How Much Time AI Saves by Task (2026 Data, From Published Studies)

What "56% faster" actually measured (the Peng / GitHub Copilot study)

The most-cited AI productivity number — 55.8% faster — comes from Peng et al. 2023, a randomized controlled trial GitHub ran with 95 professional developers. The task: implement an HTTP server in JavaScript. The treatment group used GitHub Copilot; the control did not.

Result: Copilot users finished in a median of 71 minutes and 26 seconds. The control group took 160 minutes and 53 seconds. That's 55.8% faster — and the difference was statistically significant. Completion rate was also higher: 78% in the control group vs. 96% with Copilot.

Two things to note before generalizing. First, the task was well-defined, greenfield, and within Copilot's strong zone (JavaScript boilerplate-heavy work). Second, the participants were not picking the task — they were assigned it. Real developer time is dominated by reading code, debugging, meetings, and decisions. The HTTP-server task represents the fraction of the day that's pure typing, which is exactly the part Copilot accelerates most. So 56% on that task does not mean 56% on a sprint — it means 56% on the type of work that the type of tool most directly automates.

Writing tasks: Noy & Zhang's 37% — and the quality lift

Noy and Zhang's 2023 MIT study (published in Science) ran 444 college-educated professionals through realistic writing tasks: press releases, short reports, analysis plans, and delicate emails. Half were given ChatGPT-3.5; half were not.

Average time on task dropped from 27 minutes to 17 minutes — a 37% reduction. But the more interesting finding was on quality: independent graders, blind to condition, rated ChatGPT-assisted writing higher on average. The lift was largest for workers whose baseline writing was weakest, narrowing the gap between strong and weak writers.

This is the cleanest "AI saves time AND improves quality" result in the literature. It also matches the intuition: short business writing is exactly where the model has seen enough examples to produce a good first draft, and the marginal human edit is faster than writing from scratch. For comparison with other professions, this 37% writing delta is the figure marketers, comms staff, and analysts should anchor on — not the 56% coding figure.

Customer support: the Brynjolfsson 14% number — and the 35% novice lift

Brynjolfsson, Li, and Raymond's 2023 NBER paper (w31161) studied 5,179 customer support agents at a Fortune 500 software firm. The intervention was a GPT-4-based assistant suggesting responses. The headline result: a 14% increase in issues resolved per hour, on average.

But the average hides the headline finding. Novice and low-skilled agents saw a 35% productivity increase. Experienced agents saw essentially zero gain — they already knew the playbook. This is the most carefully documented case of AI compressing the skill distribution: the tool acts as on-the-job training, raising the floor faster than the ceiling.

There was also a quality effect: customer satisfaction scores went up, agent retention improved, and the workforce reported lower stress. So for support specifically, the 14% headline understates the strategic impact. If you're modeling AI's effect on a support team's headcount, the right input is closer to 20% — weighted toward your junior agents.

Consulting: Dell'Acqua's frontier — 25% faster inside, slower outside

Dell'Acqua et al.'s 2023 study (Harvard Business School working paper 24-013) ran 758 Boston Consulting Group consultants through 18 realistic business tasks. Half had access to GPT-4; half did not. The result has two halves that everyone quotes selectively.

Inside the "AI frontier" (tasks the model handled well — creative product ideation, persuasive writing, structured analysis), consultants with GPT-4 completed 12.2% more tasks, finished 25.1% faster, and produced output that independent graders rated 40% higher in quality. That's an unambiguous win.

Outside the frontier (a business case requiring numeric reasoning with a misleading dataset), consultants with GPT-4 were 19 percentage points more likely to get the wrong answer. They felt more confident while being more wrong. This is the single most important caveat in the AI productivity literature: the tool's speed-up is task-conditional, and using it on the wrong task makes you worse, not just neutral.

Where AI does not save time (and may cost it)

Some categories show no time savings in the data, and some show negative savings. Tasks involving genuine first-principles reasoning, real-world judgment about unfamiliar contexts, decisions under high regulatory liability, and creative work where the value is the human's specific voice tend to show flat or negative effects.

From Dell'Acqua's outside-frontier sample: numeric reasoning with messy data made consultants 19 percentage points worse. From practitioner reports through 2024-2025: legal contract drafting (vs. review) shows minimal savings once verification time is counted; financial modeling at the strategy level shows minimal savings; senior code review of architectural decisions shows minimal savings.

There is also a class of work where AI accelerates the drafting step but slows the cycle overall because the output requires verification by someone who could have written it faster from scratch. We treat this case explicitly in the next section.

The verification tax — when AI shifts time rather than saving it

A pattern that shows up across multiple replication attempts: AI compresses drafting time but expands review time. For tasks where the cost of being wrong is high (legal, medical, financial, security-sensitive code), the human still has to read every line. If the reviewer is more expensive than the drafter, total cost can go up even when total time goes down.

The cleanest example: a junior engineer using Copilot to produce a feature in 2 hours instead of 4 saves 2 engineering hours. But if a senior reviewer now spends an extra 30 minutes verifying that the LLM didn't introduce a subtle bug, and the senior is 2x the cost of the junior, the team-wide savings are smaller than the per-task time delta suggests.

We don't have a clean published number for the verification tax — it varies too much by domain. But the heuristic that holds up across studies: for tasks where verification is cheap (e.g., "does this draft sound good?"), AI savings are real and durable. For tasks where verification is expensive (e.g., "is this contract clause actually safe?"), the time savings degrade once you account for the reviewer.

Microsoft Research 2024: Excel, SQL, and the formula-creation jump

Microsoft Research's 2024 internal studies of Copilot for Excel and Copilot for SQL workloads found the largest per-task savings in the dataset: formula creation dropped from a 10-minute baseline to ~2 minutes, and mid-complexity SQL query authoring dropped 65-75%.

These numbers are large because the underlying task is structured language generation — exactly what an LLM is built for. The verification step is also cheap: you run the formula or query and see if the output is right. That cheap verification is why Excel and SQL formula creation top the per-task savings table.

It's also why the freelance and contractor market is repricing tasks heavily in this category. A spreadsheet macro that used to be a $200 gig is now a $40 gig, and the floor will likely keep falling. If you're a freelancer in this lane, the per-task table above is roughly your repricing schedule.

Methodology — how we built this table and what it doesn't say

Rows tagged with a peer-reviewed or working-paper source (Peng, Noy & Zhang, Brynjolfsson, Dell'Acqua) use the original study's numbers, converted to a comparable "% time saved" basis. Where the original measured throughput (chats/hr), we report it as a directional throughput delta in the % column.

Rows tagged "replication" or "industry" combine published vendor data (Microsoft Research, GitHub Copilot impact reports, OpenAI usage analyses, McKinsey State of AI 2024) with practitioner replications. These should be read as best-available point estimates with wider error bars than the peer-reviewed rows.

What this table does not say: it doesn't say that a worker doing job X gets the full per-task savings on their daily workload. A developer doesn't spend 8 hours writing HTTP servers — they spend 6-7 hours on meetings, code review, debugging, and design. The realistic per-job multiplier is much smaller than the per-task multiplier. For per-job estimates, use the AI productivity multiplier by role calculator, which applies task mix weights on top of these per-task figures.

And finally — these numbers are 2023-2024 data. Model capability has continued to improve in 2025-2026, so the floor on most rows has likely risen, but no large RCT has re-run the Peng or Noy & Zhang protocols on newer models. We will update this page when published replications land.

Frequently asked questions