The AI Workforce – What Gets Expensive When AI Gets Cheap?

The debate about whether AI replaces or augments workers misses the main point. The important question is how AI changes the costs of different parts and deliverables of a job, and which tasks become more valuable as others get cheaper. Keep your attention on this, since that’s where the real value is.

In most knowledge jobs, creating a decent first draft now costs almost nothing. Whether it’s a basic contract, a simple financial model, clean code, or a market summary, the average version has become a commodity. This isn’t just a prediction; it is already happening in companies using these tools.

Some might think this means we need fewer people, but often the opposite is true. When it becomes cheaper to produce work deliverables, demand increases. You might order five reports instead of one, or create drafts you would not have used someone’s time to write before. The work does not go away; it grows and changes.

But here is what most people miss: when the easy 80 percent of a job becomes free, the last 20 percent becomes the whole job. That 20 percent (i.e., judgment calls, tricky situations, and moments where mistakes are costly) was always the hardest part. The work did not become easier; we have just removed the simple parts and left only the tough ones.

This creates challenges that current organizational structures are often unprepared for.

This is the biggest risk people are missing right now, and very few are dealing with it.

Junior employees used to become senior by doing repetitive tasks. The associate who wrote hundreds of memos learned to spot problems over time. The analyst who built the same model many times developed a sense for when results didn’t add up. This routine work was actually training for good judgment, and now judgment is the most valuable skill.

Now that routine work is automated, some companies have lost their old way of developing senior talent. Current senior staff have gained their judgment through experience, but they will eventually retire. Newer employees are now expected to supervise work they never learned to do themselves. This leads to a group of reviewers who have never been reviewed themselves.

You might not see the effects of this for some years, so current leaders may not worry about it. If you want your company to stay strong, you need to rebuild the experience training process you lost due to AI. Give employees real chances to develop judgment, since it won’t happen by accident anymore. Make sure people do the work before they supervise others. Have them question the digital AI output and explain their thinking. Free output helps your operations, but it can hurt your talent pipeline, so you need to manage both.

When producing work was expensive, you usually trusted it because a senior person made it, it took real resources, and it was tied to their reputation. Double-checking was rare.

Now, work can be produced quickly and without a clear author, so someone must decide if the deliverables are trustworthy. That person is costly, and their time constraints limit how much value AI can provide.

Here’s the truth: for many tasks, checking AI’s work can cost more than doing it right from the start. If a model writes a clause in seconds but a senior partner spends twenty minutes checking it for errors, you haven’t saved time. You’ve just moved the work from juniors to seniors and called it efficient. Productivity only improves when checking is cheaper than creating. For each task, you need to know which situation you’re in.

Most companies don’t track the cost of checking work. It’s hidden in salaries and tasks labeled ‘review,’ so it never appears to be the real bottleneck. The companies that succeed will be the ones that measure this cost, understand what it takes to trust an output, and assign work based on that knowledge.

Don’t focus on whether a task is high-stakes or low-stakes. That can make you worry about the wrong things. Instead, decide where to use AI based on how easily and cheaply you can fix mistakes before they get worse.

Use AI freely and without human oversight in areas where it’s easy and inexpensive to check and fix mistakes. This includes drafting documents, conducting initial research, or writing code with robust testing. If errors show up quickly and are easy to correct, let the AI handle it.

Be very careful in areas where mistakes are hard to spot, can get worse over time, or cannot be undone. This includes client commitments, regulatory filings, or anything sent to customers that can’t be easily fixed. In these cases, let the AI draft, but a human must take full responsibility for the final work, with their name attached and a real reason to check it carefully.

This approach doesn’t match traditional roles or departments. Some high-status work is easy to reverse and should be automated more than people expect. Some lower-status work is hard to undo and needs careful human oversight, even if it’s not obvious. Where AI fits in your company won’t match your org chart, and that’s why many AI projects don’t work well: they follow the old structure.

When you combine these ideas, the staffing needs become clearer. It’s not just about everyone working with AI.

The middle layer of staff shrinks, not because mid-level employees lack skill, but because machines handle routine, well-defined tasks best and at the lowest cost. What you need more of are people at both ends: seniors who can make complex decisions and spot mistakes, and mid-junior roles who are skilled with AI tools and can get much more done by using them as instruments, not as sources of truth.

The risk is letting the middle layer slowly disappear as people leave and replacements aren’t hired. If that happens, you lose the link between skilled juniors and experienced seniors: there’s no way for juniors to become seniors because the steps in between are gone. This structure only works if you intentionally build connections between the two ends, which ties back to the need for proper training and systematic deliverable-focused audit and oversight. Both issues are connected and must be solved together.

Follow these three steps in order.

First, decide where to draw the line for reversibility. Sort your actual workflows by how easy and cheap it is to fix mistakes, and be honest. Many so-called ‘high-stakes’ tasks are easier to reverse than people think, while some routine processes are harder. Automate as much as possible below this line, and put strict controls above it.

Second, figure out the real cost of checking AI’s work. For your top ten workflows, measure how much senior staff time it takes to trust AI output. If checking costs more than creating, you don’t have a productivity tool. You have a risk that looks good on paper. Know the difference before you expand.

Third, rebuild the career path you removed. Decide clearly how a new hire today can become someone you trust with big decisions in a few years. If your answer is just ‘they will learn on the job’ – that will surely not be enough. Create real opportunities for practice. Make sure people do the work before they supervise others.

This isn’t about whether AI is your partner or your replacement. That’s a question for philosophers, not business leaders. Business leaders must solve a different problem. The cost of doing the work has dropped, but the value of judgment has gone up. The key is to focus your attention on the expensive, important parts and let the routine work stay low-cost.