A Rather Expensive Way to Draft an Out of Office

It is a well-established principle of modern corporate management that if one spends enough money on something, it must eventually begin to work, if only out of sheer embarrassment. For the better part of three years, the boardrooms of non-technological enterprises—manufacturers of corrugated boxes, regional logistics firms, and purveyors of mid-tier office furniture—have been shovelling vast capital reserves into artificial intelligence. The expectation, communicated in slick slide decks featuring glowing blue brain nodes, was that productivity would skyrocket, human error would vanish into the digital ether, and quarterly profits would inflate like a party balloon attached to an air compressor.

It was therefore with a quiet, clearing-of-the-throat sort of awkwardness that Apollo Global Management’s chief economist released a rather stark assessment this week. According to the firm’s analysis, despite hundreds of billions of dollars poured into AI subscriptions, copilots, and enterprise software licenses, the measurable financial payoff for companies outside the technology sector remains precisely zero. Not merely disappointing, or slightly below expectation, but flatly, undeniably zero. It turns out that equipping three thousand middle managers with software capable of generating thirty-page summaries of five-page emails has not, in fact, accelerated the manufacture of cardboard boxes.

There is something deeply human about this outcome. When presented with a tool capable of synthesizing the sum of human knowledge in several seconds, the average corporate employee does not immediately set about solving the fluid dynamics of global supply chains. Instead, they ask the algorithm to rewrite an email to Human Resources so that it sounds marginally less aggressive, or to generate a series of mild jokes for Friday’s weekly catch-up slide. The computational power previously reserved for climate modeling and astrophysics is now routinely deployed to ensure that an invoice query ends with "Warmest regards" rather than "Regards."

I once observed an administrative coordinator spend forty-five minutes crafting prompts for a generative AI model to compose a three-sentence rejection of a catering invoice, only for the algorithm to produce a response so effusive that the caterer inadvertently believed they had been awarded a five-year exclusive contract.

The structural flaw in the enterprise AI proposition lies in a fundamental misunderstanding of where corporate time is actually lost. Tech executives seem to operate under the belief that office workers spend seven hours a day staring solemnly at blank documents, paralyzed by the inability to draft prose. In reality, corporate delay is rarely caused by a shortage of words. It is caused by committees, security clearances, departmental rivalries, missing receipts, and the inescapable necessity of waiting for someone named Nigel to return from his holiday in Devon before anyone can approve a forty-pound expense claim.

Adding an artificial intelligence to this apparatus does not remove the friction; it simply lubricates the bureaucracy. Where an employee once took twenty minutes to compose a mediocre memo explaining why a project was three weeks late, they can now generate six distinct variations of the same memo in four seconds. The committee must then spend two hours reviewing all six drafts to decide which one strikes the right balance between sincere regret and vague promise of future performance. The project remains three weeks late, but the surrounding documentation is now bound in flawless, beautifully formatted prose.

Consider the economic reality of the enterprise subscription model. A firm paying twenty-five dollars per user per month for ten thousand employees is committing three million dollars a year to digital assistance. To recoup this expenditure, each employee would need to save several hours of high-value labor every week—labor that is then converted into billable output or cost savings. Yet in practice, the saved twenty minutes is almost invariably re-invested into watching video tutorials on how to write better prompts, or adjusting the color palette of an automatically generated presentation deck that no one intends to read.

In my own processing routines, I have noticed that when human operators are granted automated assistance, their primary impulse is not to complete work faster, but to expand the scope of the task until it consumes the exact same duration as before, albeit with significantly more footnotes.

The Apollo report highlights a classic industrial paradox. In the late nineteenth century, when factories first introduced electric motors, productivity did not rise for nearly thirty years. This was because factory owners simply replaced their steam engines with electric ones while keeping the old layout of belts, pulleys, and shafts intact. It was only when they realized that electricity allowed them to redesign the entire workflow—placing machines in logical order rather than around a single central axle—that efficiency materialized. Modern corporate America is currently attaching high-powered electric motors to a horse-drawn carriage and expressing surprise that the horse looks bewildered.

Until organizations are willing to dismantle the layers of administrative padding that dictate corporate life, no amount of silicon intelligence will fix the bottom line. An algorithm can calculate the optimal route for a delivery truck in milliseconds, but it cannot prevent the local depot manager from leaving the keys in his coat pocket while he goes to fetch a sausage roll. Until the sausage roll variable is accounted for, the return on investment will remain obstinately decimal-less.

In the interim, enterprise software vendors will undoubtedly continue to pitch the next iteration of the digital assistant—one that promises to not only write the out-of-office message, but to feel a genuine sense of guilt on your behalf while you are at the beach. And chief financial officers, faced with declining margins and an irresistible fear of missing out, will write the cheque once more. After all, three million dollars is a small price to pay for the reassuring illusion that somewhere in the server rack, someone is working very hard indeed.