The case for staying small in the AI era

Most companies are still hiring like the productivity curve has not moved. We have been running Buzzed as a deliberate experiment in the opposite direction. Five people, a tight AI stack, and a refusal to add headcount until we have actually exhausted what the tooling lets us do at this size. So far we have not come close to the ceiling. This post is what that has actually looked like, where the productivity stops, and the political argument hiding underneath the practical one.

We did not build it this way because small is virtuous. We built it this way because small AI-augmented teams are the most important counterweight to a future where every operator ends up working for one of three platforms. That argument lives at the bottom of this post. Skip to it if you want the thesis. The middle is the concrete version: what a five-person team can do in 2026.

What we are, deliberately

Buzzed is five people. We build production AI systems for other teams. The kind of work that used to require a twenty-person agency with three project managers, a delivery lead, four senior engineers, an ML specialist, a designer, and a scrum coordinator whose job was mostly running the standup. We do not have any of that. We have a model stack, a small set of tools we know deeply, and a habit of saying no to anything that would force us to grow.

The clients are not small. Most have engineering organizations an order of magnitude larger than ours. They hire us because we ship in weeks what their internal teams would scope for two quarters. The reason is not that we are smarter. We have built the team around the assumption that the productivity gain is real, and we have removed the coordination cost that eats most of an internal team's capacity. Coordination is what fifty engineers spend half their day doing. Five engineers have very little of it.

What five people can ship now

Work we have shipped in the last twelve months that would have been a multi-person, multi-month engagement at any agency we used to work at:

• A document-extraction pipeline running on tens of thousands of contracts a week. Two engineers ran the build: one owned the spec, the eval suite, and the production pipeline; the other ran the integration and built the dashboards. Four weeks, end to end.
• A deployment of Ava, our customer communication AI, for a support team that cut first-response time by half. Three weeks including the eval harness. The first version of this would have been a quarter of work for the client's internal team and would have been deprioritized for a roadmap item.
• A research workflow for an investment team that turns a four-hour analyst task into an eighteen-minute one. The deliverable replaces a workflow that previously consumed three associates a week.

None of these required heroics. They required choosing the boring architecture (workflows, not agents, as we argue in pick the boring one), keeping the eval suite tight, and refusing to add complexity the customer would have to maintain.

Where the productivity stops

It would be a less interesting post if we pretended this was free. Three places we hit ceilings that staying small does not magically clear.

Sales relationships. AI does not buy your team a reputation. It does not get a CTO on a call who would not have taken the call from a small shop. The early clients have to come from your existing network. We built the company on relationships from the previous decade of work.

Judgment under genuine ambiguity. The model is very good at executing a plan and surprisingly bad at deciding which plan is worth executing. The strategic conversations about scope, architecture, and which customer is right for the company still happen between humans on a call. The productivity gain on the doing has gone way up. The productivity gain on the deciding has barely moved.

Accountability. When something breaks, a human has to call the customer. A human has to take the heat. A human has to decide whether the right move is to refund, rebuild, or push back. We have not found a way to delegate that. We are not in a hurry to try.

The political argument hiding underneath

Here is the part most productivity-focused versions of this post leave out. Small AI-augmented teams are not just a cute productivity story. They are a structural counterweight to a future where every operator ends up working for one of three platforms.

The default trajectory of the AI economy looks like this. The models concentrate at the top. Three or four labs with a durable lead. The application layer concentrates next. Meta, Amazon, Google, a handful of platforms own the surfaces where users meet AI. The vast middle that used to be employed in knowledge work gets reorganized into a creator class that earns from the platforms, depending on the platforms for survival. That is corporate feudalism. It is also where we are headed if nobody bends the curve.

Small AI-augmented teams are one of the few real counterweights. Each one is a unit of economic activity that does not flow through a platform. The team owns its tools, its relationships, and its margin. The clients own their data. The capacity that used to require a fifty-person agency now sits in a team that can say no to a contract, choose its own stack, walk away from a customer it does not like, and live without algorithmic distribution. That is what economic independence looks like in the AI era.

The platform-feudal version of the future is what happens by default. The small-team version requires people deliberately choosing it. Choosing not to scale. Choosing not to take the platform check. Choosing not to hand the customer relationship over to a marketplace. This is a choice each operator makes one engagement at a time. The cumulative effect is the difference between an economy with a healthy middle and one without.

What this means if you are considering it

Practical things we have learned, for anyone weighing whether the small-team path could work for them.

• Pick a stack and stay there. The productivity comes from depth in a small set of tools, not breadth. We use fewer tools than our clients do. We know the ones we use better than they know theirs.
• Charge for outcomes, not hours. The hourly model penalizes productivity. If you can do in a day what used to take a week, price the value, not the time, and keep the surplus.
• Refuse work that requires you to grow. Every piece of work that would force you to double headcount is work that compromises the model. Sometimes the right answer is to walk.
• Own your distribution. Build a list, write publicly, do not let a platform sit between you and your next customer. The whole point of the small-team model is that you are not dependent on someone else's rails.

We are not arguing that every company should be small. Most ambitious things in the world require more than a tiny team. We are arguing that the version of the AI future where small teams thrive is materially different from the version where they do not, and that every operator who chooses the small path is casting a small vote against the alternative.

The productivity is a feature. The independence it buys is the point.