How Do You Compete When AI Can Do What You Do?

What were you actually being paid for?

The doing, most likely. The part that took skill, and hours, and people who knew how.

That part is getting cheap, fast. And there’s now a number on how fast.

Did AI really make what we do easy to copy?

The short version: it depends entirely on how long the work takes.

METR is a research nonprofit whose stated work is measuring whether and when AI systems might threaten catastrophic harm. One thing they measure along the way is what AI can actually finish on its own. They tested frontier models from 2019 through November 2025 against multi-step software tasks, timed by how long a human professional needs. Their headline finding is that the length of task an agent completes on its own with 50% reliability has been doubling roughly every seven months for six years.

But the more useful number is where the line falls.

Current models succeed on close to 100% of tasks that take a human under four minutes. On tasks that take more than about four hours, they succeed less than 10% of the time.

So the honest answer is that AI ate the short work completely, and the boundary keeps moving right.

If what you sell can be broken into four-minute pieces, it’s already commoditized. If it takes four hours of judgment, you have time. Not forever.

Why can a competitor match our features in a quarter now?

Because a generalist with a subscription now does what a specialist team used to.

Stack Overflow surveyed 33,662 developers in 2025 and found 84% using or planning to use AI tools, up from 76% the year before. Among professionals, 51% use them daily.

Read that as a supply statement. The capability that used to sit inside your team now sits inside a twenty-dollar subscription, in every company in your category, at the same time.

And it isn’t only code. Graphite analyzed 65,000 English-language articles and found roughly half of all new online articles are now AI-generated, which we covered here. The marketing function got the same upgrade the engineering function did.

When everyone gets the same upgrade at the same moment, nobody gets an advantage. Everyone gets the same output, faster.

That’s how the Sea of Sameness gets built now. It is a market in which competing companies, each following the same widely recommended best practices, become mutually indistinguishable to the buyer. It is produced by diligence rather than laziness. Best practice is now available as a subscription, so diligence is cheaper than it has ever been, and so is sameness.

So what’s left that can’t be copied?

Four things, and only one of them is free.

• Proprietary data. Real, and expensive to accumulate.
• Workflow lock-in. Real, and slow.
• Distribution. Real, and usually bought.
• A point of view nobody else holds. Free, and already in your building.

The last one is the only one you already own and haven’t spent. It’s also the only one a model structurally cannot produce, because a model generates the consensus of what has been written. It cannot generate the thing you know that has never been written down.

That thing has a name. Your One Unforgettable Idea is the single idea a founder owns so completely that they stop being compared and become their own category. It comes out of them rather than from an agency, and it was there from the beginning.

Here’s the test. It takes two minutes.

Ask ChatGPT to explain what your company does and why someone should pick you. Then ask it the same question about your closest competitor.

Put the two answers next to each other and swap the company names.

Do both still make sense? Then you just read a description of your category. The model wrote it without ever hearing from you.

Now hunt for one sentence in your answer that your competitor could not honestly say about themselves.

Most founders can’t find one. That’s the whole problem, in about two minutes, for free.

What argues against this?

The strongest counterargument comes from the same survey I used to make the case.

Developers are adopting these tools and trusting them less. Stack Overflow found only 3.1% highly trust the accuracy of AI output, while 45.7% distrust it to some degree, up sharply year over year. 66% report dealing with answers that are almost right, but not quite. And 45.2% say debugging AI-generated code takes longer than writing it themselves.

So “AI can do what you do” is too strong, and I should say so plainly. What AI does is get to almost-right, quickly, on bounded work.

Two things follow, and they point in opposite directions.

Almost-right is enough to destroy a commodity business, because a buyer comparing two adequate options picks on price. Almost-right is not enough to replace judgment on long, ambiguous, high-stakes work, which is where the four-hour tasks live.

There’s a second objection worth carrying. METR itself notes that some figures in that post are out of date and that measurement uncertainty could move the timeline by around two years. The direction is well evidenced. The date on any particular capability is not.

What should you do about it this quarter?

Three things, and the first one is free.

• Sort your work by how long it takes. Anything a competent person finishes in under an hour is on the wrong side of the line. Anything that takes a day of judgment is defensible for now.
• Look at who the engine cited. If the answer named sources, those companies are teaching the engine what your category means. Right now they are doing it without you.
• Write down the thing you believe that your competitors would argue with. Not your values. A claim. If nobody in your category would dispute it, it isn’t a position.

The founders who get through this aren’t going to be the ones who adopted the tools fastest. Everyone adopted the tools. They’re going to be the ones who had something to say that the tools couldn’t say for them.

That’s a strange kind of good news. The one asset that still works is the one you’ve had the whole time and never wrote down. Getting it out is excavation work, and it starts with a conversation rather than a content calendar.