Less routine work. More confidence in what you produce. A clearer next step. Choose what matters to you, then test one small change.
Your choices adapt the example. They stay in this page and do not change an assessment score.
Fictional workflow example · Financial analyst
Write a variance commentary from fictional budget figures.
What to check: Count preparation, drafting, checking and corrections. Compare outputs against the same quality standard.
Time one small attempt and decide whether to keep, change or stop the workflow.
Showing save time on routine work for the financial analyst example.
A faster draft is only part of the story. Count setup, checking and correction before deciding whether a workflow helps.
These are illustrative numbers. Change them to explore a possibility, then measure a real attempt. Entries stay in this page; reloading clears your changes.
Use whole minutes, up to 1,440 for either complete approach, and 1–40 repeats. Leave an unknown time blank; it is not zero.
Usual approach30 min
AI + checking20 min
12 min setup and drafting + 8 min checking and correction
10 minutes less per task.
At 5 repeats, that would be 50 minutes less per week. This is arithmetic from your inputs, not a forecast. Quality must still pass your check.
Give any saved time a purpose. Focus on a harder task, practise a skill or aim to finish within your planned working hours. Faster output alone does not decide how your workload changes.
Try one approved, low-risk task using fictional or permitted information. Keep the quality standard fixed. If checking takes longer or errors remain, change the task or stop. Record a real attempt in the free worksheet.
Faster writing. More resolved tickets. Sometimes, more work. These studies explain why the task and the checks matter more than a blanket promise about AI.
Historical research snapshots, checked . Charts show reported effects normalized to a baseline of 100, not raw group averages. Different tasks and metrics cannot rank professions. These are rounded estimates, not live results or predictions for you.
40% less time on selected writing tasks
A randomized experiment also reported 18% higher evaluator grades. Those ratings are not a measure of factual accuracy.
Short, generic assignments did not require precise factual accuracy or company-specific knowledge. The headline estimates concern the variable-time groups, not every participant or a whole working day.
Read the author’s full manuscript5,172 agents · 2020–21 rollout · GPT-3-based assistant
Brynjolfsson, Li & Raymond · QJE, February 2025About 15% more issues resolved per hour
A staggered rollout at one company compared changes in assisted workers with workers without access. Effects differed by experience and skill.
The full study included 5,172 agents; the throughput measure covers a smaller subset. This is assisted technical support at one firm, not autonomous bots, retail support or an employment forecast.
758 consultants · two task groups · 2023 GPT-4
Dell’Acqua et al. · Organization Science, March 202625.1% less time to reach the final subtask
The same research found correctness 19 percentage points lower with AI on a separate managerial problem. Different tasks produced opposite results.
The time result is time to reach question 18, not full-session duration. The two task groups had different participants. This is one firm's historical experiment; the correctness change is percentage points, not a relative percent.
16 developers · 246 tasks · early-2025 tools
Becker et al. · METR, July 202519% more time to complete issues
A randomized study of experienced contributors working in familiar repositories found a slowdown (95% interval: 2–39% more time).
This is an early-2025 snapshot. Later estimates have measurement limitations: read METR’s February 2026 follow-up.
Small, selected sample using primarily Cursor with Claude 3.5/3.7. METR's February 2026 follow-up suggests newer speedups but says selection and time-measurement problems prevent a reliable magnitude. This is not a current-tools verdict.
This page uses AI-assisted editorial research. The workflow examples above are fictional and do not reproduce these studies. We update the evidence when a substantive source changes; there is no continuous profession-improvement dataset behind these charts.
Want a starting point based on the work you actually do? The free assessment gives you an explained exposure profile and a small task experiment. No account or payment required.
Assess my own workPreparing for a bigger move? Build a 30-day preparation plan.