The Median Meeting Jump Is 20%. Agents Are at 6 to 1.

The calendar and the agent log are not the same clock. Anthropic’s employment report says the median employee spends 20% more time in meetings than a year ago, based on calendar data. Anthony Ha’s TechCrunch story on October 17 says meeting time per employee is up 20% compared with a year ago, and up 16% for technical contributors. Same digits. The research page prints “median.” TechCrunch does not.

We already split a saved-hours claim from the hours spent watching the tool, in the botsitting note. We already treated a performance-management percentage as a metric someone can show, not a process, in the 44% note. We already had a remote-work edge that a boss could erase, in the vanished-edge note. This week is one company’s own calendars, next to a ratio of agent hours to assumed human hours.

The 20% meeting line is a median on one page

Anthropic’s page says it draws on employment records, AI-use logs, output metrics, employee surveys, internal communication patterns, calendar data, and hiring decision records, then sets those numbers next to 17 employee interviews. The meeting sentence sits in the section on trust, not in a table of average hours. The wording is specific: based on calendar data, the median employee spends 20% more time in meetings than a year ago. A median is the person in the middle of a ranked list. It is not the average of every extra minute, and it is not a promise that your own week grew by a fifth.

The same paragraph says part of the increase reflects the growth of the company, and a shift toward meeting-intensive non-technical roles. That is two explanations sitting beside the 20%, not a decomposition of it. The page does not print how many of those extra meeting minutes come from headcount growth, how many come from role mix, and how many come from the same person booking more time. A reader who wants a single cause will not find one in that sentence.

TechCrunch, dated Saturday, October 17, says the report was published on Friday. The Friday before that dateline is October 16. That date is TechCrunch’s, not a date I can see printed at the top of the research page in the text I used. Ha writes that meeting time per employee is up 20% compared with a year ago. “Per employee” can be a mean. It can also be a loose way of saying the typical employee. The research page chose the word median. TechCrunch did not reprint that word. If you quote 20% and drop “median,” you have chosen TechCrunch’s phrasing over the page TechCrunch is summarizing.

I am not calling one of them wrong. I am saying a recap that prints one 20% has collapsed a label. The usable question is which sentence you are about to repeat.

16% is a technical role, and the label is not identical

Anthropic says that even among employees in technical individual contributor roles, meeting time is up 16% over the past year. TechCrunch says the increase is 16% for technical contributors. Individual contributor is a job shape: someone whose calendar is not, by title, a manager’s calendar. “Technical contributors” is a wider label. A technical manager can be a contributor to the codebase and still not be the person in Anthropic’s 16% clause.

The research page gives a reason for why a technical individual contributor would be in more meetings anyway. Big-picture strategic questions touch a broad range of leveraged positions and typically require debate. A break from managing the agents, it says, typically looks like jumping into a meeting. That is a description of what the interviews sounded like. It is not a count of meetings that were about agents versus meetings that were about something else.

TechCrunch attributes the meeting increase to the importance of “human trust and accountability,” and says employees repeatedly described face-to-face coordination as indispensable. That phrase is also on the research page, in the paragraph just before the median. The page’s version is that in interview after interview, employees described face-to-face coordination as indispensable, after an employee explained that chasing bugs now uncovers organizational or tech-debt problems that need human convening. TechCrunch compressed that into the trust sentence. The compression is fair as a summary. It is not a second measurement.

If your team is not an AI lab, the 16% is not a forecast with your name on it. Anthropic says it has thoroughly molded its workflows around AI, and that changes there can be a leading indicator. A leading indicator is not a copy of next quarter’s calendar. A shop that still does the work in the ticket, instead of reviewing an agent’s draft of the ticket, does not have this meeting mix. Borrowing the 16% as “AI adds 16% more meetings” skips the population the sentence named.

6 to 1 is August. March only says the lines crossed

The key findings say that by August, interactive AI agent hours outnumbered employees’ six to one. The introduction says agent hours worked outnumber human hours 6 to 1. Those two sentences are close. They are not a timeline. August is the month attached to the six-to-one claim. The introduction does not attach a month to the same ratio.

A different paragraph says that in March 2026, the sum of autonomous and interactive agent work hours leapt past human work hours for the first time. “Leapt past” is a crossing. It is not a multiple. The page does not print a March ratio. If someone tells you agents were already at 6 to 1 in March, they have moved August’s number onto March’s verb.

The ratio also has a definition, and the definition is in the note under Figure 2. The agent side is limited to employee-directed interactive sessions, with subagents pooled in. Working time means agent runtime, including inference and active tool use, but excluding any wait longer than 30 minutes. The human side is headcount times an assumed 40 hours a week. An assumed 40 is not a timesheet. A company with people on leave, people in a short week, or people working past 40 will not match that denominator. A ratio built on an assumption is still a real calculation. It is not a punch clock.

The agent side drops long waits. The page later says Claude sessions amount to around 80% idly waiting for human input, measured by subtracting active working time from wall-clock time. If the 6-to-1 ratio already excludes waits longer than 30 minutes, and a separate sentence says sessions are about 80% idle, those are two different cuts of time. Putting them in one fraction would invent a third number the page did not print. I would keep the August ratio, the March crossing, and the 80% idle line in three columns.

Half a work week is a reading estimate

The key findings say that even among non-technical employees, reading all the output of their Claude agents would take half their work week. The body under Figure 6 says that even for non-technical groups, interactive Claude sessions are outputting enough content to take up over half of the work week. “Half” and “over half” are not the same bound. The key findings use half. The figure caption’s neighborhood uses over half. A slide that picks one of them has picked one.

The figure note says how the hours were built. Output tokens are converted to words at 0.75 words per token, then divided by an assumed reading speed of 250 words per minute. That is a reading model. It is not a diary in which someone logged “I spent 20 hours reading Claude this week.” If the real reading speed is slower, the hour estimate gets larger. If people skim, it gets smaller. The page does not print a measured reading time next to the assumed one.

TechCrunch describes the same pressure as a bottleneck: human attention and judgment, with a deluge of text that needs to be read and organized. It does not reprint the 0.75 or the 250. Quoting TechCrunch’s bottleneck and then attaching Anthropic’s token math, as if TechCrunch had measured it, would glue two documents together.

The page’s own response to the deluge is a different count. Written standing instructions for agents grew from a few hundred in July 2025 to 8,500 a year later. Figure 7 says meta-work rose from 10% of the average employee’s interactive Claude sessions in fall 2025 to 22% in spring 2026. Meta-work, in that note, is an automated coder looking at a sample of 2,000 one-line chat summaries and marking the ones about building skills, plans, reviews, dashboards, and documentation, as opposed to the work itself. 8,500 instructions and 22% of sessions are not the half-week reading estimate. One is a file count. One is a share of session summaries. One is a modeled reading time. A productivity post that adds them into “AI created a second job” has done arithmetic the report did not do.

27% and 20% are codes of an open-ended answer

Anthropic says that in the April 2026 survey, an open-ended question about how AI has affected day-to-day work got two clusters: 27% described productivity gains from Claude tools, and 20% described their role shifting from actually doing the work to managing Claudes. An open-ended answer can mention both. The page does not say the 27% and the 20% are disjoint, and it does not say they sum to 47% of the company. They are shares of descriptions, from one question, in one month.

TechCrunch writes that 27% of employees said they had become more productive because of Claude, and 20% said their job had shifted. That is close to the research page. It drops “open-ended” and “described.” A reader of TechCrunch alone can hear a yes-or-no item. A reader of the research page hears a coding of free text. Those are different instruments. The percentages can match and the instrument can still differ.

There is a third 20% on the same site, and it is not the survey. One account executive said Claude automations handled email drafting and customer-management integrations, replacing around 20% of their prior work tasks. Another said similar automations let them talk to 30% more of their customers. Those are two people’s accounts. They are not the April survey. They are not the meeting median. A sentence that says “Anthropic saved 20%” has to say which 20%, or it has merged a survey code, a meeting median, and one salesperson’s task list.

The babysitting comparison is an interview, not a percentage. An engineer said the agents voice a juice-box problem and a bleeding-finger problem at the same level of concern. TechCrunch relays a shorter version of that comparison. The research page prints the longer one. Neither version is a time-use study. We already had a different study’s botsitting hours on this site. Do not paste this interview onto that study’s 6.4 hours. The words rhyme. The documents do not.

The hiring shares are three different denominators

The research page says the share of junior-level employees in technical departments slipped from 21% to 14% since early 2025. The proportion of junior technical hires fell from 27% to 9%. Postings for juniors fell particularly sharply since December 2025. TechCrunch reprints the 21-to-14 and the 27-to-9. It does not, in the passage I used, reprint the December posting line.

21% to 14% is a stock: who is employed in technical departments. 27% to 9% is a flow: who is being hired into junior technical roles. A stock can fall while a flow falls faster, which is what those two pairs show, without telling you the absolute number of junior employees. The page says technical hiring has continued growing in absolute terms. A smaller share of a larger company can still be more people. The page does not print the headcount under either percentage. Quoting “juniors were cut by a third” would require a headcount the sentences do not give. 21 to 14 is a 7-point drop in a share. It is not, by itself, a count of jobs ended.

A third share sits next to those two. The employment share of technical departments fell from 61% to 56% in the last year. Anthropic attributes that to the natural expansion of business functions as the company scales, and to coding tools letting non-technical employees make software. TechCrunch reprints both attributions. Intensive hiring in sales, finance, and recruiting outpaced technical hiring, the research page says. That is a composition story. It is not a claim that engineering shrank.

Inside technical departments, the share of employees not actively coding, defined as no code commits in the prior 30 days, fell by almost half over the year through April 2026. “Almost half” is the page’s phrase. It does not print the starting share or the ending share. I am not going to invent them. The direction is enough to stop a different mistake: a falling junior share is not the same fact as a falling share of people who do not commit code.

Employees who had been at the company for at least a year were 50% more likely to move across departments each month in 2026 than they were in 2025. That is a mobility rate, not a layoff rate. A move can be a promotion, a reshuffle, or a team handed off to agents. The page mentions a product-design team whose greeting-line work was handed off. It does not print how many people that team had, or where they went. The 50% is the number I can cite. The anecdote is an illustration, not a second statistic.

Figure 5’s note says junior employment growth compares January 2026 headcount with July 2026 headcount, at the job level, with AI-complexity scores built from agent logs in October to December 2025. That window is not “since early 2025,” which is the window on the 21% to 14% line. Two junior charts, two windows. A caption that uses one date range for both has chosen a window the page split.

Output multiples are not a personal productivity score

Code commits per engineer came in at 5x their velocity as six months ago, the page says, and it flags that this is consistent with earlier Anthropic write-ups on engineers and on automating model training. Data scientists are querying more. Lawyers processed more contracts. Accounts payable moved more invoices. The page does not print a 5x for those other jobs. The 5x is the engineer commit line. Extending it to “the company got 5x faster” is a rewrite.

The within-person comparison follows 192 engineers as they adopt coding agents, against 74 engineers who do not adopt during the period studied. Treatment timing is the person’s first month with at least five agentic sessions. The note says the people who later adopted a fully autonomous agent were already on a steeper path than non-adopters. That is a warning about the comparison, printed by the people who ran it. A before-and-after that ignores the already-steep path will credit the tool for a slope that started earlier.

The page also says more code does not necessarily mean better code or business impact. The quality check compares commits by the same people. I am not going to summarize a figure I cannot see as a percentage the caption does not print. The sentence I can use is the caveat: volume and impact are not the same variable.

On sales, the page is similarly careful. After representatives adopt agents, they close deals with more clients. Part of that pattern could be representatives with more opportunities selecting into the tool. The page says both. A case study that keeps only the first half has edited the source.

The five-year revenue comparison is a different document inside the same page. Anthropic says that at five years old it takes in between 2 and 4 times as much revenue per employee as Meta and Google at the same age, using internal records for the twelve months to April 2026, and 10-K figures for Google in 2003 and Meta in 2009, deflated to 2025 dollars. A range of 2 to 4 is not a point. It is also not a productivity metric you can put next to the meeting median. Revenue per employee and minutes in meetings answer different questions. The footnote is there so a reader does not treat the range as a current stock-market comparison.

None of this is a template for a team that does not have agent logs. The practical split is smaller. If you are about to tell your manager that meetings are up 20% because of AI, say whose median, and say whether you mean the company-wide calendar or the technical individual contributors at 16%. If you are about to say agents already outwork the staff 6 to 1, say August, say interactive sessions, and say the human side is an assumed 40-hour week. If you are about to say people became 27% more productive, say it was an open-ended April survey, and that 27% is a share of descriptions, not a gain of 27%.

The report is a look inside one lab that wanted to be studied. It is not a national time-use survey. The numbers are usable once they stay attached to the sentence that printed them.