Longer Days Among AI Users Are Not Proof

A tool that is supposed to save time should leave a margin. A new analysis of digital activity finds longer observed workdays among employees who use artificial intelligence than among those who do not. The contrast is worth a manager's time. It is also easy to over-read. The tables compare groups. They do not show that AI caused one person to work an extra hour.

On 29 October 2026 ActivTrak's Productivity Lab reported patterns from 9,140 knowledge workers at 467 enterprises, observed from 1 April to 25 September 2026. The press release says workday spans range from 8.4 to 9.1 hours, compared with 7.6 hours for non-users, a difference of up to 90 minutes. The blog post the same day prints 8.42 to 9.08 hours against 7.58. The tenths and the hundredths are the same comparison at two levels of rounding. The upper end is not the average increase for every AI user, and it is not a before-and-after for one employee.¹²

The comparison groups are not one experiment

The lab's definitions matter more than the headline. Heavy AI users use AI for more than 45 minutes and at least five sessions a day. Light users spend less than 10 minutes a day. Regular users fall between 10 and 45 minutes. Non-users do not use generative AI tools, in the study's definition. A workday span is not paid hours, not output, and not time spent inside an AI product. The data are observable digital behaviour on ActivTrak's platform. Offline conversation, physical work and the reason a day ran long are outside that record.¹

People who use AI may already hold different jobs. Managers and specialists with hard deadlines may adopt a tool and also work late. Busy teams may be first in line for a rollout. Without controls for those differences, a gap between users and non-users is not a causal effect. The company's own quotation goes further than the design. Gabriela Mauch, chief customer officer, said the research shows AI "contributes to longer workdays, more application switching and shorter focus sessions." "Contributes" is the vendor's reading. This article does not adopt it as a demonstrated cause.¹

The release also changes the comparison group from one bullet to the next. Heavy users switch applications 1.1 times per hour, against 0.95 for non-users, which the release calls 15.2 percent more often. The blog prints 1.09 rather than 1.1. Uninterrupted focus sessions average 37.8 minutes for heavy users and 48.2 minutes for light users, about 22 percent shorter, or 10.4 minutes in the blog's arithmetic. Focus is heavy against light. Switching is heavy against non-users. Putting those two sentences into one controlled experiment overstates the file.¹²

Heavy users spend 3 percent of work time in formal meetings, against 5 percent for non-users. They spend 32.3 percent of the day after hours, against 31.9 percent for non-users, and the longer day adds 18 minutes of after-hours activity. A similar share of a longer day produces more evening minutes without a large change in the proportion. None of these figures say whether a customer got a better answer, a report got more accurate, or a person learned a skill that transfers.¹

ActivTrak sells the software that produces the measurement. That commercial interest does not make the counts false. It does mean a buyer should ask who was in the 467 enterprises and what the platform cannot see. CIO, on 3 November 2026, quotes Sanchit Vir Gogia of Greyhound Research on that limit: an employee can use AI and still be classed as a non-user if the instrumentation misses it; a personal device, a background agent, an offline meeting and paid hours are different measures; switching applications is easier to observe than switching attention inside one application. This article did not audit ActivTrak's customer list.⁴

The March report is not the other half of this sample

On 11 March 2026 the same lab released a State of the Workplace report covering more than 443 million hours, 163,638 employees and 1,111 organizations, from 1 January 2023 through 31 December 2025. Under productivity, the release says the workday shrank 2 percent, from 8 hours 53 minutes to 8 hours 44 minutes, and productive hours increased 5 percent, by 19 minutes, to 6 hours 36 minutes. Collaboration surged 34 percent. Multitasking rose 12 percent. Focus time fell to a three-year low. A footnote says before-and-after findings on AI adoption use a subset of 376 companies and 10,584 users, comparing 180 days before and after adoption. The 2 percent and the 5 percent are not labelled as that subset.³

CIO notes that the latest research and the earlier "working less" result can be read as a contradiction, and that the article was updated to clarify that some statistics are not related to AI. The two releases are not successive measures of one population. One describes a broad sample over three years. The other compares user categories inside a 2026 window, on a much smaller group. They can both be true without telling a manager whether a specific workflow got better. The question of where any saved minutes should go is a different article: who gets the hours back.⁴

The October blog adds a distinction the press bullets only sketch. Employees using AI embedded in a primary application average 0.6 application switches per hour. Browser-based tools range from 1.7 to 4.2. At the top of that range, switching is more than four times the 0.95 rate for non-users. The lab calls the extra movement a "swivel-chair tax". Software engineers, more likely to use embedded tools, average 0.6 switches per hour in the press release and 0.64 in the blog, and spend 78.5 percent of active time in core coding applications. Operations managers spend 75.5 percent of active time in core applications, average 1.1 switches per hour, and rise to 4.1 during heavy coordination. The blog, not the press release, adds that engineers complete 16.8 percent of work in the evening and operations managers 30.5 percent. Those role lines are still associations. They are not proof that the browser caused the evening.¹²

Saved minutes can be spent twice

Consider a hypothetical ten-person nonprofit that uses an assistant on grant applications. A coordinator used to spend six hours drafting and two hours checking evidence. Drafting falls to three hours. The team then submits twice as many applications and needs extra checks for unsupported claims and private information. The coordinator's day can get longer while each first draft gets faster. The arithmetic is an illustration. It is not a result from the 9,140-person file.

More applications can raise funds if they are relevant and evidenced. They can also exhaust staff and crowd out work with beneficiaries. Productivity is useful output against all the resources used, including review, corrections, training and supervision. A count of prompts does not settle it. The same pattern can appear in a small accounting practice: classification speeds up, and the professional remains accountable for the judgment. If every saved minute becomes another assignment, the bottleneck moves. The review time that eats a raw gain is discussed in why human review becomes the bottleneck once agents are fast.

A practical record for a smaller organization starts with one workflow. Write down elapsed time, human effort, errors, rework and what the client or beneficiary actually received, before and after. Keep changes in staffing and case complexity visible. Measure the first draft and the approved output separately. A ten-minute draft that takes forty minutes to verify saves ten minutes against an hour of manual work, not fifty. Agree in advance where a verified gain will go: less overtime, training, mentoring, or a better service. If people are punished for reporting a saving, they will stop reporting it.

Individual keystroke monitoring is a poor way to answer that question. Prefer aggregated measures tied to a specific operational problem. Say what is collected, who can see it and how long it is kept. Do not treat keyboard activity, or how often someone opens an AI tool, as the value of the employee. For a European employer, Annex III, point 4(b), of the EU AI Act lists systems intended to monitor and evaluate performance and behaviour in a work relationship. The Commission service desk says those Annex III rules apply from 2 December 2027. A team log of errors on one workflow is not, by that description alone, such a system. Classification depends on the intended purpose. This paragraph is not legal advice. The staffing of oversight is a separate question, in who oversees the algorithmic manager.⁵⁶

How hilighter can help

hilighter does not measure a workday, install activity monitoring, or decide whether AI lengthened someone's hours. The ActivTrak gap leaves a management question: which tasks still need a person, and what should happen to any time a tool appears to release.

hilighter helps teams map skills by role, compare people with those requirements in Skills & Talent Insights, record proof of capability in Skills Evidence & Recognition and set personalized growth paths. Its AI Job Impact view is indicative. It is not a timesheet and it is not a reading of the 90-minute gap. A free plan covers up to three job roles, with unlimited employees.

Did AI add 90 minutes to every user's day?

No. The 29 October release says spans for AI users range from 8.4 to 9.1 hours, against 7.6 hours for non-users, a difference of up to 90 minutes. The blog prints 8.42 to 9.08 against 7.58. That is a comparison of groups in a file of 9,140 knowledge workers. It is not the average increase for every user, and it is not a before-and-after.¹²

Do the March figures contradict October?

They answer different questions. The 11 March release says that, in a three-year sample of 163,638 employees, the observed workday shrank 2 percent and productive hours rose 5 percent. Those lines are not labelled as the AI-user comparison in the October study. CIO says some statistics in the earlier report are not related to AI. A shorter average day in one sample and a longer span among AI users in another can both be reported without proving a cause.³⁴

Should an employer track each person's AI clicks?

Not as a measure of value. The study itself is digital behaviour on a vendor platform, and CIO's quoted analyst points out what that instrumentation misses. A smaller organization can record errors, rework and the destination of any verified time gain on one workflow, in aggregate. A system intended to monitor and evaluate a person's performance may fall under Annex III, point 4(b), with those rules applying from 2 December 2027. A workflow count is not that system merely because it mentions AI.⁴⁵⁶

Sources

1. ActivTrak Productivity Lab, AI users have workdays up to 90 minutes longer than non-users, 29 October 2026. 9,140 knowledge workers, 467 enterprises, 1 April to 25 September 2026. Spans 8.4 to 9.1 hours versus 7.6 for non-users. Heavy, light, regular and non-user definitions. Switching 1.1 versus 0.95 per hour (15.2 percent), heavy versus non-users. Focus 37.8 versus 48.2 minutes, heavy versus light. Meetings 3 percent versus 5 percent. After hours 32.3 percent versus 31.9 percent, plus 18 minutes. Embedded AI 0.6 switches per hour; browser-based 1.7 to 4.2. Mauch's "contributes" is the company's reading, not a causal result this article adopts. ActivTrak sells workforce analytics.
2. ActivTrak Productivity Lab blog, AI users' work days are up to 90 minutes longer, Sarah Altemus and Productivity Lab, 29 October 2026. 8.42 to 9.08 hours versus 7.58. Switching printed as 1.09 versus 0.95. Focus gap of 10.4 minutes, about 22 percent. Software engineers 0.64 switches and 16.8 percent of work in the evening; operations managers 30.5 percent in the evening. Those evening shares are not in the press release read for source 1.
3. ActivTrak Productivity Lab, AI is accelerating work, not replacing it, 11 March 2026. More than 443 million hours, 163,638 employees, 1,111 organizations, 1 January 2023 to 31 December 2025. Workday shrank 2 percent (8 hours 53 minutes to 8 hours 44 minutes). Productive hours increased 5 percent (plus 19 minutes) to 6 hours 36 minutes. Collaboration surged 34 percent. Multitasking rose 12 percent. Focus time at a three-year low. Footnote: before-and-after AI adoption uses 376 companies and 10,584 users. The 2 percent and 5 percent lines are not labelled as that subset.
4. CIO, Taryn Plumb, Does AI actually save employees time? Data is contradictory, 3 November 2026. Compares the October and March samples. Notes an update clarifying that some statistics are not related to AI. Quotes Sanchit Vir Gogia on instrumentation limits and on ActivTrak selling the measurement. The page image credit says the picture was generated with a tool; that credit is not used as a finding about workdays.
5. European Commission, AI Act Service Desk, Annex III, point 4(b). Read on 8 November 2026. Systems intended to monitor and evaluate performance and behaviour in a work relationship. Not legal advice. A workflow error log is not that system by description alone.
6. European Commission, AI Act Service Desk, When does enforcement start?. Annex III rules apply from 2 December 2027.

Sources last reviewed: 8 November 2026. The 90-minute figure is the top of a range between groups, not a causal effect and not an average for every user. Focus and switching use different comparison groups. The March 2 percent and 5 percent lines are not the October AI-user comparison. ActivTrak sells the measurement. EU dates follow the Commission service desk. This article is not legal advice. Neither an AI-detector score nor the absence of Google penalties can be guaranteed.