Where AI actually saves time: what the research says

The short answer: AI saves real time on writing, idea work and routine questions, and it helps less experienced people most. Used on the wrong task, it can also make capable people worse. Four studies show both sides, and together they point to a sensible way to start.

That’s a live question for most organisations. The ONS found that 29% of UK businesses used at least one AI technology in June 2026, up 8 percentage points on June 2025, so most haven’t started.

Customer support: 15% more issues resolved per hour

Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,172 customer-support agents given an AI assistant (Quarterly Journal of Economics, 2025). It suggested replies and linked to relevant internal documentation. The agent decided whether to use the suggestions and stayed responsible for the conversation.

Issues resolved per hour rose 15% on average. For less experienced and lower-skilled agents, the gain was around 30%. The most experienced and highest-skilled agents saw small gains in speed and a small dip in quality.

Professional writing: 40% faster, 18% better

Shakked Noy and Whitney Zhang gave 453 university-educated professionals writing tasks matched to their occupations, and let half of them use ChatGPT (Science, 2023). Average time taken fell 40% and quality rose 18%. The gap between stronger and weaker workers narrowed, too.

Consultants: better and faster, until the task changed

Fabrizio Dell’Acqua and colleagues ran an experiment with 758 Boston Consulting Group consultants, using GPT-4 as it was in April 2023 (Organization Science, 2026).

One group worked on product-development tasks inside the AI’s abilities: ideas for a new shoe, segmenting a market, a press release, a persuasive memo. With AI, they completed 12.2% more tasks, 25.1% faster, and the quality of their work rose by about 32% on average. Consultants who scored in the bottom half on an earlier assessment task gained most, though the top half improved too.

The other group got a business problem designed to sit outside those abilities. The spreadsheet looked complete, but the crucial details were in interview notes, and the AI, given the same files, would reach the wrong answer. Without AI, consultants got it right about 84.5% of the time. The two groups using AI scored 60% and 70.6%: on average, 19 percentage points less likely to be right.

The authors call this the “jagged technological frontier”. AI is strong at some tasks and weak at others that look just as hard, and the line isn’t obvious, even to skilled people doing familiar work. They also found consultants tended to over-rely on the AI. And they describe the frontier as expanding and changing, so the 2023 line isn’t today’s line. That’s exactly why it’s worth testing on your own work rather than assuming.

Government: 26 minutes a day, self-reported

From 30 September to 31 December 2024, the UK government gave 20,000 Microsoft 365 Copilot licences to staff in 12 organisations (GOV.UK, 2025). Users reported saving an average of 26 minutes a day, with the biggest savings on drafting documents. Two caveats come from the report itself: the savings were self-reported, and Copilot “struggles with nuanced or context-heavy data requiring human judgement”.

What the studies agree on

• The time is saved at the first-draft stage. Drafts, replies and ideas, where a person can judge the result quickly.
• Less experienced and less skilled people gain most. That’s good news for new starters. The catch is that someone new to a task is also less likely to spot a wrong answer, so checking has to be part of the process, not left to instinct.
• Experts gain less, and on the wrong task, anyone can do worse. The consultants were highly skilled and were still caught out.
• Measured beats estimated. Three of the studies measured output directly. The government figures are people’s own estimates: useful, but softer.

Which tasks to try first

Good first tasks share four traits:

1. They happen often, daily or weekly, so small savings add up to hours back.
2. A person can check the result quickly. Reading a draft is faster than writing one.
3. A mistake is cheap to catch before it reaches a customer, a funder or a court.
4. The information is in front of the AI: the email thread, the document, the notes.

That points to first drafts of routine letters and emails, summaries of long documents or meeting notes, suggested replies to common questions drawn from your own help pages, and pulling key details out of documents into a spreadsheet. These are the boring bits of the work, which is why we say pick something boring for a first project.

Be more careful with:

• tasks where the answer hangs on a subtle detail, like the consultants’ interview notes;
• decisions about people, money or legal rights, where a plausible wrong answer is expensive (our UK GDPR checklist for AI tools covers the legal side);
• anything nobody has time to check.

How to test it on your own work

Because the frontier is jagged, the only reliable test is your own tasks. We’d start like this:

1. Pick one task and write down how often it happens and how long it takes now. Time a sample of real examples, not just the tidy ones.
2. Run a short pilot with a few people. The AI does the first draft; a person finishes and checks it.
3. Check a sample of the results against your usual standard, and note the errors.
4. Compare time per task, error rate and whether people kept using it.
5. Decide: keep it, adjust it, or stop.

There’s more on the measuring in how to tell if an AI project worked. For a second opinion on which task to start with, should you automate this? asks the six questions we would, and the readiness self-check takes five minutes.