Does automating the admin really save time?
For specific tasks, yes, though less than the hype. Moving tax filing online cut firms’ tax time by 40%, AI cut professional writing time by 40% in an experiment, and an AI assistant saved office workers about two hours a week on email. Across the wider economy the gains aren’t visible yet. People overestimate the time AI saves, AI drafts can take longer to check, and the one trial with small-business owners found AI advice helped strong owners and hurt weaker ones. Automate specific, repetitive tasks, check what the AI says, and measure the time saved rather than assume it.
What to do
- Automate specific, repetitive tasks first: reminders, booking confirmations and routine answers on price and hours.
- Check what your chatbot tells customers about prices and policies: the business is responsible for it.
- Time a task before and after automating it, rather than guessing the saving.
- Expect a few weeks of set-up before the savings show.
- Keep booking lead times short: long gaps between booking and visit lead to no-shows.
What the research says
- Peer-reviewed or official
Technology, Taxation, and Corruption: Evidence from the Introduction of Electronic Tax Filing
Okunogbe & Pouliquen, AEJ: Economic Policy 2022
Moving a routine admin task (filing tax returns) from in-person to online cut the time firms spent on taxes by 40%. The working paper puts this at five fewer hours a month. Firms that had been less likely to evade also paid fewer bribes.
What it means for you: Moving a recurring admin task online saves real time: firms spent 40% less time on taxes once filing went online.
Keep in mind: Tajikistan and a government process; part of the saving came from no longer queueing in person.
Read the sourceMethod and quote
"Using experimental variation, we examine the impact of electronic tax filing (to replace in-person submission to tax officials) using data from Tajikistan firms." The abstract does not give the sample size.
“Journal abstract: "E-filing reduces the time firms spend on taxes by 40 percent." Working-paper abstract: "Firms that e-file have lower compliance costs, spending five fewer hours each month on fulfilling tax obligations.”
- Peer-reviewed or official
Experimental evidence on the productivity effects of generative artificial intelligence
Noy & Zhang, Science 2023
On realistic professional writing tasks, ChatGPT access cut time taken by 40% and raised quality by 18%. It also narrowed the gap between weaker and stronger writers.
What it means for you: For writing tasks like replies, reports and emails, AI cut time by about 40% and improved quality, most of all for weaker writers.
Keep in mind: Short online writing tasks graded by evaluators; savings at real workplaces are much smaller.
Read the sourceMethod and quote
Preregistered online experiment. 453 college-educated professionals did occupation-specific, incentivized writing tasks, and half were randomly given ChatGPT.
“Our results show that ChatGPT substantially raised productivity: The average time taken decreased by 40% and output quality rose by 18%. Inequality between workers decreased…”
- Peer-reviewed or official
Shifting Work Patterns with Generative AI
Dillon, Jaffe, Immorlica & Stanton, NBER WP 33795, rev. Nov 2025
Workers who were given and used an AI assistant built into email, meetings and writing tools spent two fewer hours a week on email and did less work outside regular hours. Their mix of tasks did not change.
What it means for you: An AI assistant in email and documents saved about two hours a week on email and cut work outside regular hours.
Keep in mind: Office workers at large firms; some authors worked for Microsoft, which makes the tool.
Read the sourceMethod and quote
Randomized access to Microsoft 365 Copilot across 66 firms and 7,137 knowledge workers over 6 months. Time use was measured from application telemetry, not from surveys.
“We present evidence from a field experiment across 66 firms and 7,137 knowledge workers… In the second half of the 6-month experiment, the 80% of treated workers who used this tool spent two fewer hours on email each week and reduced their time working outside of regular hours. Apart from these individual time savings, we do not detect shifts in the quantity or composition of workers' tasks resulting from individual-level AI provision.”
- Peer-reviewed or official
Doing Well by Doing Good: Improving Retail Store Performance with Responsible Scheduling Practices at the Gap, Inc.
Kesavan, Lambert, Williams & Pendem, Management Science 2022
Making staff schedules more consistent and predictable, and giving employees some control over them, raised store productivity by 5.1%: sales rose 3.3% and labour hours fell 1.8%.
What it means for you: Predictable staff schedules, set well ahead, raised store productivity by 5% and sales by 3%.
Keep in mind: A large US clothing chain testing a scheduling policy, not software. Gap helped fund the study.
Read the sourceMethod and quote
Randomized field experiment in 28 Gap stores in the San Francisco and Chicago areas, Nov 2015–Aug 2016. Supported by qualitative interviews and shift-level data.
“Intent-to-treat (ITT) analyses indicate that implementing responsible scheduling practices increased store productivity by 5.1%, a result of increasing sales (by 3.3%) and decreasing labor (by 1.8%).”
- Peer-reviewed or official
Generative AI and Sales Productivity: Field Experiments in Online Retail
Fang, Yuan, Zhang, Donati & Sárváry, CESifo WP 12201, 2026
Across seven AI deployments at a large cross-border e-commerce platform, sales effects ranged from zero to +16.3%. The largest was a pre-sale customer-service chatbot. Returns and ratings did not worsen. In search-query refinement, small sellers gained significantly while large sellers did not.
What it means for you: An AI chatbot answering shoppers before they buy raised sales by up to 16%, compared with offering no support at all.
Keep in mind: One e-commerce platform. The comparison is with no support, not with a person, and some authors advised the platform.
Read the sourceMethod and quote
Large randomized field experiments (millions of users and products), Sep 2023–Jun 2024, across seven workflows: pre-sale chatbot, search-query refinement, product descriptions, marketing push messages, Google ad titles, chargeback defence and live-chat translation.
“We find that GenAI adoption increases sales in most workflows, with effects ranging from no detectable impact to 16.3%, depending on GenAI's marginal contribution relative to baseline firm practices." Also: "sales effects range from no detectable impact in the advertising workflows to gains of up to 16.3% in the Pre-sale Service Chatbot." And: "these effects are not associated with worse post-purchase outcomes, as product return rates and customer ratings do not deteriorate.”
- Peer-reviewed or official
No-shows in appointment scheduling – a systematic literature review
Dantas, Fleck, Cyrino Oliveira & Hamacher, Health Policy 2018
The average no-show rate is about 23% (Africa 43.0%, Oceania 13.2%). The most commonly reported drivers are long lead time (the gap between booking and appointment) and a history of previous no-shows.
What it means for you: About one appointment in four is missed on average. Long gaps between booking and visit, and past no-shows, are the strongest warning signs.
Keep in mind: Healthcare studies worldwide: an average across very different settings, not a benchmark for salons or studios.
Read the sourceMethod and quote
Systematic review: 105 of 727 articles retrieved from Scopus (which includes MEDLINE), covering all medical specialties.
“The results indicate that the average no-show rate is of the order of 23%, being highest in the African continent (43.0%) and lowest in Oceania (13.2%)… Furthermore, the most commonly reported significant determinants of no-show were high lead time and prior no-show history.”
- Peer-reviewed or official
Annual Survey of Unincorporated Sector Enterprises, ASUSE 2025
National Statistics Office (MoSPI), press note, May 2026
39.37% of India's unincorporated non-farm establishments used the internet for business in 2025, up from 26.68% in 2023-24: 31.06% in rural areas and 48.94% in urban areas. More than half of traders used it. The sector has 7.92 crore (79.2 million) establishments.
What it means for you: In India, four in ten small unincorporated businesses used the internet for business in 2025, up from about a quarter a year earlier.
Keep in mind: Government survey. “Using the internet” is broad and doesn’t measure automation.
Read the sourceMethod and quote
Area-frame survey of 6,70,289 establishments in 24,153 first-stage units, by oral enquiry using CAPI.
“Use of internet, for entrepreneurial purposes, has increased from 17.94% in 2023-24 to 31.06% in 2025 in rural and from 37.01% to 48.94% in urban sector. Overall, it increased from 26.68% to 39.37% during ASUSE 2025 as compared to ASUSE 2023-24. Among the broad activity categories, more than 50% of trading establishments have used internet for entrepreneurial purpose during the period Jan-Dec, 2025.”
- Peer-reviewed or official
Use of artificial intelligence in enterprises
Eurostat, Statistics Explained, data 2025
19.95% of EU enterprises with 10+ employees used AI in 2025, up 6.47 percentage points on 2024. Use by size: 17% of small, 30.36% of medium and 55.03% of large enterprises. Among AI users, 28.49% of small enterprises (against 43.38% of large) used it to organise business administration. The top barrier, among firms that had considered AI, was lack of expertise (70.89%).
What it means for you: In the EU, small firms use AI far less than large ones (17% against 55%), and lack of expertise is the top barrier.
Keep in mind: Firms with 10 or more staff only; micro-businesses are likely lower.
Read the sourceMethod and quote
EU survey on ICT usage and e-commerce in enterprises (official statistics).
“In 2025, 19.95% of enterprises in the EU, with 10 or more employees and self-employed persons, used at least one of the following AI…" Also: "In 2025, 17% of small enterprises, 30.36% of medium enterprises and 55.03% of large enterprises used AI." And: "the most common reason for not using AI technologies was the lack of relevant expertise (70.89%), followed by lack of clarity about the legal consequences (52.52%), and concerns regarding violation of data protection and privacy (48.83%).”
What cuts the other way
Research that complicates the picture, and what to take from it.
- Peer-reviewed or official
The Uneven Impact of Generative AI on Entrepreneurial Performance
Otis, Clarke, Delecourt, Holtz & Koning, preprint, v3 Oct 2025
A GPT-4 business mentor on WhatsApp had no average effect on small-business revenue or profit. Owners who were already performing well may have gained over 15%; low performers did nearly 10% worse. The difference came from which advice owners chose to act on.
What it means for you: An AI business adviser on WhatsApp helped strong small-business owners and hurt weaker ones; on average it made no difference.
Keep in mind: Kenya, a preprint. The AI gave advice rather than doing the work.
Read the sourceMethod and quote
Five-month RCT with 640 Kenyan small-business owners from 44 of 47 counties (median age 26). The control group received a standard business guide.
“While we are unable to reject the null hypothesis of no average treatment effect on firm revenues and profits, we find that the effect for entrepreneurs who were low-performing at baseline is nearly 0.25 standard deviations lower than for initial high performers. Subsample analyses show that low performers did nearly 10% worse due to the AI assistant, whereas high performers may have benefited by over 15%. This differential impact does not appear to result from differences in the questions posed to or the advice received from the AI, but rather from which pieces of AI advice entrepreneurs selected and implemented." The full text describes "a randomized controlled trial (RCT) with 640 Kenyan small business entrepreneurs, half of whom were randomly assigned to receive access to a GPT-4-powered AI assistant over WhatsApp.”
- Peer-reviewed or official
Ambient AI Scribes in Clinical Practice: A Randomized Trial
Lukac et al., NEJM AI 2025
One AI scribe (Nabla) cut doctors' note-writing time by 9.5% against controls. The other (Microsoft DAX) showed no significant change. Each was used in only about 30% of visits. Burnout scores improved, and doctors reported clinically significant inaccuracies "occasionally".
What it means for you: Results vary by tool: one AI note-taker cut doctors’ writing time by about 10%, another made no difference, and doctors used them in only about 30% of visits.
Keep in mind: Two months at one US hospital system.
Read the sourceMethod and quote
Pragmatic three-arm RCT with 238 outpatient physicians in 14 specialties, randomized 1:1:1 to DAX, Nabla or usual care, 4 Nov 2024–3 Jan 2025, in an academic health system. Funded by the UCLA Department of Medicine and others.
“DAX was used in 33.5% of 24,696 visits; Nabla was used in 29.5% of 23,653 visits. Nabla users experienced a 9.5% (95% confidence interval [CI], -17.2% to -1.8%; P=0.02) decrease in time-in-note versus the control group, whereas DAX users exhibited no significant change versus the control group (-1.7%; 95% CI, -9.4% to +5.9%; P=0.66)." The abstract adds that "clinically significant inaccuracies were noted "occasionally"".”
- Peer-reviewed or official
Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI
Humlum & Vestergaard, NBER WP 33777, rev. Mar 2026
Two years after ChatGPT launched, AI chatbots had no measurable effect on earnings or recorded hours for Danish workers or workplaces. The estimates rule out effects above 2%, even though adoption was widespread and workers reported benefits.
What it means for you: Two years after ChatGPT, Danish data show no measurable effect on workers’ earnings or hours. Expect gains to show up slowly.
Keep in mind: Denmark, with 2023–2024 tools. Work is being reorganised, so this means “not yet”, not “never”.
Read the sourceMethod and quote
Large adoption surveys linked to Danish administrative labour-market records. Difference-in-differences.
“most employers in exposed occupations have adopted chatbot initiatives, workers report productivity benefits, and new AI-related tasks are widespread. Yet these currents have not broken the surface… we estimate precise null effects on earnings and recorded hours at both the worker and workplace levels, ruling out effects larger than 2% two years after the launch of ChatGPT.”
- Independent research
Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity
Becker, Rush, Barnes & Rein (METR), arXiv 2025
Experienced developers expected AI to save 24% of their time and afterwards believed it had saved 20%. In fact it made them 19% slower.
What it means for you: People overestimate the time AI saves: developers believed it saved them 20% when it actually made them 19% slower. Measure, don’t guess.
Keep in mind: Experienced software developers using early-2025 tools; a small study.
Read the sourceMethod and quote
RCT at the task level: 16 experienced developers worked 246 tasks in mature projects, with each task randomly allowed or disallowed AI tools (Feb–Jun 2025 tools).
“Before starting tasks, developers forecast that allowing AI will reduce completion time by 24%. After completing the study, developers estimate that allowing AI reduced completion time by 20%. Surprisingly, we find that allowing AI actually increases completion time by 19%--AI tooling slowed developers down.”
- Peer-reviewed or official
AI-Generated Draft Replies Integrated Into Health Records and Physicians' Electronic Communication
Tai-Seale et al., JAMA Network Open 2024
AI-drafted replies to patient messages increased reading time by 21.8%, did not change reply time and made replies 17.9% longer. Doctors still valued the drafts.
What it means for you: AI-drafted replies didn’t save doctors time: reading the drafts took longer, though doctors still valued them.
Keep in mind: Replies to patient messages, with a 2023-era model.
Read the sourceMethod and quote
Randomized waiting-list quality-improvement study in an academic health system, Jun–Aug 2023: 52 primary-care physicians (25 immediate, 27 delayed) plus 70 contemporary controls.
“The estimated association of GenAI was a 21.8% increase in read time (95% CI, 5.2% to 41.0%; P = .008), a -5.9% change in reply time (95% CI, -16.6% to 6.2%; P = .33), and a 17.9% increase in reply length (95% CI, 10.1% to 26.2%; P < .001).”
- Peer-reviewed or official
The Productivity J-Curve: How Intangibles Complement General Purpose Technologies
Brynjolfsson, Rock & Syverson, AEJ: Macroeconomics 2021
Technologies like AI need large, mostly unmeasured complementary investments, such as new processes and training. Measured productivity therefore dips or lags first and is overstated later. Adjusting US data for software intangibles raised TFP by 15.9% by 2017.
What it means for you: New tools need set-up, training and new routines first, so productivity often dips before it rises.
Keep in mind: An economy-wide finding, though the same set-up cost applies to a single shop.
Read the sourceMethod and quote
Growth-accounting model applied to US national data.
“General purpose technologies (GPTs) like AI enable and require significant complementary investments. These investments are often intangible and poorly measured in national accounts… this can lead to underestimation of productivity growth in a new GPTs early years and, later, when the benefits of intangible investments are harvested, productivity growth overestimation… adjusting for intangibles related to computer hardware and software yields a TFP level that is 15.9 percent higher than official measures by the end of 2017.”
- Peer-reviewed or official
Moffatt v. Air Canada, 2024 BCCRT 149
Civil Resolution Tribunal (British Columbia)
A company is liable for what its chatbot tells customers. Air Canada had to pay C$812.02 (C$650.88 in damages, plus interest and fees) after its bot misstated the bereavement-fare refund policy.
What it means for you: A business is responsible for what its chatbot tells customers. Keep your bot’s answers on prices and policies accurate.
Keep in mind: A small-claims decision in British Columbia, not binding elsewhere.
Read the sourceMethod and quote
Small-claims tribunal decision (negligent misrepresentation).
“In effect, Air Canada suggests the chatbot is a separate legal entity that is responsible for its own actions. This is a remarkable submission. While a chatbot has an interactive component, it is still just a part of Air Canada's website. It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot." The order: "I order Air Canada to pay Mr. Moffatt a total of $812.02, broken down as follows: a. $650.88 in damages, b. $36.14 in pre-judgment interest… and c. $125 in CRT fees.”
Sources checked as of October 25, 2026. Links go to the original publication.