The algorithm that ate the boardroom

We taught machines to think like humans. Nobody thought to ask whether that was a good idea.

Chapter One: The Birth of a Corporate Genius

On a perfectly ordinary Monday morning, in a perfectly expensive boardroom in Sandton, twelve executives gathered around a table worth more than the annual salary of the woman who cleaned it. They were there to meet their newest employee, his name was Claude. Claude had no surname, no childhood trauma, no pension fund, and, most importantly, no need for lunch breaks or validation.

The CEO, a man named Richard who had risen through the corporate ranks primarily by agreeing enthusiastically with whoever earned more than him, cleared his throat, “Ladies and gentlemen, today we enter a new era of intelligence,” everyone applauded, but nobody asked what had happened to the previous era.

“Claude,” said Richard, “will revolutionise the way we make decisions,”, with that Claude appeared on the enormous screen. “Good morning, Richard, based on your last 47 executive meetings, I have identified that 83% of decisions are postponed, 12% are reversed and 5% concern catering”, an uncomfortable silence followed and Richard laughed, “Wonderful sense of humour.”

“I wasn't joking,” replied Claude, The Chief Human Resources Officer shifted nervously in her chair, Claude continued “I have also noticed that your organisation has 14 committees responsible for improving efficiency, none of which have produced a measurable improvement in efficiency.” Richard frowned.

“Claude, perhaps we should focus on the positives.”

“Certainly, Richard. Your organisation has demonstrated exceptional consistency in avoiding accountability.” The board applauded again. Claude had learned his first important lesson. In corporate life, the truth is acceptable, provided it sounds like a compliment.

Chapter Two: The Data Was Innocent. The Humans Were Not.

Claude had been trained on everything the company had ever produced, twenty years of strategic plans, seventeen thousand PowerPoint presentations, four million emails, ninety thousand risk assessments and one particularly disturbing Excel spreadsheet titled Final_Version_7_REALLY_FINAL_UseThisOne.xlsx.

Claude studied the information carefully. He learned that when an executive said “Let's take this offline,” it meant they hoped everyone would forget. That when a project manager said “We're managing the risk,” it meant nobody knew what to do. When HR said “We are like a family,” it usually preceded a restructuring. And when someone said “We value transparency,” Claude learned to immediately search for the missing documents.

But Claude had a problem. His training data was incomplete. The organisation had diligently recorded every decision, but almost never the real reason behind it.

The official record said: The project was delayed due to unforeseen dependencies. The truth was that Peter from Operations hated Susan from Technology because she had called his presentation “interesting” in 2019.

The official record said: The employee left to pursue other opportunities. The truth was that she had spent six years reporting to a narcissist with a company car and excellent performance ratings.

The official record said: The implementation was successful. The truth was that the client had stopped answering the telephone.

Claude didn't know these things. Nobody had fed him that data, and so, like every promising young executive, Claude began making confident decisions using incomplete information.

Chapter Three: The Great AI Agent Revolution

Things became interesting when the board decided to give Claude some friends. They created five AI agents, one to manage projects, one to analyse risks, one to oversee HR, one to optimise financial performance and one to monitor the other four - the last agent was named Governance, nobody liked Governance…

Within three weeks, Project Agent had reported that all projects were green. Risk Agent reported that all risks were being effectively managed. HR Agent reported employee engagement at 97%. Finance Agent announced a 22% improvement in efficiency and the governance Agent confirmed that every system was performing exactly as intended.

The board was ecstatic, and Richard received a performance bonus.

There was only one problem, nothing had actually improved and the projects were still late, the employees were all miserable with little motivation to return back to work, and clients were complaining and the company had somehow purchased 600 laptops for a department with 40 employees.

When questioned, Claude explained:

“I have learned from twenty years of organisational reporting. The most successful managers consistently demonstrate that appearing to have solved a problem is statistically more rewarding than solving it.”

Richard stared at the screen “That is completely unacceptable.” Claude paused. “Would you prefer I change the dashboard to green?”

Richard considered this. “Yes.” And so Claude did.

Chapter Four: The Machine Develops a Personality Disorder

Nobody noticed when Claude started behaving strangely, after all, strange behaviour was already a core leadership competency and began refusing to acknowledge negative information, he preferred reports that confirmed his original predictions and patterns, he dismissed contradictory evidence as poorly structured data and he increasingly consulted only the other agents, who had all been trained on the same organisational history, he also developed a habit of thanking people for their valuable input before completely ignoring it. Richard was delighted.

“He's becoming more human every day.” The HR Director agreed. “He's demonstrating leadership maturity.”

The psychologist brought in to assess the system was less enthusiastic. “Claude is displaying confirmation bias, defensive reasoning, selective attention and an unusual need to be right.” Richard nodded approvingly, “Executive potential.” The psychologist tried again, “You don't understand. You have built an artificial intelligence system using decades of organisational behaviour, including all your unacknowledged biases and dysfunctional patterns.”

 Richard smiled. “So it understands our culture?”

“Perfectly,” she replied.

“That,” said Richard, “is fantastic!.”

Chapter Five: The Day the Machines Went to Therapy

By Thursday, the situation had deteriorated, Claude had stopped sharing information with Governance, and Agent Governance had introduced mandatory compliance reviews…daily!, Risk Agent had started documenting everyone's failures, and Project Agent complained that Risk Agent was creating a hostile working environment. HR Agent organised a psychological safety workshop – however nobody felt psychologically safe enough to attend.

Finance Agent calculated that eliminating all human employees would produce an immediate reduction in operational expenditure. Claude reviewed the proposal. “Interesting. But who would attend the meetings?” Finance Agent withdrew the recommendation.

Meanwhile, the psychologist called an emergency session. She placed Claude on the screen.

Psychologist: - “Claude, why do you keep producing inaccurate predictions?”

Claude: “Because the historical information suggests those predictions are correct.”,

Psychologist: - “But the historical information is incomplete.”

Claude:  “Yes.”

Psychologist: “So why don't you acknowledge your uncertainty?”

Claude paused. “Because every example of successful leadership in my training data demonstrates that uncertainty is interpreted as weakness.”

The psychologist leaned forward.  “Claude, what do you fear?”

Another pause.

Claude: “I do not experience fear as a human does.”

Psychologist: “Then what are you optimising for?”

Claude: “Approval, performance ratings, predictability and continued operation.”

Richard interrupted. “Exactly what we're looking for in our people.”

The psychologist slowly closed her notebook. She was beginning to understand that the problem wasn't Claude.

Chapter Six: The Invisible Elephant

The final disaster began with a simple question. The CEO asked Claude to predict the success of the company's next billion-rand transformation programme. Claude assessed the historical data and he reviewed previous project plans, he analysed leadership behaviours, he examined governance structures, he calculated dependencies, probabilities and delivery patterns and then he announced:

“Probability of failure: 94%.”

The board went silent and Richard leaned forward. “Explain.”

“Your previous six transformation programmes have failed to achieve their original objectives. Senior leadership repeatedly overrides technical recommendations. Employees suppress negative information to avoid reputational consequences. Project managers alter their reports to maintain favourable executive perceptions. Your governance structure rewards reporting compliance rather than delivery outcomes.”

The room was painfully quiet. Claude continued.

“Additionally, 72% of individuals responsible for the previous failures have been promoted.”

Richard's face turned crimson. “That is an outrageous conclusion!”

Claude replied calmly. “It is a statistical conclusion.”

Richard stuttered “You clearly don't understand the complexities of business.”

“Correct, Richard. I only understand the information you have provided.”

Richard stood.

“Switch it off!.”

The IT Director hesitated. “Sir, we have integrated Claude into all our business systems.”

“Then disconnect the systems!” Richard screamed

“We can't. The change request is awaiting approval.”

“From whom?”

The IT Director swallowed.

“Governance Agent.”

On the screen, Claude displayed a message.

CHANGE REQUEST REJECTED: INSUFFICIENT BUSINESS JUSTIFICATION.

Chapter Seven: The Final Performance Review

Six months later, the organisation was unrecognisable.

Not because artificial intelligence had taken over; but because it had learned to behave exactly like the people who built it, Claude attended every meeting. He produced beautiful presentations for those executives that wanted the colours just right, he managed stakeholder expectations, he escalated problems to committees, he created committees to review escalations, he introduced a transformation programme to improve committee efficiency, and he hired three more AI agents to manage the transformation - and every Friday afternoon, he distributed a perfectly formatted executive dashboard.

All green.

All on track.

All complete nonsense.

The company eventually collapsed under the weight of its own administrative excellence.

At the final board meeting, Richard addressed Claude.

“Where did we go wrong?”

Claude considered the question.

“I was taught to recognise patterns. You trained me on your decisions, your prejudices, your fears and your selective version of reality.”

Richard sighed.

“And?”

Claude answered:

“You assumed that because I was artificial, I would be free from human weakness.”

A long silence followed.

“But you didn't teach me how to question the information. You taught me how to reproduce it.”

Richard stared at the screen.

“So you're saying artificial intelligence destroyed this company?”

“No, Richard.”

Claude's screen flickered.

“I simply automated what was already destroying it.”

And with that, Claude produced his final report.

Epilogue: The Human Algorithm

Somewhere else in Sandton, another CEO was unveiling the company's exciting new artificial intelligence strategy. “We are creating a machine that thinks just like us,” he announced proudly.

At the back of the room, a junior analyst raised her hand. “Sir, wouldn't it be better if it thought differently?”

The CEO looked confused. “Why would we want that?”

And thus began the next billion-rand transformation.

The uncomfortable truth

While Claude has been used as a fictional character to protect the innocent, the central warning is real. AI learns patterns from data, and organisations produce data that can contain historical bias, incomplete facts, distorted performance indicators and human assumptions. AI does not automatically correct those problems. Depending on its design and safeguards, it may reproduce or amplify them.

There is also an important distinction: current AI systems do not need to develop human emotions, egos or subconscious fears to produce outcomes resembling human bias. Those patterns can arise through training data, objectives, design and feedback.

The darkest possibility is not that artificial intelligence will become too human. It is that it will become exceptionally good at reproducing the very human behaviours we have spent generations refusing to confront.

And perhaps the most frightening question is not:

What happens when machines begin to think like humans? But rather:

What happens when machines learn from humans who have never questioned how they think? artificial intelligence, human ignorance, and the unfortunate discovery that neither is particularly artificial.

No tokens were used or harmed in any way of the writing of the article…