Will AI Oversight Be the New Email Inbox Burnout?

Anyone who has used LLMs for any kind of extended project over the past three years has faced the dilemma of how much of the time gained through the efficiency of generative AI should be reinvested in verifying and fact-checking the outputs. And this makes me think of email.

I got my first job after college as an Editorial Assistant at Bantam Doubleday Dell publishing in New York. It was back in the days before they merged with Random House, when BDD was in that big building on Fifth Avenue. The one Jared Kushner’s father owned—666 Fifth Avenue with its famous “Top of the Sixes” restaurant.

I was the fresh-faced college kid, and low man on the totem pole, so not many people spoke to me except my boss. I kind of looked forward every day to several visits from Alicia, the mail clerk, who would swing by my desk with the memo cart. We were both low social status at the company so we kept track of each other.

Alicia would sometimes hand me a big brown envelope marked “Inter-office Mail” that had a bunch of names written on it. All but one of them would be crossed out: “Beverly Horowitz,” the Editor-in-Chief of the Books for Young Readers division, my boss. I would unwind the little waxy string wrapped around the paper buttons that held it closed, remove the Memorandums inside and set them on Beverly’s desk for her review.

By then, people were printing their memos at the departmental printer and having their assistant (me) make photocopies down the hall for distribution in the Inter-office Mail cart. The days of the typing pool, where hand written memos were sent to be edited and typed were fading from memory. So the total time it took to send and receive a business communication had fallen from about 3 or 4 days down to one. But it still took a day.

This was the late 90s, more than 30 years after Ray Tomlinson combined two existing programs—SENDMSG and CPYNET—to send individual electronic messages by combining the host’s name and the computer name with an @ symbol.

Email and Its Unintended Consequences

Email first took hold in academic and government research circles, like the ARPANET and DARPA communities. But the early business systems like IBMs Office System (OFS) were expensive and technical. A handful of breakthrough protocols made email more accessible: SMTP (Simple Mail Transfer Protocol) allowed messages to be safely and efficiently sent across networks. POP3 (Post Office Protocol 3) made it possible for email messages to be stored on a mail server and downloaded to personal devices. And IMAP (Internet Message Access Protocol) provided users access to email servers from multiple devices.

It took the success of personal computing to finally pave the way for the business use of email. The IBM PC first launched in 1981. Apple’s Macintosh introduced the mouse-driven graphical user interface in 1984, followed by Microsoft’s own Windows OS, in 1985, that came bundled with Microsoft Mail. Then came the internet browsers. Netscape Navigator launched in December of 1994 followed by Microsoft’s Internet Explorer the next year. Browser-based email services like AOL, Hotmail, and Yahoo Mail emerged finally making email practical for the masses.

Email had so many advantages over the previous forms of business communication. It was instantaneous compared to drafting, typing, and mailing which used to take days or weeks. It cost almost nothing. And, unlike phone conversations, it left a written trail that could be archived and even searched. Companies began using email not just for distributing internal memos, but for customer support, external business communications and, finally, even marketing campaigns.

However, the widespread adoption of email as a powerful “internet communication tool” (to adapt Steve Jobs’ classic description), helped launch another industry, as well: email inbox overload mitigation services.

Anti-Email

The first “spam” filters were keyword-based. Microsoft’s Outlook 97 started by flagging emails that contained words like free or win or guaranteed. Or they were IP-based. The “Mail Abuse Prevention System” (MAPS) started as a “blackhole list” of IPs that typically sent unwanted commercial messages. Email services started allowing users to create “whitelists,” as well, to help surface important communications from work or family amid the rising tide of less important communications.

Amateur advertisers had learned how to harvest email addresses from chat rooms, web pages, newsgroup archives, and service provider directories to send commercial messages blindly to millions of people at a time. All at virtually no cost.

“Anti-spam” startups and research groups began allocating significant resources to build more sophisticated ways of filtering out junk messages. Scoring and threshold systems were developed to combine suspicious indicators in the IP address, domain name, header information and envelope, routing directions, subject line, body text, and more to identify and filter out unwanted messages.

By 1996, an academic paper was published on the newly compelling topic of “email overload.” Steve Whitaker and Candace Sidner of Lotus Corp. described how the new task of email management was crippling executives, who were still trying to build virtual filing systems with topical folders for different email communications. One executive was quoted as saying, “I dedicate somewhere between minimally two hours at the outlying range, up to ten hours on any given day trying to stay on top of email.” Whitaker and Sidner’s research found that 53% of emails remained in the inbox, unfiled. 35% of folders contained only 1 or 2 emails (what they call “failed folders”). And users struggled to remember folder names and definitions. The result, the authors claimed, was that a system that was supposed to streamline business communications and decision making had had the opposite effect, with all the critical business information, tasks, and pending decisions lost somewhere in the inbox. Representative quotes from their study subjects included, “I don’t know where to put it. And.. by making a wrong decision, I could really forget about it…” and “It might as well be deleted as buried in this pile of junk.”

Corporate communications—meeting minutes, interoffice memos, letters to customers and partners, etc.—used to take a lot of effort, what with drafting, typing, addressing, mailing or distributing, etc. Email made that process fast and cheap. But instead of freeing up all that time for other more productive business activities, email simply moved the chokepoint from the pool of typists and mail couriers to the email inbox itself.

And, in fact, it magnified the problem, because email is so easy to draft and send that it invited all kinds of actors into the pool that weren’t there before. Not just malicious actors like spammers, either. Regular employees who wouldn’t bother to draft, type, and distribute an Inter-Office Memo could, without much thought at all, fire off an email to the whole team, setting off reply-alls like rows of dominoes that took even more time to read and reply to.

By the early 2000s, companies like Intel and the Wall Street Journal were spending millions of dollars on major corporate initiatives to train their employees to send fewer, more effective emails, as well. They branded such sessions “Email Etiquette Training.” Email footers began sporting reminders like, “No reply-all unless necessary,” and “No one-word replies.” No Email Day or NED was implemented in 2006 by PBD, an Atlanta-based fulfillment services company, and swept through the business press that year leading to hundreds of copycat programs. Business executives were fighting to reclaim their productivity from the scourge of email.

In 2011, CEO Thierry Breton of the IT company Atos even announced an ambitious plan to reduce internal email to “zero” over the next three years when internal analysis showed that his typical employee was receiving more than 100 emails each day, most of which added little value to the company. Breton drafted a plan and appointed thousands of internal “ambassadors” to evangelize a new approach featuring new tools. The main new tool was an internal messaging system called blueKiwi where the company could enforce more rigorous boundaries around internal communications. This led, of course, to other credential-based internal messaging systems for company communications like Slack and Microsoft Teams that have replaced the internal emails that replaced those Inter-office Memos way back when.

Constructive Friction

Stanford professors Robert I. Sutton and Huggy Rao call this retro-active imposition of credentials on otherwise open systems “constructive friction.” They point out that successful organizations manage friction, not just by removing it from key business processes that benefit from greater efficiency, but also by imposing friction on communications, processes, and even products that should be harder to make, or require MORE thought and review or vetting before being initiated: “Unfettered and overconfident leaders can squander a lot [of] cash…when they fall in love with flawed ideas and there are insufficient organizational speed bumps…” (The Friction Project, p. 7).

Sutton and Rao outline five costs to excessive speed that build up, like “organizational debt” the longer they go unaddressed:

1. Employee Burnout

2. Selfishness (because stressed people don’t have the bandwidth to be kind)

3. Bullying by Leadership

4. Bad Decisions

5. Lost Creativity (because creativity takes time)

Steven Covey is credited with having coined the phrase, “fast is slow and slow is fast” when talking about business processes. But some business leaders also point to the Navy SEAL training motto, “slow is smooth, smooth is fast” or even the old woodworker’s adage, “measure twice, cut once” to help build more careful, deliberative business cultures as a means to better market outcomes. Studies and use cases have been built around companies like LEGO and Basecamp and Patagonia as examples of organizations that take deliberate “slow growth” approaches as a more sure means of building shareholder value than their “ready, fire, aim” counterparts in the “move fast and break things” world of Silicon Valley. Dana Kanze at the London Business School has even shown a demonstrable correlation of unlawful discrimination suits against companies that have words like do it, fast, urgent, hurry, can’t wait and launch in their mission statements versus companies with words like careful, consider, right, evaluate, think and thorough in their mission statements.

Send All

I remember back in my Bantam Doubleday Dell days, when the company introduced a new messaging system. You could send a direct message to any other employee by name or any department or even the whole company.

This was several years later. Random House had been bought by the German multi-national conglomerate Bertelsmann and we had moved from Fifth Avenue to a swanky new building on Broadway in Times Square. Our communications had upgraded, too. We all had email addresses and were already fighting inbox fatigue so the new company messaging system felt like an end-around to get quick answers to questions without having to wait for someone to drain their inbox before finding your question in an email.

By then, I had moved up the ladder from Editorial Assistant to Assistant Editor. In addition to my title acquisition and editing responsibilities, I was serving on the “New Media” committee, figuring out how building educational internet resources could support the use of our books in schools and in book clubs. I was even volunteering as the team captain for an elementary school in the Bronx where BDD donated children’s books and paid for employees to read to the kids at lunchtime. My chief duty was to persuade fellow editors to give up their lunch hours for a good cause, and I was using the new-fangled messaging system to pester some likely do-gooders.

I’ll never forget one exchange I had with another assistant editor. Let’s call her Maureen. She was willing to give up some free time but she had been raised as an only child and didn’t know how she would handle a room full of children. It made her kind of panicky just to think about it. At the fateful moment, she typed, “I don’t know… I can just see myself trying to read with all those little monkeys climbing all over me.” And then, instead of clicking “Send,” she clicked “Send all” by mistake.

As I mentioned, the technology was new. These days there would be a “constructive friction” feature to invite you to pause and reflect. A dialogue box might pop up and ask, “Do you really want to send to all?” That’s all it would have taken to spare Maureen the pain and discomfort that was about to follow. If she had had a moment to reflect, she wouldn’t have sent that message.

But no such dialogue box popped up. Maureen’s message essentially went straight from her brain to the entire company unfiltered—all 1,300 Random House employees or so, most of whom had never met me or Maureen. That very afternoon Maureen and I were summoned into the office of the Vice President of Human Resources on the Executive Floor for a series of one on one “conversations.” I had never been to that floor before.

The location of the school in question and the unfortunate selection of the word “monkey” had combined to carry unavoidable racially disparaging overtones, and people were offended and mad. They let me off the hook relatively quickly. I hadn’t sent the message and was, after all, in charge of recruiting volunteers to help the school. But Maureen was written up and had a warning put into her file. Despite her protestations that she just meant that children can act like uncontrollable monkeys and that sending the message to everyone was the last thing on earth she would have wanted, she was unable to convince a number of important people that her comment wasn’t calculated to be offensive.

A little friction would have helped a lot that day.

Implications for AI Integration

All of this brings us to AI. There are a couple of things AI integrators can learn from the somewhat cautionary story of email.

One is that technologies take time to work their way into the culture.

I remember that as late as 1992, my wife’s boss at Alliance Capital was still dictating his emails. He had learned to rely on his assistant for all communications—minutes, memos, letters, and now emails. At the time, we couldn’t imagine but that his replacement would eventually type his or her own emails. And even at Random House, the interoffice memo cart continued making its rounds, living side by side with email and messaging for several more years.

But the most important lesson is that when you increase capacity at a bottleneck without redesigning the system around it, you don’t free the system. You simply move the bottleneck.

For the first time ever, businesses have access to professional tools that not only process, distribute, and store information, but can do research, pattern recognition, reasoning, and even decision-making itself. And it only takes minutes, sometimes seconds, to produce business ready documents. But all those documents still need to be read, reviewed, fact-checked and integrated into existing systems and operations.

Recently, I have had the new and singular experience of spending a week reading through and fact-checking materials that it took me or a colleague only part of a day to produce—hundreds, even thousands of pages. It struck me that in my former life, I would spend half a day fact-checking material that it took me a week to produce. It is not hard to see how organizations without a clear sense of what questions should be answered efficiently and which questions would be better off loaded with friction will get bogged down in all the material, just like our email inboxes.

And what will be the institutional pressures for move-fast organizations to NOT do that careful work of reviewing and vetting AI-produced materials? Along with hiring executives and consultants who understand the capabilities and limitations of new AI technologies and platforms and how to apply them to the unique goals and processes of their business, successful organizations will ALSO need to hire executives and consultants who understand how to build the human organizations around these new AI-Assisted workflows to maintain the efficiency gains.

The New AI Jobs: Architects, Not Just Editors

This summer, NYTimes reporter Robert Capps attempted to imagine a list of jobs that would be created by AI. He grouped them into three broad categories: Taste, Integration, and Trust.

While Capps is directionally correct, his categorization inadvertently reinforces the “Email Trap”—the idea that humans must manually process the flood of new information. His “Trust” category, for example, is filled with “auditors, editors, and fact-checkers.” This suggests a future where AI generates millions of documents, and we simply hire more humans to read them. That is not a sustainable future; that is just high-speed burnout.

If we look closer at the labor market required to sustain these systems, a more rigorous, functional taxonomy emerges—one that solves for the bottleneck rather than just staffing it. We don’t just need “tasters” and “checkers”; we need five distinct categories of builders and guardians:

1. Creation & Curation (Defining the “What”) Capps calls this “Taste,” but it’s really about Intent. We don’t just need designers; we need AI Experience Designers and Outcome Orchestrators. These aren’t people who tweak the final output; they are the people who design the prompt chains and workflows to ensure the AI produces differentiated value rather than generic commodity text. They move the “quality check” from the end of the process to the very beginning.

2. Implementation & Operations (The “Plumbing”) Remarkably, most discussions about AI jobs miss the blue-collar work of keeping the systems running. We don’t just need generic “integrators”; we need AIOps Leads and AI Integration Engineers. Just as we hire electricians to manage the power grid rather than hiring people to watch the lightbulbs, we need technical staff to manage the model performance, data pipelines, and latency.

3. Assurance & Governance (The “Guardrails”) This is where the “Burnout” battle will be won or lost. Capps suggests we need “editors” to verify AI output. But in a world generating billions of tokens, manual verification is impossible. Instead, we need to distinguish between Systemic Friction and Transactional Friction.

   1. Systemic Friction is for the 80% of volume. Guardrails Engineers write code that prevents the model from generating bad outputs (Software 3.0). This allows us to trust the system without reading every word.

   2. Transactional Friction** is for the 20% of high-stakes decisions. Here, we deliberately insert human review.

   3. By automating the safety for the volume, we preserve human attention for the decisions that matter.

4. Institutional Mediation (The “Social Contract”) AI systems don’t exist in a vacuum. We will see the rise of AI Risk & Policy Officers and Workforce Negotiators who manage the friction between the algorithm and society. These roles don’t touch the code; they define the liability envelope and the labor agreements that allow the code to run.

5. Education & Enablement (The “Upgrade”) Finally, to prevent the entire workforce from drowning in these tools, we need AI Fluency Coaches. These aren’t IT support staff; they are mentors who teach professionals how to act as the “Senior Partner” to their AI “Associates,” ensuring humans remain in control of the autonomy slider.

Moving the Choke Point

This distinction matters because it changes how we organize for the future.

If we follow the “Email Model,” we will hire armies of people to read, review, and fret over AI content, recreating the inbox misery of the late 90s on a grander scale. But if we follow the “Industrial Model,” we will hire architects to build systems that are safe and reliable by design.

The organizations that win won’t be the ones with the fastest “fact-checkers.” They will be the ones that have built the Assurance & Governance layers that allow them to trust the system without reading every word.

Sources:

• Choudary, S. (2025). Reshuffle: Who Wins when AI Restacks the Knowledge Economy. FutureFirst Books.

• Sutton, R., & Rao, H. (2024) The Friction Project: How Smart Leaders Make the Right Things Easier and the Wrong Things Harder. St. Martin’s Press.

• The History of Email and Its Impact on Communication. (2024). MailChimp. https://mailchimp.com/resources/history-of-email/.

• Rise of Personal Computing and the Internet. (2025). Fiveable. https://fiveable.me/history-of-american-business/unit-16/rise-personal-computing-internet/study-guide/E3VhIlx3Xm8QV0XO.

• Zeldes, N. (2012). No Email Day: a Misunderstood but Promising Solution to Information Overload. https://www.nathanzeldes.com/blog/2012/09/no-email-day-a-misunderstood-but-promising-solution-to-information-overload/.

• Morgan, J. (2015, April 10). How This Company Is Trying To Eradicate Email. https://www.forbes.com/sites/jacobmorgan/2014/09/29/how-this-company-is-trying-to-eradicate-email/.

• Capps, R. (2025, June 17). A.I. Might Take Your Job. Here Are 22 New Ones It Could Give You. https://www.nytimes.com/2025/06/17/magazine/ai-new-jobs.html.