Why your AI bill is twice what it should be

Why the engineer who knows your codebase in their heart is the only one worth hiring in 2026.

A client engagement, a few weeks back. Routine work. Update a design spec across a React Native app. The kind of task AI is supposed to be incredible at.

I let the agent run. It came back and told me confidently it had updated the design. The summary looked clean. The markdown file describing the work was beautiful.

Whole pages were untouched.

If I had not been in that codebase for months, if I did not know it, I would have shipped a broken update to a client and learned about it from a screenshot on Slack three days later.

That moment is the entire hiring problem in 2026. And almost nobody is hiring for it.

The CEO's blind spot

There is a specific failure mode I have started calling AI psychosis, and CEOs are uniquely prone to it.

You are the furthest person in your company from the last mile of delivery. You read the status update. You watch the demo. You see the agent dashboard light up green. You assume the thing is done.

Right now, a lot of things that look done are not done. AI-generated code that nobody truly understands is not just technical debt. It is liability. The bill comes due at 11pm on a Friday when something breaks and nobody on your team has the muscle memory to find it.

This is the framing every founder needs to internalize before they post the next job rec:

You are not hiring for code output. You are hiring for the judgment that catches the agent when it lies.

Savoir and connaître

The French have two ways of saying "to know." Savoir and connaître. To know intellectually, and to know by acquaintance.

I know Flutter intellectually. I can read the docs, build a component, tell you how it works on paper.

I know React Native in my heart.

The difference shows up the second something goes wrong. When a screen feels choppy, I do not open Stack Overflow. I already know. Someone used a ScrollView when they should have used a FlatList. I can feel it before I trace it.

That kind of knowing is insurance against AI liability. The engineer who knows the codebase in their heart spots the gap in 30 seconds. The engineer who only knows intellectually believes the summary and ships the bug.

This is not a soft claim. It is structural. Your agent has a million-token context window that is not actually a million tokens of useful memory. It is filled with the agent's working understanding of your codebase at that moment. It does not have the decision your team made in sprint 16. It does not have the new design direction the CEO laid out last Tuesday. When the agent has been running for 40 minutes, a lot of that gets compressed and thrown out. The judgment layer cannot live in the tool. It lives in the engineer.

The uncomfortable truth: we are training a generation of juniors on a workflow that skips the part where you get burned. They prompt their way out of every fire instead of sitting in one. They never develop the calluses. Meanwhile, the seniors who already have those calluses are retiring out of the industry.

The supply of engineers who know things in their heart is shrinking at both ends. The demand just spiked. You should be reading this and adjusting your compensation bands accordingly.

The economics of knowing

A great engineer needs $200-$300 a month in AI tooling now. That is the cost of equipment.

Watch what the same companies are saying right now, weeks apart. One week: use as many credits as possible, never even open the IDE, go crazy. The next week: we are using too many credits, the bill is bigger than doubling the workforce. Both messages are wrong. Both come from companies that have never asked the actual question.

Using a scarce resource efficiently is the skill. Engineering has always been about cycles, memory, throughput, server cost. AI credits are just the latest scarce resource.

The senior who knows the codebase in her heart gets the agent to ship the right diff on the first try, for maybe a dollar in tokens. The mediocre engineer prompts and re-prompts and burns $1,000 on a feature that mostly works. On the scoreboard of who used the most credits this month, the bad engineer looks like a star. In reality, you just paid $999 to ship a worse outcome.

You are not paying for AI credits. You are paying for the judgment that decides where to spend them.

The hiring filter

Stop running the 2019 playbook. LeetCode is dead. Take-homes are dead. Every candidate has Claude open in another tab. You are not testing them. You are testing the model.

Here is what actually works:

Hire internally first. Engineers hate working with mediocre engineers. They will never recommend someone they do not want to sit next to. Their reputation is on the line. Highest signal hire you can make.

Require a real cover letter. Not AI-generated. Have them write about one specific thing that broke on them in production and how they figured it out, plus one lateral interest outside of code. Not video games. Something else. Brazilian jiu-jitsu. Woodworking. Yo-yo. You will interview on both, and you will find out fast whether they actually know what they claim. If they skip the cover letter, throw it away.

Plant a weird instruction in the post. Ask for their favorite mac and cheese recipe. Ask them to include the Contra skip code somewhere in the cover letter. Something specific and small that has nothing to do with the job. This sieves out everyone using automated submission tools. Not foolproof. Just necessary.

First interview is conversational. No tests. Just talk about the depth thing and the lateral thing. If they cannot speak about either with weight, you are done.

Then bring them in for real work, 30 to 60 days, paid. With AI eating every coding test, watching judgment on your actual codebase is the only signal that survives. Pair program. See how they get unstuck. See how they spend credits. Everything else is theater.

Hire for obsession

Underneath judgment, the trait you are actually hiring for is obsession.

You cannot fake it. You cannot pick it up over a weekend with Claude. The candidate who is obsessed about the platform AND obsessed about their lateral thing — that pattern almost always plays out in production.

We actually have a yo-yo champion on the Vermillion team. He is also a great designer. Not coincidence. The person who becomes a champion at one obscure thing has the wiring to become a champion at the next thing they decide matters. That wiring is what you are buying.

Obsession breaks through every AI coding interview. It breaks through ten years on a resume. It breaks through credentials.

If you find it, give them the trial. If they perform, lock them in.

The story being told right now is that AI will automate engineering. It will not. It will automate the parts of engineering that were already automatable. Judgment, obsession, and heart-knowing got more valuable, not less.

The market just has not repriced yet.

You can wait for it to. Or you can hire ahead of it.

The Cold Start, by Ken Vermeille. Forward this to the founder who just asked you why they need engineers anymore.