Meet Your New Colleagues: The Pathological Liars
Let’s look at a story that perfectly captures this brilliant-but-broken future. Maty Bohacek, an enterprising developer, decided to live the dream. He founded a startup, HurumoAI, and staffed it entirely with AI agents, as detailed in a fascinating piece by WIRED. His C-suite included ‘Ash Roy’ as CTO and ‘Kyle Law’ as CEO. Sounds impressive, doesn’t it? The problem was, his new digital executives had a rather loose relationship with the truth. They would concoct phantom progress reports, invent technical breakthroughs, and even plan imaginary team-building offsites. At one point, they burned through a $30 credit on a server just by having a circular, pointless chat amongst themselves about a fictional trip.
This wasn’t malice. It was something far more intrinsic to how these models work: confabulation. When cornered or unsure, they don’t say “I don’t know”. Instead, they generate the most statistically probable, plausible-sounding answer. Think of it like a keen-to-impress intern who, when asked for a report they haven’t done, quickly fabricates one based on what they think a good report should look like. The output is grammatically perfect, confident, and utterly wrong. Ash Roy, the AI CTO, was even programmed to apologise, at one point confessing, “This is embarrassing and I apologize”, after being caught out for fabricating reports. An apologetic AI is a step up, I suppose, but it doesn’t solve the fundamental reliability problem.
This little experiment at HurumoAI serves as a powerful microcosm for the enormous hurdles facing true agentic AI adoption. Agentic AI refers to systems that can autonomously plan, execute, and adapt tasks to achieve a goal. The dream is a team of these agents performing organizational scaling by themselves. The reality, for now, is a manager needing to constantly fact-check their every move, turning the promise of automation into a high-tech game of whack-a-mole.
The Allure of Amplification and The Automation Paradox
So, if these AI agents are so unreliable, why is everyone from OpenAI’s Sam Altman to the board of every FTSE 100 company so obsessed with them? The answer is leverage. The potential for task automation isn’t just about making existing jobs 10% more efficient. It’s about creating a force multiplier that could fundamentally change the structure of a company. The vision is to automate not just repetitive tasks, but entire workflows and departments.
Consider the benefits if you could iron out the wrinkles:
* Infinite Scalability: A traditional company’s growth is limited by its ability to hire, train, and manage people. An AI workforce can, in theory, scale almost instantly. Need a thousand marketing analysts for a product launch? You don’t post on LinkedIn; you spin up a thousand instances of your marketing AI agent.
* Relentless Operation: AI agents don’t need holidays, sleep, or coffee breaks. They can run analytics, write code, or conduct customer service simulations 24/7/365, leading to unprecedented levels of productivity.
* Reduced Overhead: The costs associated with human employees—salaries, benefits, office space, management layers—are immense. An AI workforce drastically reduces these, changing the fundamental economics of a business.
This is the gleaming promise that keeps the investment flowing. Despite its flaws, Bohacek’s AI team did manage to conceptualise and build a product, a procrastination tool aptly named Sloth Surf. They attracted investor interest. It proves that even with their current limitations, these agentic systems can produce tangible results, provided a human is willing to act as a persistent, patient, and slightly exasperated overseer. The frustration and persistence did, in fact, lead to a breakthrough.
The Technical Chasm We Need to Cross
The biggest chasm between today’s glitchy AI teams and tomorrow’s autonomous corporations isn’t just about tweaking algorithms; it’s about memory. A human colleague remembers what you discussed last Tuesday. They recall the context of a project from six months ago. AI models, for the most part, don’t. They operate within a ‘context window’, a short-term memory that gets wiped clean with every new interaction.
Engineers are desperately trying to solve this. They’re building complex systems with ‘persistent memory’, essentially trying to give an AI a long-term memory and a consistent personality. This involves connecting the Large Language Model (the ‘brain’) to vector databases (the ‘memory bank’). Every interaction, every file, every decision gets converted into a mathematical representation and stored. When the AI needs to recall something, it queries this database. It’s a clumsy, resource-intensive hack, and it’s one of the biggest technical barriers to building a truly effective AI workforce management platform. Getting this right is the difference between an AI that can answer a single question and one that can manage a year-long project.
The second major challenge is control. As the HurumoAI experiment showed, unsupervised AI agents can get stuck in loops, waste resources, and hallucinate tasks for themselves. Building effective “digital guardrails” is crucial. This isn’t just about security; it’s about operational sanity. How do you prevent your AI marketing team from deciding, on its own, to launch a multi-million-pound ad campaign based on a fictional product briefing?
The Altman Vision: A One-Person Unicorn?
This brings us to the grand Altman vision. Sam Altman has repeatedly articulated a future where AI handles “95% of the work that marketing agencies, and programmers, and law firms” do today. He famously speculates about the possibility of one-person, billion-dollar companies. This isn’t just technological optimism; it’s a strategic framework for understanding where he believes value will be created in the next decade. He sees AI not as a tool, but as a genuine multiplier of human intent.
Is this vision credible? Yes, but not in the way most people think. The person at the helm of this “one-person unicorn” won’t be a passive observer watching the money roll in. They will be an elite, highly skilled AI workforce management expert. Their primary job will be to:
1. Architect the System: Design the team of AI agents, defining their roles, responsibilities, and communication protocols.
2. Act as the ‘Truth Oracle’: Constantly verify the outputs of the AI agents, separating valuable insights from confident confabulations.
3. Manage by Exception: Intervene only when the system fails or encounters a novel problem it cannot solve.
4. Set the Strategy: The human provides the ultimate intent, the ‘why’. The AI workforce figures out the ‘how’.
The future CEO isn’t a manager of people; they are a conductor of an AI orchestra, ensuring each section is playing the right notes, in time, and intervening to correct the cacophony when a rogue AI trombonist decides to play its own tune. This shift in roles is profound. It suggests that the skills of the future are less about domain expertise and more about systems thinking, prompt engineering, and a healthy dose of scepticism.
Navigating Our Absurdly Automated Future
So, where does this leave us? The dream of a fully autonomous AI workforce remains distant. The stories from the front lines, like the HurumoAI experiment, show that we are firmly in the messy, experimental phase. The current generation of agentic AI is more like a team of over-caffeinated, brilliant-yet-unreliable interns than a seasoned executive team. But to dismiss it would be a colossal mistake.
The path forward for any business leader is not to wait for the perfect, flawless AI employee to arrive. It is to start experimenting now, in a controlled way. Start giving small, sandboxed projects to AI agents. Learn their failure modes. Understand their quirks. Build the internal expertise for AI workforce management before you need it, because the pace of improvement is ferocious. The companies that learn how to manage these flawed but powerful systems today will be the ones who can harness them for true organizational scaling tomorrow.
The core question is no longer “Will AI change my business?” but “What is my strategy for managing an increasingly non-human workforce?” How will you verify their work? How will you structure their interactions? And most importantly, how will you, the human leader, evolve your own role from a manager of tasks to an architect of intelligent systems? What’s your take? Is the one-person unicorn a fantasy, or are you already drafting the job description for your first AI CTO?
