Essay & podcast · AI & democratization
AI lowers the barrier to build
A few years ago I had something I wanted to build and could not. It wasn't the idea. It was everything between the idea and a working version of it: the team, the capital, the specialized skills I didn't have and couldn't fake. So it stayed an idea.
In June I shipped a website and an agent without writing a line of code, and wrote about it here. That was a small thing. What I want to describe this time is a larger one.
What got built
Over the summer I designed one integrated set of agentic rails for a wealth team. Mail, calendar, messages, projects and tasks, people and households, files, with an agent in every surface. It exists today as a single file, about 425 kilobytes, every view wired and running on seeded data. Click anywhere and it does what it should.
I have never coded anything before. I started building web applications in earnest in the spring. The prototype went through four versions in nine weeks, and the fourth is now frozen as the build target for the engineer who is putting a real platform underneath it.
What changed isn't me. It's the floor.
Where the work went
The work was deciding, and there was a lot of it.
A prototype like this is a few hundred decisions, and each one was mine. Whether there is one composer or many. What happens to a pasted link. Whether a list can be folded. What color an agent is. Which verbs exist. Most of those decisions took longer to make than they took to build, and the building was the part I did not do.
Some of them I will write down here, because they are the product.
One interface. A single composer sits at the bottom of every view. I do not want a hunt-around-for-buttons experience. You type; routing is invisible on the way in and visible on the way out, so you can redirect it.
The agent is a user, not a feature. Every person has their own agent. The firm has a set of permissioned agents by discipline. Agents propose; people accept. Every write has a named actor.
Nothing is deleted. Archive is the only verb. A correction is a new record that supersedes the old one, which is the difference between an audit trail and a document someone can edit.
Simplicity over borrowed features. Keep what makes Gmail or Slack useful and drop the customization surface. We're using Gmail through API — don't deviate wildly from its core function. Databases get a filter in every column header and never a separate search box. No emoji picker. No taglines anywhere inside the product.
Never rebuild. Copy the prior version, change only what was asked, verify it.
None of those are engineering decisions. They are taste, and rules, and a view about how a team should work. That is the new literacy: learning to direct, and to supply the judgment and the context the tool cannot supply for itself.
Why it matters that one person could do this
Step back from my project.
The tools that used to take a team and a decade to assemble are now available on demand. For most of this industry's history, the reason nobody built rails like this for a wealth team was cost. Enterprise software solved this problem for other industries. This industry could never afford enterprise software, so it got solved by people instead.
The cost of acting on an idea is falling toward zero for anyone willing to wrestle with the decisions. There is a massive advantage in starting today and building what ought to exist, rather than waiting for a vendor to sell you a version of it built for a broad audience.
Two ways to meet the moment
The technology amplifies whichever posture you bring.
One person uses it as a shortcut: hand over the judgment, ship something they don't understand and can't stand behind. They get faster at being wrong.
The other uses it to build and to learn, to do things they couldn't do alone and to understand more by doing more. They compound.
The case against
Judgment matters more now, not less. The agent is only as good as the person directing it, and you have to know what good looks like and care enough to fix what isn't. There is a real risk of a flood of confident work from people who skip that part. Against that, building and learning from mistakes has never cost less, and that is what opens new learning and innovation curves.
The tools that lower the barrier for me could concentrate power rather than spread it, if the largest models end up owned by a few. An outcome centered on one provider is suboptimal, because the services are not likely to be tailored to any one craft.
What the tools don't supply is the wanting. The reason to build in the first place is the opportunity for those who practice a specific craft, like evidence-based investing, to define the future of wealth and investment services.
The invitation
The first user of this system is us. The next step is to move our own work onto it and off the separate tools we run today, and to find out in public what is missing. I'll write that down too. After us comes a small founding consortium of advisors, by application.
Whatever you have been carrying as someday, when I have the team and the money and the skill: the barrier is lower than you think, and it is still dropping.
If you're tired of janky systems, come help us make things.