large codebases they knew well. With AI tools allowed, tasks took 19% longer. Before the study the developers expected a 24% speed-up; afterwards they still believed AI had made them about 20% faster.

METR stresses the limits: a small group of experts, in familiar codebases, using early-2025 tools, and it says the results may not reflect newer tools. The striking finding is the gap between perceived and measured productivity.

Lesson for ROI: do not rely on self-reported time savings. Measure them.

5. The big picture: large potential, uneven results

McKinsey estimated in 2023 that generative AI could add $2.6–4.4 trillion a year in value across the use cases it analysed, and that current technology could in principle automate work activities absorbing 60–70% of employees' time. Yet MIT's Project NANDA reported in 2025 that around 95% of organisations it studied were seeing no measurable business return from their generative AI investments. Both can be true: the potential is real, but capturing it depends on choosing the right tasks and changing the workflow, not just buying licences. We look at why in why AI projects fail to deliver ROI.

What this means for your numbers

Use 20–35% time savings on targeted tasks for planning, unless you have pilot data.
Expect bigger gains for newer staff and standardised work; smaller for experts.
Treat self-reported savings with caution — measure time per task before and after.
Remember quality: a faster process that produces more errors can cost more than it saves.

Open the ROI calculator

Sources

Lawrence Goh, Director, Autonex Pte Ltd. Lawrence designs AI and workflow automation for small and mid-sized businesses. About the author · How we write and check our guides