Do you still need an agency? What one AI-native senior replaces (and what they can't)

A client's front end had gone five years without its upgrades. Old runtime, old framework, build tooling nobody would touch. Work like this is normally scoped as a quarter for a small team, and agencies price it accordingly, because the risk earns the quote.

It went through in stages last year: codemods and AI tooling chewing the mechanical bulk, tests between each stage, a human deciding the order of operations and what to risk when. Around fourteen thousand lines changed. The app started four times faster afterwards.

And the small team never existed.

If you run a product company, sooner or later you price the same three options: an agency, a full-time hire, or one very good fractional person. The menu is familiar. What each line on it buys has changed, mostly in the last two years, and the story above is half of the change. This piece is about both halves.

In brief

• Most of an agency invoice is production work — and production got cheap, and is still getting cheaper. That purchase is being repriced whoever you buy it from.
• Judgment didn't get cheaper. Models are agreeable by design: ask one to help you build the wrong thing and it will help you build it faster.
• The twenty-year bundle of judgment and production is coming apart. The real question is no longer "agency, hire, or fractional" but where the judgment sits and what it's amplified by.
• Two evaluation questions do most of the work: what has the tooling shipped for you, and when did you last say no to something?

The half that collapsed

Take an agency engagement apart and most of the invoice is production. Hands writing code, process keeping the code moving, specialists covering the corners, review catching the mistakes. On the product where I do much of my work, a data-heavy US health platform, here is what happened to those lines.

Releases used to run down a manual checklist, which is another way of saying they occasionally failed in human ways. A release is now one command: it triggers from the terminal, links itself to the tickets, and writes its own notes on the way out. Half a day of senior attention became minutes, and the checklist category of mistake left with it.

Translations used to route through external tools by hand. Now AI drafts them and a human reviews them. Scheduled agents read the meeting notes, the tickets and the code, and surface what everyone missed while they were busy. Small things. But teams are mostly made of small things, and none of these needs a person any more.

So production got cheap, and it is still getting cheaper. If production is what you were buying, from anyone, that purchase is being repriced while you read this.

The half that didn't

The same year, on the same product, a different list. I demoted a P0 that nobody wanted to demote. I refused to estimate a feature until its scope stopped moving. I pulled a half-built capability out of a release rather than ship it soft. I killed a roadmap item we had already invested in, because the partner dependency underneath it was not going to hold, and I wrote the memo explaining why.

None of that came from a model.

Models are agreeable by design. Ask one to help you build the wrong thing and it will help you build it faster. And inside the code, the pattern repeats: on well-structured work the first pass is strong, but on a gnarly refactor in ambiguous territory it comes back right about half the time. Half right sounds usable. It is worse than nothing, because confidently wrong work that nobody can cheaply check costs more to unwind than it cost to produce.

So the tools amplify whatever judgment is in the room. That sentence cuts both ways, and which way it cuts for you depends entirely on who is in the room.

The question underneath the question

For twenty years the practical reason to hire an agency was that judgment and production came bundled. You could not buy one without the other, so you bought the bundle and hoped the seniority was spread thick enough. That bundle is now coming apart. Production keeps dropping in price. Judgment does not, and a bad call compounds faster than it used to, because everything downstream of it gets built at tool speed.

Which turns "agency, hire, or fractional" into a different question: where will the judgment sit, and what will it be amplified by? For a meaningful class of products, one senior person with full context and an AI-native toolkit now covers what a small team used to. Past a certain scale the team wins again; the threshold has moved a long way from where most instincts still put it. And whichever way you go, the answer partly depends on your own systems: whether AI tooling can work in your software at all is an audit question now.

Whoever you are evaluating, two questions do most of the work. What has the tooling shipped for you? Anyone competent now has a good answer. Then ask when they last said no to something, and what the no saved.

Sit with how long the second answer takes to arrive.

---

I'm Kiah — an independent product engineer and technical partner. I build complex, data-heavy products with an AI-native toolkit and the judgment layer that makes it safe. If the unbundling above sounds like the conversation you're currently having, I'm easy to talk to.