AI Efficiency Calculator
Estimate how much time AI could realistically save you each week — anchored to peer-reviewed research and real-world studies, not vendor marketing.
How this works
- Select the tasks you do regularly. Each shows a potential saving — from a published study where one exists, and labelled as our estimate where one doesn’t.
- Set your usage reality — how often you'll use AI, how senior you are, how much you need to check its output.
- Read three numbers — the upper bound (potential), your estimate (realistic), and what US workers report on average (national).
Where should we send you from here?
Tell us who you are to see your three numbers, how to read them, and your AI opportunities for each task you selected. It takes ten seconds and you only do it once.
Already signed up on this device? Your results unlock automatically next time.
Your estimated hours saved per week
Three baselines. Read all three — they tell different stories.
Upper bound. Assumes full adoption, no oversight cost. From controlled studies where they exist, our estimates where they don’t.
Your estimate. Applies your adoption, oversight, and experience level to the potential number.
Reality check. What US workers report on the St. Louis Fed’s national survey — about 2.2% of all working hours, including people who don’t use AI.
How to read these numbers
Select tasks and hit Calculate to see your result.
Estimates are illustrative, not forecasts. Task-level percentages come from published studies where they exist and are labelled as our estimates where they don’t (sources below). Outcomes vary significantly by tool quality, task complexity, workflow design, and individual skill level.
Research sources & methodology
How the three baselines are calculated:
- Potential — Weekly hours × % on selected tasks × average potential saving (from research).
- Realistic — Potential × adoption × (1 − oversight) × experience factor.
- National average — Weekly hours × 2.2% (St. Louis Fed Real-Time Population Survey, mid-2026; self-reported, all employed adults including non-users).
Key studies referenced:
- Noy & Zhang (2023, Science) — Writing tasks: 40% time reduction, 18% quality improvement in controlled experiment with 453 professionals.
- Peng et al. / GitHub (2023) — Coding: 55.8% faster on a single controlled task; confidence interval 21–89%.
- Cui, Demirer, Jaffe, Musolff, Peng & Salz (2025) — Three field experiments with 4,867 developers at Microsoft, Accenture and a Fortune 100 firm: 26% more completed tasks with an AI coding assistant; less experienced developers gained most.
- METR (2025, updated 2026) — Experienced open-source developers took 19% longer with AI on codebases they knew well in early 2025. A late-2025 follow-up pointed towards a speed-up, but METR calls that evidence weak.
- Brynjolfsson, Li & Raymond (2025, QJE) — Customer support: 15% average productivity gain; juniors improved most, senior experts saw small or negative effects.
- Bick, Blandin & Deming (Federal Reserve Bank of St. Louis, Real-Time Population Survey) — By mid-2026, 39% of employed US adults had used generative AI for work in the past week, and self-reported time savings were 2.2% of all work hours, up from 1.6%. In the November 2024 survey, people who used AI reported saving 5.4% of their hours.
- Humlum & Vestergaard (2025) — Danish workers using AI chatbots reported saving 2.8% of work hours on average.
- Microsoft (2024) — Copilot users reported saving about 11 minutes per day (self-reported survey).
- Dell'Acqua et al. (2023, HBS "Jagged Frontier") — Productivity gains vary sharply by task type; tasks inside AI's capability saw large gains, tasks outside saw quality decline.
- Vendraminelli et al. (2025, HBS Working Paper 26-011, "The GenAI Wall Effect") — AI helps people stretch into adjacent work, but does not turn outsiders into specialists on tasks far from their expertise.
Caveats worth remembering: Most real-world figures, including the national average, are self-reported, and people tend to overestimate — in METR’s trial, developers believed AI had made them 20% faster when they were measured as 19% slower. Time saved doesn't always become more output — some becomes slack. And senior experts gain less than juniors on specialist tasks they already know well.