In 1975, engineer Fred Brooks wrote that adding programmers to a backlog project sets it back even more. His metaphor became famous: “nine women cannot carry a baby in a month.”
Half a century later, hundreds of companies are repeating the mistake with artificial intelligence. They deploy agents hoping to speed up their processes and discover, invoice after invoice, that coordination costs more than technology promises to save.
The lesson IBM learned the hard way
By: Gabriel E. Levy B.
In the mid-1960s, IBM bet its future on a family of computers called the System/360.
Fortune magazine described the project as a $5 billion gamble, a huge figure for the time. The hardware progressed as planned, but the operating system, dubbed OS/360, began to add additional delays to the existing ones.
Fred Brooks, the engineer leading the project, did what any manager would have done: he called for reinforcements. More than a thousand people came to work on the project, which consumed about five thousand man-years of accumulated effort.
Each new addition, required many weeks of training by the veterans and multiplied the meetings necessary to keep the project on track.
The schedule continued to run exponentially. A decade later, Brooks recounted the experience in detail in a book he called: “The Mythical Man Month”, where he formulated the law that today bears his name: “adding staff to a delayed software project delays it even more”.
The literary figure that he applied to explain this phenomenon was written in marble in the culture of engineering: “gestating a baby takes nine months no matter how many women are assigned to the task.” Brooks died in 2022, after being a Turing Award winner.
What is an agent and why is it trendy?
An AI agent is a program that goes a step beyond a simple automated chat agent. While the chatbot answers questions, the agent executes multiple or specific tasks: it reads emails, consults databases, fills out forms, compares prices and is meticulously in charge of meticulating those steps without a person intervening in each one.
The commercial promise seems to be essentially seductive. If an agent works as a digital employee who neither sleeps nor gets paid overtime, ten agents should perform as ten employees and a hundred as a small army. The investment figures reflect that enthusiasm.
The consulting firm Gartner calculated that global spending on generative artificial intelligence would be around 644,000 million dollars in 2025 and that total spending on this technology would approach one and a half trillion dollars that same year. Under this strange logic of abundant money, thousands of companies deployed agents in customer service, sales, accounting, logistics and human resources.
They finally ended up repeating the mistake Brooks described half a century ago: “It assumes that work can be broken into infinite pieces and that each new piece of the system adds capacity without adding costs.”
The arithmetic that no one wants to do
Brooks debunked the myth of the month man with basic elementary school arithmetic.
When two people collaborate, there is a single channel of conversation.
With five, the possible channels go up to ten. With twenty, to one hundred and ninety.
The formula, “n times n minus one divided by two,” grows much faster than the size of the team, with each channel consuming hours of explanations and misunderstandings.
Added to that cost is the on-ramp of learning: a newcomer produces little during his first weeks and, while he learns, steals time from the experts who are supposed to guide him. There remains, moreover, the nature of the work itself.
Some tasks could eventually work very well if they are distributed among many hands, such as harvesting a field or painting a building or even dismantling a machine, but others are sequential and do not admit of any division, no matter how many resources are allocated to them.
Artificial intelligence agents have so far demonstrated an inability to support efficient dispersed division of labor.
It is important to clarify that this does not imply that in the future they will be able to do so if the memorization systems of AI languages improve, but for now it is much further away than it seems.
Every new agent needs context, permissions, clean data, and operating rules—its own on-ramp.
Each agent connected to other agents or to people opens additional channels of coordination, now between humans and machines.
And the final step of almost any process, the human review of what the machine produced, remains as indivisible as ever. Someone must verify the result and answer for it, and that someone becomes the new bottleneck and that is repeated over and over and over again.
The numbers that nowhere add up
The independent laboratory METR put the theory to the test in 2025 with a controlled experiment: sixteen veteran programmers solved 246 real tasks, sometimes with artificial intelligence assistants and sometimes without.
Participants believed they had worked twenty percent faster thanks to the machine.
The stopwatch showed the opposite: they were nineteen percent slower, because reviewing and correcting the assistant’s proposals consumed more time than the tool saved.
The METR itself qualified the finding in 2026, with a larger sample and more recent models, and today argues that the negative effect was reduced, although evidence of a clear acceleration remains weak.
An MIT report calculated that ninety-five percent of corporate generative AI pilots produced no measurable return, a number that several academics question for their methodology, though few dispute the trend.
S&P Global found, particularly strikingly, that the proportion of companies that abandoned most of their AI initiatives jumped from seventeen percent to forty-two percent in the short period of a year.
Yes, only one year.
Gartner predicted that more than forty percent of projects with agents will be canceled by 2027 and denounced “agent washing,” the practice of renaming simple chatbots as agents.
Researchers from Stanford and BetterUp baptized the everyday symptom as workslop: machine content that appears to be finished work and forces the recipient to redo it, with an estimated cost of $186 per month per employee, that is, it was cheaper to have never implemented them.
A global reversal of the corporate world
Cases with their own names abound.
Swedish financial company Klarna announced in 2024 that its auto attendant was doing the work of 700 customer service agents. A year later, its executive director, Sebastián Siemiatkowski, recognized that the bet had gone too far and hired people again, because the quality of the service deteriorated.
Perhaps the most notorious and ridiculed case on social networks is that of the famous fast food company: McDonald’s, which tested for three years a voice ordering system developed together with IBM in more than a hundred restaurants in the United States and finally decided to turn it off in 2024, after videos of customers receiving absurd orders, including 2,510 nuggets, went viral and became the laughingstock of the internet.
In more systematic tasks they can work
Economist Erik Brynjolfsson measured the performance of more than 5,000 human call center operators and found that AI assistance boosted their productivity by 14 percent, with improvements of up to 34 percent among newbies.
The tool performs when the task is bounded, repetitive, divisible, and easy to supervise.
The debate is still open between those who believe that agents will bend Brooks’ law by allowing smaller teams and those who, like engineer Wes McKinney, creator of programming tools used around the world, warn that the bottleneck was never in the hands that type and that coordination only changed shape.
In short, the law that Fred Brooks formulated in 1975 is still in force because it describes human boundaries that technology does not erase. AI agents promise speed, but each demands context, permissions, monitoring, and coordination, and those costs grow faster than the benefits when deployment lacks method. Companies that cancel projects today learn the old lesson of OS/360: there are jobs that no budget can compress. Nine mothers are still unable to make a baby in a month.
References
- Brooks, F. P. (1975). The Mythical Man-Month: Essays on Software Engineering. Addison-Wesley.
- Association for Computing Machinery. (1999). Frederick Brooks, A.M. Turing Award Laureate. https://amturing.acm.org/award_winners/brooks_1002187.cfm
- Becker, J., Rush, N., Barnes, E. y Rein, D. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. METR. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
- METR. (2026). We are Changing our Developer Productivity Experiment Design. https://metr.org/blog/2026-02-24-uplift-update/
- MIT Project NANDA. (2025). The GenAI Divide: State of AI in Business 2025. MIT Media Lab.
- Gartner. (2025). Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
- Niederhoffer, K., Kellerman, G. R., Lee, A. y Liebscher, A. (2025). AI-Generated Workslop Is Destroying Productivity. Harvard Business Review. https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity
- S&P Global Market Intelligence. (2025). Voice of the Enterprise: AI & Machine Learning 2025. S&P Global.
- Forbes. (2025). Klarna Reverses AI Push, Says Customers Prefer Human Support. https://www.forbes.com/sites/quickerbettertech/2025/05/18/business-tech-news-klarna-reverses-on-ai-says-customers-like-talking-to-people/
- CNBC. (2024). McDonald’s to end AI drive-thru test with IBM. https://www.cnbc.com/2024/06/17/mcdonalds-to-end-ibm-ai-drive-thru-test.html
- Brynjolfsson, E., Li, D. y Raymond, L. (2025). Generative AI at Work. Quarterly Journal of Economics, 140(2), 889-942. https://www.nber.org/papers/w31161
- Gartner. (2025). Gartner Forecasts Worldwide GenAI Spending to Reach $644 Billion in 2025. https://www.gartner.com/en/newsroom/press-releases/2025-03-31-gartner-forecasts-worldwide-genai-spending-to-reach-644-billion-in-2025
- Gartner. (2025). Gartner Says Worldwide AI Spending Will Total $1.5 Trillion in 2025. https://www.gartner.com/en/newsroom/press-releases/2025-09-17-gartner-says-worldwide-ai-spending-will-total-1-point-5-trillion-in-2025
- McKinney, W. (2026). The Mythical Agent-Month. O’Reilly Radar. https://www.oreilly.com/radar/the-mythical-agent-month/
