What an AI-First Company Actually Looks Like From the Inside

An AI-first company is one where the operation is designed around AI before people are hired into it. You automate the work first and hire for what is left. An AI-assisted company does the opposite: it keeps its structure and gives the same people better tools. Both can run the same software. The difference is the order of decisions.

That order is how I describe Dronehub on my LinkedIn profile: the company "was rebuilt around AI before it was staffed: automate the operation first, hire against what is left." This piece is what that means from the inside, where it breaks, and what the best-known public examples, Shopify and Duolingo, show about the difference between a memo and an operating model.

AI-first vs AI-assisted

The table is the definition. If a company answers the first three rows the same way it did before AI tools arrived, it is AI-assisted, whatever its website says.

Where my definition comes from

I didn't start with a theory. Dronehub builds autonomous drone-in-a-box systems: the docking station, the automatic battery swap, and the AI that inspects what the drone sees. On my profile I make the link to the business directly: "Autonomy works at this price because the company behind it was rebuilt around AI before it was staffed."

The phrase that matters is "before it was staffed." Dronehub existed long before today's AI tools, so the rebuild was a decision about the order of work, not about software. In practice the rule reads like this: when a function needs more capacity, the first question is which part a machine can do, and the hire goes to the remainder.

It sits on top of a simpler rule I've used for years and put on my About page: if I do something twice, I think about automating it; if three times, I automate it. I wrote up how that rule works separately. The AI-first version applies the same counting to a whole company instead of one person's week.

What it looked like at Dronehub

I described it in public in a July 2024 interview with Biznes i Styl. Dronehub then employed twenty-something people in Poland and the US, and I said AI had transformed our sales, marketing, documentation and code writing. My words in Polish were "rewolucyjne oszczędności": revolutionary savings. In the same interview I said we were planning to hire more people.

Those four functions are where AI-first shows up first in almost any company, because the work is text-heavy and it repeats. What the split between machine and human means in each, in general terms:

• Sales. The machine does research, first drafts of outreach and proposals, and pipeline hygiene. People own the relationship and the close.
• Marketing. The machine produces drafts, variants and schedules. People decide what the company says and check that it is true.
• Documentation. The machine produces first drafts from structured sources: specifications, reports, technical files. An engineer reviews and signs.
• Code. Engineers write with AI and review every change. The review is where the accountability sits.

The hiring point surprises people, so I'll repeat it. AI-first did not mean no hiring. It meant each hire went to work the automation could not do: judgment, customer relationships, engineering decisions, and building and maintaining the automation itself.

What stays human

Automation moves work. It does not move accountability. Someone still signs the inspection report, owns the customer and answers for every number in a program report. Dronehub's R&D has included programs backed by the European Space Agency, the European Defence Agency and Horizon 2020, and in programs like those the company answers for everything it signs. That is where most popular AI advice falls apart: it assumes nothing has to be signed.

The design rule that follows is simple. Split every process into judgment and execution, automate the execution, and put a named person in charge of every automated output. I wrote about that split in Human + Machine, and about keeping a human gate on anything that touches a customer or money in How to Build Your First Useful AI Agent. It is the same thing we do with drones: the robot takes the climb, and a person decides what the finding means, as I described in Autonomy and the Future of Work. Agents will take more of the execution over the next few years; I wrote about where they are actually going. They don't change who signs.

What AI-first costs

AI-first is not free, and the costs are not where people expect them.

Writing the process down. You can't automate work that lives in someone's head. Before anything is automated, someone has to describe how the function actually runs, exceptions included. It is slow, unglamorous work, and it is most of the job.

Owning the automations. Every workflow you build is software you now maintain: a model that changes, a vendor that changes its API, an edge case at 2 a.m. The build cost was never the number that mattered; the cost to own it forever is. I made that case in Build vs. Buy.

Building too much. When building gets cheap, you build everything and end up with a dozen half-maintained tools. The line I repeat from a 2024 PARP interview applies here too: "a company doing everything is a company doing nothing." Automate the work that repeats and matters, not everything that could be automated.

Hiring for a different profile. The person you hire into an AI-first function has to be able to build, supervise and fix the automation, not only do the task by hand. That profile is harder to find, and it is worth paying for.

AI-first company examples: Shopify and Duolingo

Two public cases are the ones most people mean when they search for AI-first companies.

Shopify. In a memo he posted publicly in April 2025, CEO Tobi Lütke said employees must show why they "cannot get what they want done using AI" before asking for more headcount and resources, called AI use a "fundamental expectation," and said AI usage would factor into performance reviews, CNBC reported. His framing question: "What would this area look like if autonomous AI agents were already part of the team?"

Duolingo. On April 28, 2025, CEO Luis von Ahn told staff the company would become "AI-first." Duolingo would "gradually stop using contractors to do work that AI can handle," and "headcount will only be given if a team cannot automate more of their work," TechCrunch reported. The backlash was loud. By August he said the memo "did not give enough context" and that Duolingo had "never laid off any full-time employees" (TechCrunch). In April 2026 Fortune reported that Duolingo had dropped AI use as a performance-review metric. On a podcast, von Ahn put it plainly: "if it can't, I'm not going to force you to do that."

My read: both memos state the rule I use, automate first and hire for the rest. What the Duolingo story shows is that a hiring rule announced by memo is not an operating model. If the operation hasn't been redesigned, "automate first" reads as a threat to the people doing the work, and it gets walked back. Measuring AI usage also measures the wrong thing, and von Ahn's own explanation, as Fortune reported it, says why: "rather than being held accountable for the actual outcome, we're trying to just push something that in some cases did not fit." I care whether the output got better and cheaper, not whether someone opened a chatbot.

How to tell whether a company is really AI-first

Five questions I would ask from the inside:

1. How was the last hire justified? If nobody asked what could be automated first, the company is AI-assisted.
2. Can I see the process document? Automation needs explicit processes. Work that lives in people's heads can't be automated.
3. Who maintains the automations? If the answer is "whoever built it, in their spare time," it will break the week that person leaves.
4. Who signs the output? Every automated output needs a named human owner.
5. What is measured? Output per person, cycle time and error rate, not seats and prompts.

How to start

Write down your operation, function by function, and mark each step as judgment or execution. Automate the execution that repeats: twice, think; three times, automate. Decide whether to build or buy each workflow, and borrow from the patterns that actually save hours. Put a named owner on every automated output. Then hire against what is left.

If you are a small team, start with the first high-leverage automations, and don't confuse a demo with a working system; I wrote about what entrepreneurs get wrong about AI for exactly that reason.

This is the method I teach through VADYM.AI in Ukrainian and KIERUNEK.AI in Polish. The tools will keep changing. The order of decisions is the part that lasts.