The Evolution of Remote Work Communications Through AI Technology

How CodoraTech is funded: CodoraTech is supported by advertising and, in some articles, by affiliate links. Where an article contains affiliate links, we say so at the top of that article.

Microsoft’s telemetry says the average worker receives 153 Teams messages each weekday and 117 emails, that 57 percent of meetings are ad hoc calls with no calendar invite, and that meetings after 8 pm have risen 16 percent year over year. That is the condition AI meeting tools are sold against, and it is measured rather than asserted. Whether the tools relieve it is a separate question with much weaker evidence behind it. What exists divides cleanly into figures a vendor published about its own product and figures an independent researcher published about the field, and the two do not agree.

153 messages a weekday, and meetings after 8 pm

The best-sourced description of the problem is Microsoft’s own, published on 17 June 2025 as “Breaking down the infinite workday”. It draws on aggregated, anonymised Microsoft 365 telemetry through 15 February 2025 alongside a Work Trend Index survey of 31,000 knowledge workers across 31 markets. Half of all meetings fall in the 9 to 11 am and 1 to 3 pm blocks. A third now span multiple time zones, up 35 percent since 2021. Employees are interrupted roughly every two minutes, about 275 times a day, and the average worker handles more than 50 messages outside core hours.

It is worth noting who collected that. Microsoft measures the fragmentation on its own platform and sells the assistant marketed as the remedy. The telemetry is still the most concrete account of the problem available. The caution applies to the remedy, not the diagnosis.

Whose number is it

The most quoted result in this space belongs to Lumen Technologies and reached the public through Microsoft’s customer story page. Ashley Haynes-Gaspar, Lumen’s Chief Revenue Officer, is quoted saying that “it typically takes a seller four hours to do research for customer outreach, and with generative AI, they can now do that in 15 minutes”, and that “four hours back each week is worth $50 million in revenue over a 12-month period”. That 50 million dollar figure travels widely without its label. It is a Lumen projection published by Microsoft, not an audited result, and the same page attaches no metric to the meeting-recap use Lumen’s marketing chief Ryan Asdourian describes.

Zoom’s disclosures follow the same pattern. On its fourth-quarter fiscal 2026 earnings call on 25 February 2026, chief executive Eric Yuan said AI Companion monthly active users more than tripled year over year, side-panel users more than doubled quarter over quarter, and Zoom Phone AI usage rose 35 percent sequentially. Those are growth multiples with no absolute base disclosed, so they cannot be turned into a user count. On the same call Zoom named two organisations deploying Custom AI Companion across the enterprise: Harmonic, for knowledge retention, sales enablement and onboarding, and Grand Valley State University, for help desk and student-facing processes.

Zoom has also published a contact-centre figure, crediting InflectionCX with cutting after-call work by 3.5 minutes through AI summaries and auto-transcription, alongside research from the analyst firm Metrigy showing agent call time falling from 16.2 to 10.4 minutes. Metrigy produced that research; Zoom distributes it. None of these numbers is wrong on its face. All reach the public through a channel the seller controls.

Two point eight percent, and two hours of rework

The largest independent field measurement points somewhere else entirely. Anders Humlum of the University of Chicago Booth School and Emilie Vestergaard of the University of Copenhagen studied around 25,000 workers across 7,000 Danish workplaces in 11 occupations, linking two survey rounds to administrative labour-market records through June 2024. Their paper, NBER Working Paper 33777, appeared in May 2025. Between 64 and 90 percent of users in every exposed occupation reported saving time, and the average saving was 2.8 percent of total work hours. The headline finding is blunter: the chatbots had no significant impact on earnings or recorded hours in any occupation, with confidence intervals ruling out effects larger than 1 percent. Only 3 to 7 percent of reported savings passed through to higher earnings.

The authors themselves set that 2.8 percent against the 15 to 50 percent gains reported by controlled experiments. The gap between a laboratory task and a workplace is the single most useful fact in this literature.

There is also a cost that lands on the recipient rather than the user. Research by BetterUp Labs with the Stanford Social Media Lab, published in Harvard Business Review in September 2025, named it workslop: content that appears polished but lacks real substance and offloads cognitive labour onto colleagues. Forty-one percent of workers reported receiving it, at a cost of nearly two hours of rework per instance, which the researchers estimated at $186 per employee per month. Forty-two percent trusted the sender less afterwards. The sample size and field dates were not published, so the percentages should be quoted without an implied n.

The transcript that kept recording

The category’s best-documented failure is not hypothetical and involves named parties. In September 2024, Alex Bilzerian, an AI researcher and engineer, took a Zoom meeting with a venture capital firm that recorded the call using Otter.ai. After the meeting ended, Otter automatically emailed him the transcript, which contained hours of the firm’s private conversation held after he left, including confidential discussion of their business. He posted the account on 26 September 2024 and walked away from the deal. Otter’s position was that the firm could have disabled automatic sharing. The Washington Post reported it on 2 October 2024 and the AI Incident Database catalogues it as Incident 811. Otter reported more than 35 million users and over a billion meetings processed as of December 2025.

The transcription layer underneath these products is also demonstrably fallible. Associated Press reporting published on 26 October 2024 collected researchers’ findings on OpenAI’s Whisper model: a University of Michigan researcher studying public meetings found hallucinations in eight of every ten audio transcriptions, a machine learning engineer found them in more than half of over 100 hours of transcriptions, and a developer found them in nearly all of 26,000 transcriptions he created. The fabrications ranged from racial commentary to imagined medical treatments. OpenAI said it was continually working to reduce hallucinations and that its usage policies prohibit Whisper in certain high-stakes decision-making contexts. Those studies tested Whisper rather than any particular commercial meeting product, and no equivalent published error rate exists for Zoom, Teams or Otter. What they establish is that the speech layer everything else is built on can invent content.

Consent is the binding constraint

The open legal questions are more likely to shape deployment over the next year than any productivity figure.

• In re Otter.AI Privacy Litigation, No. 5:25-cv-06911 in the Northern District of California before Judge Eumi K. Lee, consolidated four suits on 22 October 2025 and heard a motion to dismiss on 20 May 2026. No ruling had been located as of 1 September 2026, so every claim in it remains an allegation.
• Those suits allege automatic joining of Google Meet, Zoom and Microsoft Teams calls without consent, indefinite storage, and collection of biometric voiceprints without written notice, under the federal wiretap statute, Illinois BIPA and California’s CIPA. Otter denies unauthorised access and says account holders must obtain permissions.
• Cruz v. Fireflies.AI Corp., No. 3:25-cv-03399, filed in December 2025 under Illinois BIPA, targets the Speaker Recognition feature and alleges voiceprints were taken from meeting participants who held no account. It is also undecided.
• Twelve US states require all parties to consent to a recording: California, Connecticut, Delaware, Florida, Illinois, Maryland, Massachusetts, Montana, New Hampshire, Oregon, Pennsylvania and Washington. A meeting that crosses one of those lines is a compliance question before it is a productivity question.
• Article 5(1)(f) of the EU AI Act prohibits systems that infer a person’s emotions from biometric data in the workplace or in education, and has applied since 2 February 2025. The Future of Privacy Forum’s analysis notes only two narrow exceptions, for medical purposes and for safety limited to life and health, both to be read narrowly.

Sources: Microsoft WorkLab · National Bureau of Economic Research · Harvard Business Review · TechCrunch · National Law Review