If AI has increased your workload instead of reducing it, you are not doing it wrong. The interface is lying to you.
The Remote Intern Problem
A recent Wall Street Journal article describes a consistent gap: top executives across organisations believe AI is saving their workforce eight hours a week. The workforce says it is saving them zero.
This is not a measurement error. It is a design failure.
Big Tech sold AI as a co-pilot — a senior partner who reads the room and jams on ideas in real time. What we actually got is a remote intern. This intern has read every book in the world but has never worked a day in its life. It has zero context about your business, your workflows, your team dynamics, or your clients’ quirks. It is eager to please, completely confident, and frequently wrong.
After a year of heavy use, I stopped treating AI as a tool. I started treating it as a subordinate. Without realising it, I had become a manager of a new kind of employee — one who cannot collaborate the way a colleague does and who needs clear written instructions for everything. The cost I was paying was not a technology cost. It was a management debt.
Sync vs Async Physics
To understand why this happens, borrow a concept from software engineering: synchronous versus asynchronous execution.
Sync work is how we collaborate with people in real time. High bandwidth, shared context, implicit understanding. I can say “fix that issue” to a team member and they know exactly what I mean because they were in the room when it broke last week.
Async work is like an API call. Zero shared context unless you specify it. You cannot send a vague request to an API and expect useful output. You need to specify the endpoint, the parameters, and the expected format precisely.
AI is async. The chat box is designed to look sync. That gap is where productivity dies.
The Chat Is a Trap
I treated AI as sync. It functions as async. Most days I access AI through a chat interface that mimics how I communicate with my team on WhatsApp. That interface signals to my brain that we are in a sync relationship.
So I chat. I brainstorm. I iterate on ideas as though I am talking with a colleague. But AI has no memory of my company’s history. It does not know what I mean. It only knows what I wrote in that specific prompt.
When you use a sync workflow for an async tool, you create what Chroma Research calls context rot.
I discovered this while building O9X. I would spend hours chatting with AI about code architecture. It felt productive. But as the conversation extended, the AI drifted. It forgot the constraints I set an hour earlier. It started producing solutions that violated my core business logic. In the early months, I spent days debugging code that looked correct but was functionally broken. That was the AI tax.
The AI Tax Is Real
A Workday survey of 3,200 employees in late 2025 found that 85% saved one to seven hours weekly using AI. But nearly 40% of that time disappeared into rework — correcting errors, rewriting content, verifying outputs.
Another study found knowledge workers spend 4.3 hours per week just checking AI outputs for accuracy. Half a workday auditing a tool that was supposed to save time.
Only 17% of people never have to rewrite prompts to fix false information. The other 83% are iterating blindly.
This produces an odd split in who pays the cost. The CEO uses AI to compress a thousand-page report into a one-page brief. For them, it works. They consume the final output, shielded from the mess of creation. The junior uses AI to get a task off their plate fast, skims the output, and hits send. The middle manager receives that work, spots the hallucination, finds the tone is wrong and the facts are based on an older pricing model, and now has to reverse-engineer the prompt to understand why the AI lied.
The junior saved 30 minutes. The manager lost three hours. The AI tax is paid entirely at the operator layer.
Hallucinations are not rare edge cases. Average rates sit between 3.7% for top models and 16.9% overall, with some newer reasoning models hitting 33% to 48% on specific benchmarks. In 2024, 47% of enterprise AI users made at least one major business decision based on hallucinated content.
The Spec Solution
Over the past year I stopped using chat as my primary workflow. I shifted to a spec-based approach and treat AI the way you would treat a capable freelancer who lives in a different time zone, does not speak your language perfectly, and has never seen your business before.
You cannot chat with that person. You have to send them a brief.
The Remote Intern Protocol has three rules.
The first rule is no prompting without a brief. I do not open any AI tool until I have written down what I want — my rules, my constraints, my expected output format. If I cannot articulate that clearly in writing, I am not ready to ask the AI. This exposes lazy thinking before it becomes broken output.
The second rule is treat context as a file, not a memory. Never rely on AI to remember what you said ten messages ago. Context must be provided as a static file or refer to a living Google Doc at the start of every session. If context changes, update the file. Research on context rot shows this is not optional — it is essential.
The third rule is review your spec, not the output. The spec is the work. If the brief is good, the output will likely be usable. If the brief is lazy, the output will be wrong in ways that are hard to catch. Writing a good spec takes skill that most of us never learned. We learned how to do the work. Now we need to learn how to delegate it.
The Cost of Judgment
AI drives the cost of production toward zero. It drives the cost of judgment toward its real market value — which turns out to be high.
In a world where anyone can generate ten thousand words of code in seconds, the value of the person who can spot the single fatal flaw in a mountain of output skyrockets. We do not need fewer people. We need fewer typers and more thinkers. AI is not replacing humans — it is replacing unvetted execution.
If you want to survive the AI transition, stop hoping for a co-pilot. Start training yourself to be a precise manager of a fast, mediocre, remote workforce that never sleeps, never complains, and confidently produces plausible nonsense when given bad instructions.
The chat box is a trap. The document is the discipline.
P.S: Vanderbilt researchers developed 16 prompt patterns specifically for software tasks. They function as reusable solutions to common problems.