AI-first company

Measurement · Adopt

Measure time saved per task against a baseline

Record how long a task takes and how good the result is before you add AI. Without that starting figure you cannot tell a real gain from a hopeful one.

By
Last reviewed
Corrections
Report an error
When Do first. Before you start with AI, or as you start.
Effort Low (days, one person)
Evidence Common practice

Widely recommended. No study has tested whether companies that measure against a baseline do better. A randomised trial shows that people's own estimates can point the wrong way, which is why a figure from somewhere else, or from staff, is no substitute for a measured one.

Fits
Industries All industries
In the guide Micro and small businesses

What it is

Before you add AI, record the figure you want to improve: minutes per case, cases per hour, error rate. That figure is your baseline. After AI is in place, measure the same thing the same way and compare.

What the evidence shows

Published results differ widely, so someone else’s figure tells you little about your own.

The three do not contradict each other. The first two studies measured one well-chosen task. The third measured everyone who had access to a chatbot.

People’s own estimates are a poor guide, so measure rather than ask.

  • Staff can feel faster and be slower. In a randomised trial with 16 experienced software developers, each of 246 real tasks was randomly allowed or barred from using AI. Tasks took 19% longer with AI. Beforehand, the developers expected AI to make them 24% faster. Afterwards, they believed it had made them about 20% faster.
  • National self-reports are small. In an online survey series of US adults run by the economists Bick, Blandin and Deming, employed people put the time generative AI saves them at 2.17% of all their work hours in the second quarter of 2026. That figure is self-reported.

Reported returns are a different thing again. In a survey by IDC for Microsoft, which sells AI tools, business leaders reported an average return of $3.50 for each $1 invested in AI. That figure is what leaders said, not a measurement. When nearly 6,000 executives in four countries were asked, nine in ten reported no effect of AI on their firm’s productivity or employment over the past three years.

How to do it

  1. Pick one or two measures per task: time per case, and one quality measure.
  2. Record them for a few weeks before the change. Use system logs where you can, not estimates. You cannot rebuild a baseline afterwards. When the developer trial above was rerun in 2026, many developers would not work without AI, and 30–50% held back tasks they did not want to do without it, so the comparison failed. At four endoscopy centres in Poland, doctors’ detection rate in colonoscopies done without AI had already fallen from 28.4% to 22.4% three months after AI arrived. A “without AI” figure taken after rollout makes the AI look better than it is.
  3. Change one thing at a time.
  4. Measure again once people are past the learning period.
  5. If you can, keep a group working the old way for comparison. Set it up at the start. Once people are used to the tool, they refuse the old way or choose which tasks go into it.
  6. Count checking and correction time as part of the new process.
  7. Report the result even when it is small or negative.

What to measure

  • Time per case or cases per hour.
  • A quality measure: error rate, rework, customer complaints.
  • Checking time added by AI.
  • Sales or fees per job, and whether your prices changed. Time saved and money earned can diverge. Among US small employer firms using AI, 71% told a Federal Reserve survey their productivity had risen, and 31% that their sales had risen.
  • Outside spending avoided. In the same survey, 18% of firms using AI had cut their spending on outside services.
  • Work that was not being done before: unanswered enquiries, products with no description. In randomised experiments on a large online retail platform, the largest sales gain, 16.3%, came from a chatbot answering buyers who had previously had no live help at all.
  • An objective quality measure, not only customer ratings. In a field experiment in Alibaba’s customer service, AI help raised customer ratings, but customers came back with the same problem just as often. The study is a working paper, not yet peer-reviewed.
  • Output per day, when people run several AI tasks at once. In the 2026 rerun of the developer trial, time per task stopped being reliable once developers ran several AI agents in parallel.

To see how many minutes a week each person must save to cover the costs, put your figures into Will AI pay back?.

Where it fails

  • Self-reported time savings. People’s estimates can get the direction wrong, not only the size. The developers above felt faster and were slower. Two UK government evaluations of the same office assistant also differ. Across government, users said Microsoft 365 Copilot saved them 26 minutes a day. HMRC, the UK tax authority, gave 3,000 licences to staff chosen at random and cut the self-reported figures by about 20% for bias. It found about 60 minutes a week. Both figures are what staff said. See the claim check on the 26 minutes. Use logs or timed samples.
  • Speed without quality. A faster process that produces more errors is not a gain.
  • Saved time that goes nowhere. Minutes saved per task matter only if the time is put to use. In a field experiment with 7,137 workers at 66 firms, workers given an office AI assistant spent about two hours less on email a week, and the amount and mix of their other work did not change. Three of the four authors worked for Microsoft, which sells the assistant. Decide in advance what the time will be used for.
  • Billing by the hour. If you charge clients for time, an hour AI saves is an hour you can no longer bill, unless you change how you price. Clio, which sells legal software, reports that in its 2026 survey 86% of solo lawyers and 78% of small law firms had made no pricing change because of AI. Before you count the time saved as a gain, decide whether you will charge by the job, a fixed fee or the hour, and what that does to your fees.
  • Work that is hard to count. For varied expert work, compare samples of output, not minutes.

By company size

  • Micro and small: a spreadsheet and a two-week tally are enough. Note your sales or fees per job in the same sheet. If you bill by the hour, decide how you will price before you bank the time saved.
  • Mid-market and enterprise: use workflow system data, and agree the measures with finance before the project starts, so the result is accepted afterwards.

See it applied

Company cases

Companies that did this

What happened without it

Sources

  1. Humlum and Vestergaard (2025). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777 (revised March 2026). What this study did and found
  2. Brynjolfsson, Li and Raymond (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942. What this study did and found
  3. Noy and Zhang (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192. What this study did and found
  4. Becker, Rush, Barnes and Rein (METR) (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. METR, published 10 July 2025 (arXiv 2507.09089); follow-up published 24 February 2026. What this study did and found
  5. Dillon, Jaffe, Immorlica and Stanton (2025). Shifting Work Patterns with Generative AI. NBER Working Paper 33795; arXiv 2504.11436, version 4 (November 2025). What this study did and found
  6. Budzyń, Romańczyk, Kitala and colleagues (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. The Lancet Gastroenterology & Hepatology, 10(10), 896–903. What this study did and found
  7. Federal Reserve Banks (2026). Small Business Credit Survey: 2026 Report on Employer Firms. Federal Reserve Banks, Small Business Credit Survey (published March 2026). What this study did and found
  8. Fang, Yuan, Zhang, Donati and Sarvary (2025). Generative AI and Sales Productivity: Field Experiments in Online Retail. arXiv 2510.12049, version 6 (29 June 2026). What this study did and found
  9. Yotzov, Barrero, Bloom, Bunn, Davis, Foster, Jalca, Meyer, Mizen, Navarrete, Smietanka, Thwaites and Wang (2026). Firm Data on AI. NBER Working Paper 34836 (February 2026, revised March 2026). What this study did and found
  10. Microsoft (2023), reporting IDC survey research: an average return of $3.50 per $1 invested in AI, as reported by business leaders. blogs.microsoft.com
  11. Bick, Blandin and Deming. Generative AI, time savings: employed adults (series RPSGENAITSALL), from the Real-Time Population Survey, an online survey of US adults. Federal Reserve Bank of St. Louis, FRED. fred.stlouisfed.org
  12. UK Government Digital Service (2 June 2025). Microsoft 365 Copilot experiment: cross-government findings report. gov.uk
  13. Ni, Wang, Feng, Lu, Wang and Zhou (2026). Generative AI in Action: Field Experimental Evidence from Alibaba's Customer Service Operations. arXiv 2603.29888. Working paper, not peer-reviewed. arxiv.org
  14. Clio (2026). Legal Trends for Solo and Small Law Firms. Press release. Clio sells legal software with AI built in. clio.com

Claims on this topic, checked against these studies: